arXiv CSopen access

Force-Aware Neural Tangent Kernels for Scalable and Robust Active Learning of MLIPs

Force-Aware Neural Tangent Kernels for Scalable and Robust Active Learning of MLIPs

Eszter Varga-Umbrich1,2 , Zachary Weller-Davies1,† , Paul Duckworth1

arXiv:2605.13788v1 [cs.LG] 13 May 2026

Jules Tilly1 , Olivier Peltre1 , Shikha Surana1,†,∗ 2

1 InstaDeep, UK Department of Engineering, University of Cambridge, Cambridge, UK † These authors supervised equally. ∗ Corresponding author: [email protected]

Abstract Active learning for machine-learning interatomic potentials (MLIPs) must address several challenges to be practical: scaling to large candidate pools, leveraging energy–force supervision, and maintaining robustness when candidate pools are biased relative to the target distribution. In this work, we jointly address these challenges. We first introduce a linearly scaling acquisition framework based on chunked feature-space posterior-variance shortlisting. By avoiding materialisation of the candidate and train set kernels, this approach enables screening of ∼200k structures within hours and applies broadly to acquisition strategies that score candidates based on molecular similarity metrics. We then extend the Neural Tangent Kernel (NTK) to a force-aware setting via mixed parameter–coordinate derivatives, yielding a force NTK and a joint energy–force NTK that provide natural similarity metrics for vector-field prediction. We demonstrate the effectiveness of the joint energy–force NTK on the OC20 dataset, where force-aware acquisition is crucial: it achieves the lowest energy and force MAE and RMSE across all metrics and distribution splits. Across T1x, PMechDB, and RGD benchmarks, our force NTK methods remain competitive with established baselines while being significantly more efficient than committee-based approaches. Under a controlled candidate-pool shift case study on T1x, acquisition based on pretrained MLIP embeddings and NTKs remains robust, whereas committee-based methods exhibit higher variance. Overall, these results show that a single pretrained MLIP can enable scalable, force-aware, and distribution-robust active learning for foundationmodel fine-tuning.

1

Introduction

Machine-learning interatomic potentials (MLIPs) Behler and Parrinello [2007] predict energies and forces at orders-of-magnitude lower cost than density functional theory (DFT) and are now standard tools for atomistic simulation across molecular and materials systems Jacobs et al. [2025], Wang et al. [2024], Li et al. [2025]. Their accuracy is tightly coupled to the coverage and quality of the training data, a dependence that persists for pretrained atomistic foundation models Wood et al. [2026], Batatia et al. [2023a]: downstream accuracy still hinges on whether the curated fine-tuning set covers the target task Anstine et al. [2025]. Modern simulation workflows, including molecular dynamics, metadynamics Laio and Parrinello [2002], nudged elastic band methods Jónsson et al. [1998], Henkelman and Jónsson [2000], and uncertainty-driven samplers Zaverkin et al. [2024], routinely produce thousands to millions of candidate structures, since proposing geometries is orders of magnitude cheaper than DFT labelling. Dataset curation is therefore the practical bottleneck for deploying MLIPs in new chemical domains. Active learning (AL) Settles [2012] addresses this Preprint.

bottleneck by prioritising informative candidates for labelling; in offline pool-based AL the unlabelled pool is fixed in advance Zaverkin et al. [2022], Zou and Marzouk [2026], decoupling candidate generation from acquisition. Effective offline pool-based AL for MLIPs raises three coupled challenges. First, acquisition must scale to candidate pools containing hundreds of thousands of structures: committee ensembles add a retraining cost that grows with the labelled set, while kernel-based methods become memory-bound once similarity matrices no longer fit in memory. Second, MLIP labels comprise both scalar energies and per-atom forces, and the two are not interchangeable acquisition signals: forces, not energies, govern molecular dynamics and geometry relaxation, and a structure that most reduces force error need not be the one with the most uncertain energy. Energy-only acquisition can therefore systematically miss force-informative candidates. Third, candidate pools inherit the bias of the generators that produce them: molecular dynamics oversamples near-equilibrium basins, reaction-path methods emphasise chosen endpoints, and enhanced samplers encode bias through their collective variables and initialisation, so the pool distribution can differ sharply from the deployment target Deng et al. [2025], Cui et al. [2025]. Recent work has shown that model-based kernels constructed from pretrained equivariant MLIPs Varga-Umbrich et al. [2026], specifically energy NTKs Jacot et al. [2018], Zaverkin et al. [2022] and hidden-activation features Ouyang et al. [2024], outperform committee and fixeddescriptor baselines on small reactive-chemistry pools. Two limitations, however, constrain their applicability in large scale and diverse settings: pairwise similarity evaluations scale as O(nT nP ) and O(n2P ) in the training-set (nT ) and candidate-pool (nP ) sizes, and existing constructions capture only energy sensitivity. Neither strategy has been evaluated under controlled candidate-pool bias. We make the following contributions that together enable model-based acquisition to operate effectively in large, biased candidate pools: • Scalable feature-space acquisition via shortlisting: We introduce a feature-space posteriorvariance (PV) pipeline with shortlist-based selection that remains conditioned on the training set while avoiding O(nT nP ) train–pool and O(n2P ) pool–pool kernel computations. Candidates are scored in chunks using explicit d-dimensional feature representations, and a top-K subset is first selected by PV acquisition in O(nT + nP ) time. A diversity-based Largest Cluster Maximum Distance (LCMD) rule Holzmüller et al. [2023] is then applied to this shortlist to select a batch of size B in O(KnT ). By shortlisting in feature space, this approach scales linearly in both the training and pool sizes, enabling acquisition on ∼ 200k candidates and, in principle, much larger pools. • Force-aware NTK representations: We extend the Neural Tangent Kernel Jacot et al. [2018], Zaverkin et al. [2022], Varga-Umbrich et al. [2026] to force prediction by incorporating mixed parameter–coordinate derivatives ∇θ ∇r Eθ , yielding a force-aware NTK and a joint energy–force NTK. The joint representation subsumes the energy and force NTKs as special cases and enables interpolation between energy- and force-driven acquisition. • Robustness and scaling analysis: We provide a systematic evaluation across reactive molecular benchmarks (T1x Schreiner et al. [2022], PMechDB Tavakoli et al. [2024], and RGD Zhao et al. [2023]) and large-scale catalysis (OC20 Chanussot et al. [2021], Sahoo et al. [2025]), including controlled candidate-pool bias experiments on T1x that isolate interand intra-reaction shifts. We show that model-derived kernels remain robust under pool shift, while committee-based methods exhibit higher variance due to their reliance on point-wise uncertainty. We additionally characterise the practical operating regime of our methods with respect to pool size and feature dimension. On OC20, we find that incorporating force information is essential for effective active learning. The joint energy–force NTK combined with the shortlist-LCMD acquisition achieves the lowest energy and force errors across all metrics, outperforming competing methods on both in- and outof-distribution splits. However, we find that force-awareness is not always essential: across the reactive benchmarks, energy-based acquisition already captures most of the informative variation. Incorporating force information does not provide additional benefit; however, the force-only and joint energy–force NTKs still track the performance of energy-only NTKs closely, with the energy-only variant remaining the strongest baseline in this setting. Taken together, these results highlight the importance of methods that generalise across both regimes and scales. Overall, we find that the 2

joint energy–force NTK provides a consistently strong and robust acquisition strategy across the benchmarks considered in this work.

2

Related Work

Active learning for MLIPs: Most MLIP active-learning workflows are online, where structures are generated during molecular dynamics, reaction-path exploration, or perturbations and selected using uncertainty or extrapolation criteria from the current model. Representative approaches include committee-based methods Smith et al. [2018], Schran et al. [2020], Niblett et al. [2025], extrapolation criteria for moment-tensor potentials Podryabinkin and Shapeev [2017], Bayesian and Gaussian-process force fields Jinnouchi et al. [2019], Vandermause et al. [2020], and concurrent or uncertainty-driven sampling schemes Zhang et al. [2019], Kulichenko et al. [2023], Zaverkin et al. [2024, 2026]. Offline pool-based active learning instead selects from a fixed candidate set Zaverkin et al. [2022], Zou and Marzouk [2026] and is increasingly relevant for dataset curation and distillation from large atomistic datasets Levine et al. [2026], Sahoo et al. [2025], Horton et al. [2025]. In this setting, candidate pools can be large and biased relative to the target distribution, which can degrade acquisition performance under distribution shift Deng et al. [2025], Cui et al. [2025]. Acquisition signals and model-based kernels: Common acquisition strategies include committee disagreement Schran et al. [2020], Peterson et al. [2017], as well as more efficient variants such as shallow ensembles Kellner and Ceriotti [2024], multi-head pretrained models Beck et al. [2025], learned uncertainty heads Neumann et al. [2025] and Bayesian interatomic potentials Jinnouchi et al. [2019], Vandermause et al. [2020], Coscia et al. [2026]. Descriptor-based methods that use fixed representations such as SOAP Bartók et al. [2013], De et al. [2017], Himanen et al. [2020] or molecular fingerprints with Tanimoto similarity Morgan [1965], Rogers and Hahn [2010], Ralaivola et al. [2005] are typically combined with diversity objectives Zou and Marzouk [2026], Kellner and Ceriotti [2025]. More recently, model-based kernels derived from neural representations, such as energy NTK features Zaverkin et al. [2022] and activation kernels Varga-Umbrich et al. [2026], Ouyang et al. [2024], have been shown to be useful signals for pretrained equivariant MLIPs (see Holzmüller et al. [2023] for a general discussion on deep batch AL). However, existing demonstrations that incorporate diversity still compute full train–pool or pool–pool similarity computations Zaverkin et al. [2022], Holzmüller et al. [2023]. Feature-space posterior-variance methods are well known Rasmussen and Williams [2005], Holzmüller et al. [2023], and have enabled AL for candidate pools of up to ∼ 150k Zaverkin et al. [2022], but utilising them for efficient shortlisting has not been explored in the context of MLIPs.

3

Methods

3.1

Active learning

We study offline pool-based active learning for MLIPs. The learner is given a small labelled training set T (0) and a large fixed pool of unlabelled candidate structures P (0) , together with a per-round labelling budget B and a total budget of T rounds. Each candidate is an atomic structure x = (z, r), N (x) N (x) with atomic numbers z = (zi )i=1 and Cartesian coordinates r = (ri )i=1 , where N (x) is the number of atoms. The label y(x) associated with a structure consists of its total energy E(x) ∈ R and per-atom forces Fi (x) ∈ R3 , related by Fi (x) = −∇ri E(x), and is obtained from a reference DFT calculation. At each round t = 0, 1, . . . , T − 1, an acquisition function α scores candidates in P (t) and selects a batch A(t) = arg max α(A), (1) A⊆P (t) , |A|=B

where α may depend on the current model fθ(t) , the labelled set T (t) , or fixed structural descriptors of the candidates, depending on the acquisition method. Reference labels are then queried for the selected candidates, and the training set and pool are updated as T (t+1) = T (t) ∪ {(x, y(x)) : x ∈ A(t) }, 3

P (t+1) = P (t) \ A(t) .

(2)

The model is fine-tuned on T (t+1) before the next round. The goal is to choose acquisition batches that minimise test error after T rounds, given the fixed budget T · B. Our experiments use pretrained MACE models Batatia et al. [2023b] trained via the mlip library Brunken et al. [2025] via jax Bradbury et al. [2018]. The MACE architecture predicts total energies from input structures Eθ (x), and forces are obtained by auto-differentiation. 3.2

Acquisition signals

Each candidate label consists of a scalar energy and atomic forces, so an effective acquisition signal should reflect sensitivity to both quantities. The methods we compare fall into three families: model-based kernels derived from the MLIP (energy NTK, force NTK, energy–force NTK, and latent activations), descriptor-based kernels using fixed structural representations (SOAP Bartók et al. [2013] and Morgan fingerprints Morgan [1965], Rogers and Hahn [2010]), and direct uncertainty scores from committee disagreement, alongside random selection. Kernel-based methods construct a similarity geometry over the candidate pool and pair it with a batch selection rule (Section 3.3); direct methods assign point-wise scores and do not encode redundancy or coverage among candidates. Kernel methods are derived from an underlying molecular representation ϕ(x) ∈ Rd . For binary representations such as Morgan fingerprints, we use Tanimoto similarity Rogers and Hahn [2010]. For all other embeddings, we define kernels via inner products k(x, x′ ) = ϕ(x)⊤ ϕ(x′ ) of cosinenormalised embeddings. Energy NTK (NTK-E): Following prior work Zaverkin et al. [2022], Varga-Umbrich et al. [2026], we represent a structure by the sensitivity of the predicted energy to model parameters, ϕE (x) = ∇θ Eθ (x).

(3)

Computing ∇θ Eθ (x) for a single structure requires only a standard backward pass, but the resulting feature has dimension equal to the full parameter count, so storing the d × d feature-space precision matrix or materialising pairwise NTK entries quickly becomes memory-prohibitive. We therefore restrict the NTK to a parameter subspace θP , with the embedding parameter blocks of MACE used as the default choice (see Appendix B). Although these weights are species-indexed, ∇θP Eθ (x) is computed by backpropagating through every subsequent interaction and readout layer, so the NTK aggregates whole-model sensitivities into a feature that distinguishes structures by both composition and geometry. Parameter subsets trade-offs are compared in Section 4.4. Force NTK (NTK-F): NTK-E does not directly capture sensitivity to forces. To extend the NTK to vector-field prediction, we consider the mixed parameter-coordinate derivative JθP (x) = ∇θP F(x) = −∇θP ∇r Eθ (x),

(4)

Since JθP depends on the atomic positions through atomic forces, to get a structure P level embedding, we must aggregate over the atoms. For any translationally invariant force field, i Fi = 0, and so we instead aggregate via the root mean square p 1 ϕF (x) = p JθP (x) · JθP (x), (5) N (x) where the dot product is taken over coordinate space and not parameter space. The resulting embedding ϕF (x) is thus an equivariant molecular embedding with dimension equal to the number of parameters θP . Energy-force NTK (NTK-EF): Energy and force sensitivity need not rank candidates identically (see Appendix G.3 for a visual comparison of energy and force NTK kernels), and a general molecular embedding should be explicitly sensitive to both energy and force prediction. We therefore combine them into a single representation, √ √ ϕEF (x) = [ wE ϕE (x), wF ϕF (x)] , (6) where each component is normalized separately and wE , wF ≥ 0 control the relative contributions of energy and force signals. The corresponding kernel takes the form KEF (x, x′ ) = ϕEF (x)⊤ ϕEF (x′ ) = wE kE (x, x′ ) + wF kF (x, x′ ), 4

(7)

where kE (x, x′ ) = ϕE (x)⊤ ϕE (x′ ) and kF (x, x′ ) = ϕF (x)⊤ ϕF (x′ ). As a consequence, kernel-based acquisition in Equation (6) subsumes both energy and force NTKs as special cases (wF = 0 or wE = 0) and interpolates between them depending on which signal is most informative. Although the combined kernel is a linear sum, the contributions become coupled through conditioning on the training set: posterior-variance acquisition (Equation (8)) reflects how both signals interact given the observed data. Baselines: We include activation features Varga-Umbrich et al. [2026] as a learned representation baseline. As model-independent methods, we use SOAP descriptors Bartók et al. [2013], De et al. [2017] computed with DScribe Himanen et al. [2020] (rcut = 6.0 Å, nmax = 8, lmax = 6, with outer averaging), and Morgan fingerprints Morgan [1965], Rogers and Hahn [2010] (radius r = 3, 2048 bits) with the Tanimoto kernel Ralaivola et al. [2005] computed using RDKit Landrum et al. [2025]. For direct uncertainty methods, we use pretrained committees with M = 3 models initialised from the same checkpoint, with diversity introduced via independent data-order shuffling Varga-Umbrich et al. [2026]. We report energy disagreement (Committee-E) and force disagreement (Committee-F), obtained by aggregating per-atom force variance into a structure-level score. Committee variants are evaluated using independently fine-tuned models initialised from the same pretrained checkpoint. This requires training a total of M + 1 models. We note that potentially improved committee variants for pretrained atomistic foundation models, such as multi-head committees Beck et al. [2025], are not considered in this work. We additionally include random selection as a baseline. 3.3

Scalable batch selection via shortlisting

Given a kernel, a batch selection rule chooses B candidates per round. We use Gaussian posterior variance (PV) to shortlist candidates based on uncertainty, and Largest-Cluster Maximum Distance (LCMD) Holzmüller et al. [2023] to enforce diversity within the shortlist. To make this tractable for large pools, we compute PV in the feature space, then apply LCMD to a high-uncertainty shortlist. Posterior variance. Given a kernel k and current training set T (t) , a candidate x has Gaussian posterior variance σt2 (x) = k(x, x) − kT (t) (x)⊤ (KT (t) T (t) + λI)−1 kT (t) (x),

(8)

where kT (t) (x) is the vector of similarities between x and the training set. Direct evaluation requires the n2P candidate-candidate kernel, which is infeasible at nP ∼ 200k. Feature-space form. For kernels with explicit feature maps k(x, x′ ) = ϕ(x)⊤ ϕ(x′ ), posterior variance can be computed directly in feature space Holzmüller et al. [2023]. Let ΦT ∈ RnT ×d denote the feature matrix of the labelled set, and define the feature-space precision matrix −1 MT = (Φ⊤ ∈ Rd×d , T ΦT + λI)

(9)

By the Woodbury identity Golub [2012], Rasmussen and Williams [2005], the score s(x) = ϕ(x)⊤ MT ϕ(x),

(10)

is proportional to the Gaussian posterior variance and yields the same candidate ranking, while avoiding construction of the n2P candidate–candidate kernel. We compute MT by accumulating Φ⊤ T ΦT over chunks of training features, then inverting the resulting d × d matrix. Candidate scores s(x) are computed in chunks against the cached MT , while a running top-K shortlist tracks the highest-scoring candidates. This keeps peak memory at O(d2 + Cd) for chunk size C and makes the total acquisition cost linear in nP and nT . LCMD on the shortlist. Posterior variance alone tends to favour isolated high-uncertainty points. We therefore use PV only to shortlist the K highest-uncertainty candidates, then apply LCMD on the shortlist to select the final batch. LCMD modifies greedy farthest-point sampling to favour dense but under-covered regions, mitigating redundancy and discouraging outlier selection. This separation is deliberate: PV acts as an uncertainty-based filter with full conditioning on the training set, while LCMD enforces diversity. 5

In practice, we are able to use this method to efficiently score candidate pools of ∼ 200k points with shortlists of size K = 50k, but in theory, PV shortlisting extends linearly in the candidate and training pool, in principle unlocking much larger sets. Further details and pseudocode are given in Appendix D.

4

Results

Our experiments are designed to isolate the roles of force awareness, robustness to distribution shift, and scalability in offline active learning. We first evaluate our methods in a large-scale OC20 candidate-pool setting, where we show that force-aware acquisition becomes essential for strong performance. We then study reactive benchmarks in Section 4.2 to better understand under what conditions explicit force awareness improves acquisition beyond energy-only signals. In Section 4.3, we perform controlled candidate-pool biasing on T1x to evaluate robustness under distribution shift and study how different acquisition strategies respond to biased candidate pools. Finally, Section 4.4 analyses runtime and memory costs. All experiments run on a single H100 80GB GPU. Large-pool OC20 catalyst benchmark 21

26 24

0.175 0.170 0.165

22

30 28 26

E MAE / atom [meV]

0.20

F RMSE [eV/Å]

E RMSE / atom [meV]

19 18 17

0.114

16

0.19

0.18

24

Committee-F Activation PV

0.110 0.108 0.106

0.122 0.120

22

20

18

0.118 0.116 0.114 0.112 0.110

0.17

Random Committee-E

0.112

0.104

24

32

OOD

20

F MAE [eV/Å]

28

E MAE / atom [meV]

0.180

F RMSE [eV/Å]

E RMSE / atom [meV]

ID

30

F MAE [eV/Å]

4.1

Activation LCMD NTK-E PV

NTK-E LCMD NTK-EF LCMD

NTK-F LCMD

Figure 1: OC20 final-round test errors on val_is (ID) and the mean of the three out-of-distribution splits val_oos_ads, val_oos_bulk, and val_oos_ads_bulk. The joint energy–force NTK achieves the lowest energy and force errors across all panels. We consider active learning on a randomly selected 200k subset of OC20 Chanussot et al. [2021]. To handle the pool size we use feature-space PV to shortlist candidates, followed by LCMD selection (Section 3.2). OC20 candidate structures are drawn from relaxation trajectories near, but not at, local minima, where small geometric perturbations drive large force changes that energy similarity alone cannot resolve. We use a MACE model pretrained on MPTraj Deng et al. [2023]. Each run starts from 2,000 labelled structures and acquires 1,250 per round for six rounds, averaged over three seeds. Performance is evaluated on a 20k-structure test set partitioned evenly across the four OC20 regimes Chanussot et al. [2021]: in-distribution (val_is) and three out-of-distribution splits with held-out adsorbates, bulks, or both. We use a balanced weighting wE :wF = 1:1 for the NTK-EF. Figure 1 reports final-round test errors. The NTK-EF variant attains the lowest error on every metric and every split. NTK-F trails NTK-EF by a small but consistent margin. Forces are peratom quantities sensitive to local geometry; energies are extensive and reflect the full structure. 6

The joint kernel inherits both views, so candidates that match in local force pattern but differ in overall composition or size remain distinguishable. Activation-LCMD is competitive on val_is but degrades on the OOD splits. NTK-E degrades most sharply on val_oos_ads, where energy features alone are insufficient to distinguish adsorbate chemistry and the diversity objective ends up emphasising bulk-surface geometry rather than adsorbate variation. Its OOD energy MAE remains comparable to the other methods, while its OOD energy RMSE is the worst among non-random baselines; this gap indicates a small number of large per-structure errors rather than a broad shift in accuracy, consistent with energy-only PV occasionally selecting a few high-leverage outliers on the adsorbate-shift split. Both committee variants are the weakest non-random methods on every panel, indicating that pretrained-backbone ensemble disagreement, in either form, is not a competitive acquisition signal at this scale. This is consistent with prior reports that ensemble disagreement underestimates uncertainty under distribution shift Kahle and Zipoli [2022]. With the LCMD step removed, PV acquisition performs at or near the random baseline on every split (Figure 1), indicating that diversity-aware selection is required for the PV signal to translate into informative batches at this scale. Per-split breakdowns and full learning curves are in Appendix H. 4.2

When does force-awareness matter?

We now test whether force sensitivity improves acquisition across reactive benchmarks. Following the protocol of Varga-Umbrich et al. [2026], we evaluate on three reactivity datasets: HCNO-filtered PMechDB Tavakoli et al. [2024], RGD Zhao et al. [2023], and T1x Schreiner et al. [2022], using a SPICE-2-pretrained Eastman et al. [2023], Levine et al. [2026] MACE model. The PMechDB and RGD candidate pools are 10k and 5k random subsets of OMol Levine et al. [2026], and the T1x pool comprises 100 randomly selected reaction pathways. Each run starts from 50 labelled structures and acquires 150 per round for 20 rounds, the remaining structures are split 20%/10% test/validation. Results are averaged over two random seeds. Table 1: Energy and force acquisition performance across reactive benchmarks. Reported metrics are energy/force RMSE area under the acquisition curve (AUC) and final-round RMSE. Lower is better; best values in each row are bolded. Percent differences are relative to Random. Dataset

Metric Random

Activation

Committee-E

Committee-F

NTK-E

NTK-F

NTK-EF

RGD

E AUC F AUC Final E Final F

348.0 4763.5 13.0 188.5

326.0 (-6.3%) 339.0 (-2.6%) 334.0 (-4.0%) 339.0 (-2.6%) 4677.0 (-1.8%) 4659.0 (-2.2%) 4587.5 (-3.7%) 4651.5 (-2.4%) 12.0 (-7.7%) 12.0 (-7.7%) 11.5 (-11.5%) 12.0 (-7.7%) 185.0 (-1.9%) 186.0 (-1.3%) 182.5 (-3.2%) 186.0 (-1.3%)

338.0 (-2.9%) 4665.5 (-2.1%) 12.0 (-7.7%) 184.0 (-2.4%)

344.5 (-1.0%) 4684.0 (-1.7%) 12.0 (-7.7%) 186.0 (-1.3%)

T1X

E AUC F AUC Final E Final F

246.5 2300.0 6.5 62.0

164.0 (-33.5%) 268.5 (+8.9%) 250.0 (+1.4%) 159.5 (-35.3%) 178.5 (-27.6%) 175.0 (-29.0%) 1789.0 (-22.2%) 2516.0 (+9.4%) 2182.5 (-5.1%) 1754.5 (-23.7%) 1802.5 (-21.6%) 1796.5 (-21.9%) 5.0 (-23.1%) 5.5 (-15.4%) 5.0 (-23.1%) 5.0 (-23.1%) 5.5 (-15.4%) 5.0 (-23.1%) 57.0 (-8.1%) 60.0 (-3.2%) 57.0 (-8.1%) 57.0 (-8.1%) 59.5 (-4.0%) 60.5 (-2.4%)

PMechDB

E AUC F AUC Final E Final F

453.0 2857.0 17.0 113.5

427.0 (-5.7%) 493.0 (+8.8%) 457.5 (+1.0%) 426.5 (-5.9%) 2800.5 (-2.0%) 2799.0 (-2.0%) 2634.5 (-7.8%) 2718.5 (-4.9%) 16.0 (-5.9%) 19.0 (+11.8%) 17.0 (0.0%) 16.0 (-5.9%) 109.0 (-4.0%) 107.0 (-5.7%) 102.0 (-10.1%) 106.5 (-6.2%)

433.5 (-4.3%) 2780.5 (-2.7%) 16.0 (-5.9%) 111.5 (-1.8%)

431.0 (-4.9%) 2777.0 (-2.8%) 16.5 (-2.9%) 109.5 (-3.5%)

Table 1 summarises the energy and force RMSE area under the (learning) curve (AUC), together with the final-round energy and force RMSEs across the three datasets for a range of baselines (see Appendix E for additional comparisons against fixed-descriptor methods). All four modelbased kernels (activation, NTK-E, NTK-F, NTK-EF) substantially outperform random selection, but differences between them are dataset-dependent and minute on most metrics (see Appendix E). Among these methods, activation features and energy NTKs provide the strongest and most consistent baselines overall. Committee-based methods, especially force committees, can perform extremely well (best in class) in some settings, but exhibit substantially higher variance across datasets and across the energy–force trade-off. Appendix E shows that committee MAE performance is well below random baselines, suggesting that a portion of the RMSE gains may arise from preferential selection of outlier or high-error structures. Across model-based methods, we find that improvements over random are generally smaller for MAE than for RMSE but still positive. Overall, these results indicate that, on the reactive benchmarks considered here, energy-based acquisition already captures much of the informative variation in the candidate pool. To support this, in Appendix E we find that increasing the wE : wF ratio gives better results. Importantly, 7

however, the resulting degradation is very small: force-only and joint energy–force NTKs continue to closely track the performance of the energy-only NTK across datasets, while remaining substantially better than random selection. More generally, the benefit of force-aware acquisition could depend on how much additional discrimination forces provide beyond energies, which is primarily a dataset property: within a reactive pathway, energies and forces co-vary and force features add little; across heterogeneous chemistries they carry information that energy similarities cannot recover. A second factor is force-label quality: Kuryla et al. [2025] report systematic force-component errors on molecular datasets, and force acquisition signals therefore operate on a noisier label stream than the corresponding energy channel. Improved force-aware uncertainty quantification may be achievable also through alternative normalisations of the energy and force representations (see Appendix G.3); we leave this for future work. 4.3

Robustness to candidate pool shift

While the robustness of acquisition is implicitly evaluated by evaluating AL performance across a variety of datasets, we further study robustness to candidate pool distribution shift in a controlled setting: this is relevant since real candidate pools are biased due to the inherent biases of the structure generators. T1x Schreiner et al. [2022] provides a natural setting for studying pool shift: each reaction family contains structures along NEB and climbing-image NEB paths, so the frame index serves as a proxy position along the reaction coordinate. This allows us to simulate two forms of bias: (i) inter-reaction imbalance, which skews sampling across reactions, and (ii) intra-reaction imbalance, which biases sampling along the reaction coordinate. Across both settings, we find that representation-based acquisition methods, including both learned model-derived representations and fixed descriptors, generally exhibit more stable behaviour under candidate-pool bias than uncertainty-only approaches, particularly under intra-reaction bias along the reaction coordinate. Combined with the results on OC20 and the reactive benchmarks, where committee-based methods exhibit higher variance across datasets and objectives, these findings suggest that representation-aware acquisition strategies can provide more robust behaviour under candidate-pool shift than those using only point-wise uncertainty estimates. Full experimental details and results are provided in Appendix G. 4.4

Cost, scalability, and operating range

The dominant cost in NTK-based acquisition is the feature-space PV pass over the candidate pool; LCMD selection on the resulting shortlist contributes negligible additional overhead. Two design choices determine the operating point: the parameter subset θP , which fixes the embedding dimension d, and whether the kernel encodes energy, force, or both. Figures (a,b) report the cost–accuracy frontier on OC20 across both error axes for energy and force error vs our experiments wall clock time; this additionally includes OC20 runs of the readout NTK variants (see Appendix B). Full per-method runtime and memory breakdowns are reported in Appendix I. To further characterise the θP trade-off in isolation, in Appendix I.3 we sweep four MACE parameter blocks (readouts, embeddings, interactions, last-layer) on a 5-reaction T1x subset. All NTK subsets achieve similar accuracy, and we find that larger subsets (interaction, last-layer) provide no measurable accuracy gain. A further cost constraint is memory during acquisition. A naive non-chunked PV implementation must materialise the kernel matrix K ∈ RnS ×nS , where nS = nP + nT , and so grows quadratically in the pool size, whereas chunked feature-space PV maintains only the d × d precision matrix and a small chunk buffer (Figure 2c). The bottleneck therefore shifts from pool size to feature dimension d, which restricts NTK acquisition to lightweight parameter subsets such as readouts or embeddings. Moving the bottleneck off pool size matters beyond active learning itself: dataset curation and distillation from million-scale atomistic datasets Levine et al. [2026], Sahoo et al. [2025] require acquisition signals that scale at most linearly in the candidate set.

5

Conclusion

We address three practical challenges for offline active learning of MLIPs: scaling to large candidate pools, incorporating force-aware acquisition, and maintaining robustness under candidate-pool bias. 8

Random Activation

Committee-E Committee-F

NTK-E NTK-E (reado.)

NTK-F NTK-F (reado.)

NTK-EF NTK-EF (reado.)

25.0 22.5 20.0 0

100

200

Wall-clock / round [min]

(a) Energy MAE vs cost.

Memory [GiB]

27.5

F MAE [eV/Å]

E MAE / atom [meV]

60 0.125 0.120 0.115 0.110

0

100

200

Wall-clock / round [min]

(b) Force MAE vs cost.

chunked PV non-chunked PV

40

20

0

10

3

10

4

10

5

Pool size nS = nP + nT

(c) Memory vs. pool size.

Figure 2: Cost, accuracy, and memory scaling of feature-space acquisition on OC20. (a,b) Mean per-round wall-clock vs. test energy MAE per atom and force MAE at final round. All kernel methods use LCMD selection on a PV shortlist and (reado.) denotes NTK variants on the readouts parameter subset (default is embeddings). The shaded green wedge marks the Pareto region defined through Random and NTK-EF (reado.). Committee acquisition implementations are not fully optimised, but this does not change the qualitative comparison; we evaluate committees using independently fine-tuned models initialised from the same pretrained checkpoint, requiring M + 1 models in total (including a separate evaluation model); excluding this would reduce the reported wall-clock by a factor 4/3. In addition, each ensemble member independently rebuilds candidate-pool graphs and executes sequential forward passes, while both energies and forces are computed for all committee variants regardless of the acquisition objective. (c) Acquisition peak memory of chunked vs. nonchunked feature-space PV: the non-chunked baseline grows quadratically, whereas the chunked path grows linearly until the tested nS ≈ 2×105 . Our results indicate that these challenges are coupled: scalability widens the operating regime in which the other two matter; force awareness becomes essential as pools grow more chemically heterogeneous; and bias robustness determines whether the resulting acquisition signal remains effective when the candidate pool diverges from the test distribution. We introduce force-aware NTK based on mixed parameter-coordinate derivatives, together with a joint energy-force NTK that subsumes energy-only and force-only acquisition as special cases. On OC20, the joint kernel is the only method that performs consistently across all in- and out-ofdistribution splits, demonstrating that energy-only acquisition is insufficient when force accuracy governs downstream behaviour. On the reactive benchmarks, force-aware variants remain competitive with energy-only baselines but do not uniformly improve over them; characterising the regimes in which force information contributes to acquisition signal energy sensitivity remains an open question. We further show that model-based acquisition can be scaled to candidate pools of ∼ 200k structures with linear complexity in both training-set and pool size via the chunked shortlist-based pipeline. This shifts the dominant computational bottleneck from pool size to feature dimension d, which restricts practical NTK acquisition to lightweight parameter subsets such as embeddings and readouts. Within this regime, the resulting methods are both more accurate and faster than committee-based baselines, since the per-round cost of retraining an ensemble exceeds that of computing feature representations through a single forward and backward pass over the candidate pool. Nevertheless, model-based acquisition is expensive at million-scale candidate sets, where runtime is still dominated by repeated feature extraction over the pool, and fixed-descriptor approaches may offer a cheaper alternative. Further limitations are discussed in Section A. We additionally study robustness to candidate-pool bias in a controlled setting on T1x, where reactionfamily and reaction-coordinate sampling weights can be varied independently. Across both forms of bias, model-derived representations (both NTK and pretrained activation kernels) produce more stable acquisition behaviour than committee-based disagreement, indicating that an explicit similarity geometry over the pool can dampen the impact of generator-induced bias more effectively than point-wise uncertainty estimates. There are many directions that follow from this work. The joint energy–force NTK provides a general construction in which physically distinct prediction targets are combined into a single tunable 9

acquisition kernel. A natural extension would be to incorporate additional supervision signals such as stress, Hessians, and partial charges. At the same time, many aspects of the construction remain heuristic, including how best to combine and normalise different representations. More broadly, the offline acquisition setting connects naturally to dataset distillation: the chunked feature-space pipeline introduced here can score and prune candidates from large-scale atomistic datasets, providing a route toward model-aware curation of pretraining and fine-tuning corpora.

Acknowledgements We thank Silvia Acosta Gutiérrez, Massimo Bortone, Heloise Chomet, Valentin Heyraud, Jack Simons, Tamás Lajos Tompa, Lucien Walewski, and Leon Wehrhan for insightful discussions. We also thank Scott Cameron for highlighting the potential usefulness of Neural Tangent Kernels for uncertainty quantification.

References Dylan M. Anstine, Roman Zubatyuk, and Olexandr Isayev. Data generation for machine learning interatomic potentials and beyond. Chemical Reviews, 2025. doi: 10.1021/acs.chemrev.4c00572. Albert P. Bartók, Risi Kondor, and Gábor Csányi. On representing chemical environments. Physical Review B, 87(18):184115, 2013. doi: 10.1103/PhysRevB.87.184115. Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. Elena, Dávid P. Kovács, Janosh Riebesell, Xavier R. Advincula, Mark Asta, Matthew Avaylon, et al. A foundation model for atomistic materials chemistry, 2023a. URL https://arxiv.org/abs/2401.00096. Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields, 2023b. URL https://arxiv.org/abs/2206.07697. Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications, 13:2453, 2022. doi: 10.1038/s41467-022-29939-5. Hubert Beck, Pavol Simko, Lars L. Schaaf, Ondrej Marsalek, and Christoph Schran. Multi-head committees enable direct uncertainty prediction for atomistic foundation models. The Journal of Chemical Physics, 163(23):234103, 2025. doi: 10.1063/5.0288994. Jörg Behler and Michele Parrinello. Generalized neural-network representation of high-dimensional potential-energy surfaces. Physical Review Letters, 98(14):146401, 2007. doi: 10.1103/ PhysRevLett.98.146401. James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Yash Katariya, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. URL http://github.com/jax-ml/jax. Christoph Brunken, Olivier Peltre, Heloise Chomet, Lucien Walewski, Manus McAuliffe, Valentin Heyraud, Solal Attias, Martin Maarand, Yessine Khanfir, Edan Toledo, Fabio Falcioni, Marie Bluntzer, Silvia Acosta-Gutiérrez, and Jules Tilly. Machine learning interatomic potentials: library for efficient training, model development and simulation of molecular systems, 2025. URL https://arxiv.org/abs/2505.22397. Lowik Chanussot, Abhishek Das, Siddharth Goyal, Thibaut Lavril, Muhammed Shuaibi, Morgane Riviere, Kevin Tran, Javier Heras-Domingo, Caleb Ho, Weihua Hu, Aini Palizhati, Anuroop Sriram, Brandon Wood, Junwoong Yoon, Devi Parikh, C. Lawrence Zitnick, and Zachary Ulissi. Open catalyst 2020 (oc20) dataset and community challenges. ACS Catalysis, 11(10):6059–6072, May 2021. ISSN 2155-5435. doi: 10.1021/acscatal.0c04525. URL http://dx.doi.org/10. 1021/acscatal.0c04525. 10

Dario Coscia, Pim de Haan, and Max Welling. Blips: Bayesian learned interatomic potentials, 2026. URL https://arxiv.org/abs/2508.14022. Taoyong Cui, Chenyu Tang, Dongzhan Zhou, Longtao Wang, Yu Zheng, Yu Wang, Liang Wang, Weizhi Yang, Lei Bai, and Wanli Ouyang. Online test-time adaptation for better generalization of interatomic potentials to out-of-distribution data. Nature Communications, 16:1891, 2025. doi: 10.1038/s41467-025-57101-4. Sandip De, Albert P. Bartók, Gábor Csányi, and Michele Ceriotti. Machine learning unifies the modeling of materials and molecules. Science Advances, 3(12):e1701816, 2017. doi: 10.1126/ sciadv.1701816. Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J. Bartel, and Gerbrand Ceder. CHGNet as a pretrained universal neural network potential for chargeinformed atomistic modelling. Nature Machine Intelligence, 5:1031–1041, 2023. doi: 10.1038/ s42256-023-00716-3. Bowen Deng, Yunyeong Choi, Peichen Zhong, Janosh Riebesell, Shashwat Anand, Zhuoying Li, KyuJung Jun, Kristin A. Persson, and Gerbrand Ceder. Systematic softening in universal machine learning interatomic potentials. npj Computational Materials, 11:9, 2025. doi: 10.1038/s41524-024-01500-6. Peter Eastman, Pavan Kumar Behara, David L. Dotson, Raimondas Galvelis, John E. Herr, Josh T. Horton, Yuezhi Mao, John D. Chodera, Benjamin P. Pritchard, Yuanqing Wang, Gianni De Fabritiis, and Thomas E. Markland. Spice, a dataset of drug-like molecules and peptides for training machine learning potentials. Scientific Data, 10(1), January 2023. ISSN 2052-4463. doi: 10.1038/ s41597-022-01882-6. URL http://dx.doi.org/10.1038/s41597-022-01882-6. Gene Golub. Matrix Computations. Johns Hopkins University Press, 2012. ISBN 9781421407944. doi: 10.56021/9781421407944. URL http://dx.doi.org/10.56021/9781421407944. Graeme Henkelman and Hannes Jónsson. Improved tangent estimate in the nudged elastic band method for finding minimum energy paths and saddle points. The Journal of Chemical Physics, 113(22):9978–9985, December 2000. ISSN 1089-7690. doi: 10.1063/1.1323224. URL http: //dx.doi.org/10.1063/1.1323224. Lauri Himanen, Marc O. J. Jäger, Eiaki V. Morooka, Filippo Federici Canova, Yashasvi S. Ranawat, David Z. Gao, Patrick Rinke, and Adam S. Foster. DScribe: Library of descriptors for machine learning in materials science. Computer Physics Communications, 247:106949, 2020. ISSN 0010-4655. doi: 10.1016/j.cpc.2019.106949. URL https://doi.org/10.1016/j.cpc.2019. 106949. David Holzmüller, Viktor Zaverkin, Johannes Kästner, and Ingo Steinwart. A framework and benchmark for deep batch active learning for regression. Journal of Machine Learning Research, 24(164):1–81, 2023. URL https://www.jmlr.org/papers/v24/22-0937.html. Matthew K. Horton, Patrick Huck, Ruo Xi Yang, Jason M. Munro, Shyam Dwaraknath, Alex M. Ganose, Ryan S. Kingsbury, Mingjian Wen, Jimmy X. Shen, Tyler S. Mathis, Aaron D. Kaplan, Karlo Berket, Janosh Riebesell, Janine George, Andrew S. Rosen, Evan W. C. Spotte-Smith, Matthew J. McDermott, Orion A. Cohen, Alex Dunn, Matthew C. Kuner, Gian-Marco Rignanese, Guido Petretto, David Waroquiers, Sinead M. Griffin, Jeffrey B. Neaton, Daryl C. Chrzan, Mark Asta, Geoffroy Hautier, Shreyas Cholia, Gerbrand Ceder, Shyue Ping Ong, Anubhav Jain, and Kristin A. Persson. Accelerated data-driven materials science with the materials project. Nature Materials, 24(10):1522–1532, 2025. ISSN 1476-4660. doi: 10.1038/s41563-025-02272-0. URL http://dx.doi.org/10.1038/s41563-025-02272-0. Ryan Jacobs, Dane Morgan, Siamak Attarian, Jun Meng, Chen Shen, Zhenghao Wu, Clare Yijia Xie, Julia H. Yang, Nongnuch Artrith, Ben Blaiszik, Gerbrand Ceder, Kamal Choudhary, Gabor Csanyi, Ekin Dogus Cubuk, Bowen Deng, Ralf Drautz, Xiang Fu, Jonathan Godwin, Vasant Honavar, Olexandr Isayev, Anders Johansson, Boris Kozinsky, Stefano Martiniani, Shyue Ping Ong, Igor Poltavsky, KJ Schmidt, So Takamoto, Aidan P. Thompson, Julia Westermayr, and Brandon M. Wood. A practical guide to machine learning interatomic potentials – status and future. Current Opinion in Solid State and Materials Science, 35:101214, March 2025. ISSN 1359-0286. doi: 10. 1016/j.cossms.2025.101214. URL http://dx.doi.org/10.1016/j.cossms.2025.101214. 11

Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. In Advances in Neural Information Processing Systems, 2018. URL https://arxiv.org/abs/1806.07572. Ryosuke Jinnouchi, Ferenc Karsai, and Georg Kresse. On-the-fly machine learning force field generation: Application to melting points. Physical Review B, 100(1):014105, 2019. doi: 10. 1103/PhysRevB.100.014105. Hannes Jónsson, Greg Mills, and Karsten W. Jacobsen. Nudged elastic band method for finding minimum energy paths of transitions. In Classical and Quantum Dynamics in Condensed Phase Simulations, pages 385–404. World Scientific, June 1998. doi: 10.1142/9789812839664_0016. URL http://dx.doi.org/10.1142/9789812839664_0016. Leonid Kahle and Federico Zipoli. Quality of uncertainty estimates from neural network potential ensembles. Physical Review E, 105(1), January 2022. ISSN 2470-0053. doi: 10.1103/physreve. 105.015311. URL http://dx.doi.org/10.1103/PhysRevE.105.015311. Matthias Kellner and Michele Ceriotti. Uncertainty quantification by direct propagation of shallow ensembles. Machine Learning: Science and Technology, 5(3):035006, 2024. doi: 10.1088/ 2632-2153/ad594a. Matthias Kellner and Michele Ceriotti. Enhanced representation-based sampling for the efficient generation of data sets for machine-learned interatomic potentials. Journal of Chemical Theory and Computation, 2025. doi: 10.1021/acs.jctc.5c01767. Maksim Kulichenko, Kipton Barros, Nicholas Lubbers, Ying Wai Li, Richard Messerly, Sergei Tretiak, Justin S. Smith, and Benjamin Nebgen. Uncertainty-driven dynamics for active learning of interatomic potentials. Nature Computational Science, 3:230–239, 2023. doi: 10.1038/s43588-023-00406-5. Domantas Kuryla, Fabian Berger, Gábor Csányi, and Angelos Michaelides. How accurate are dft forces? unexpectedly large uncertainties in molecular datasets, 2025. URL https://arxiv.org/ abs/2510.19774. Alessandro Laio and Michele Parrinello. Escaping free-energy minima. Proceedings of the National Academy of Sciences, 99(20):12562–12566, September 2002. ISSN 1091-6490. doi: 10.1073/ pnas.202427399. URL http://dx.doi.org/10.1073/pnas.202427399. Greg Landrum et al. RDKit: Open-source cheminformatics Software, 2025. URL https://doi. org/10.5281/zenodo.17232453. Version 2025_09_1. Daniel S. Levine, Muhammed Shuaibi, Evan Walter Clark Spotte-Smith, Michael G. Taylor, Muhammad R. Hasyim, Kyle Michel, Ilyes Batatia, Gábor Csányi, Misko Dzamba, Peter Eastman, Nathan C. Frey, Xiang Fu, Vahe Gharakhanyan, Aditi S. Krishnapriyan, Joshua A. Rackers, Sanjeev Raja, Ammar Rizvi, Andrew S. Rosen, Zachary Ulissi, Santiago Vargas, C. Lawrence Zitnick, Samuel M. Blau, and Brandon M. Wood. The open molecules 2025 (omol25) dataset, evaluations, and models, 2026. URL https://arxiv.org/abs/2505.08762. Yifan Li, Xiuying Zhang, Mingkang Liu, and Lei Shen. A critical review of machine learning interatomic potentials and hamiltonian. Journal of Materials Informatics, 5(4), 2025. ISSN 2770-372X. doi: 10.20517/jmi.2025.17. URL https://www.oaepublish.com/articles/ jmi.2025.17. H. L. Morgan. The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service. Journal of Chemical Documentation, 5(2):107–113, May 1965. ISSN 1541-5732. doi: 10.1021/c160017a018. URL http://dx.doi.org/10.1021/ c160017a018. Mark Neumann, James Gin, Benjamin Rhodes, Steven Bennett, Zhiyi Li, Hitarth Choubisa, Arthur Hussey, and Jonathan Godwin. Orb-v3: atomistic simulation at scale, 2025. URL https: //arxiv.org/abs/2504.06231. 12

Samuel P. Niblett, Panagiotis Kourtis, Ioan-Bogdan Magdău, Clare P. Grey, and Gábor Csányi. Transferability of data sets between machine-learned interatomic potential algorithms. Journal of Chemical Theory and Computation, 21(12):6096–6112, June 2025. ISSN 1549-9626. doi: 10.1021/acs.jctc.5c00272. URL http://dx.doi.org/10.1021/acs.jctc.5c00272. Xinjian Ouyang, Zhilong Wang, Xiao Jie, Feng Zhang, Yanxing Zhang, Laijun Liu, and Dawei Wang. Latent space active learning with message passing neural network: The case of hfo2 . Physical Review Materials, 8(10):103804, October 2024. ISSN 2475-9953. doi: 10.1103/PhysRevMaterials. 8.103804. URL https://doi.org/10.1103/PhysRevMaterials.8.103804. Andrew A. Peterson, Rune Christensen, and Alireza Khorshidi. Addressing uncertainty in atomistic machine learning. Physical Chemistry Chemical Physics, 19:10978–10985, 2017. doi: 10.1039/ C7CP00375G. Evgeny V. Podryabinkin and Alexander V. Shapeev. Active learning of linearly parametrized interatomic potentials. Computational Materials Science, 140:171–180, 2017. doi: 10.1016/j. commatsci.2017.08.031. Liva Ralaivola, Sanjay J. Swamidass, Hiroto Saigo, and Pierre Baldi. Graph kernels for chemical informatics. Neural Networks, 18(8):1093–1110, October 2005. ISSN 0893-6080. doi: 10.1016/j. neunet.2005.07.009. URL http://dx.doi.org/10.1016/j.neunet.2005.07.009. Carl Edward Rasmussen and Christopher K. I. Williams. Gaussian Processes for Machine Learning. The MIT Press, November 2005. ISBN 9780262256834. doi: 10.7551/mitpress/3206.001.0001. URL http://dx.doi.org/10.7551/mitpress/3206.001.0001. David Rogers and Mathew Hahn. Extended-connectivity fingerprints. Journal of Chemical Information and Modeling, 50(5):742–754, 2010. doi: 10.1021/ci100050t. Sushree Jagriti Sahoo, Mikael Maraschin, Daniel S. Levine, Zachary Ulissi, C. Lawrence Zitnick, Joel B Varley, Joseph A. Gauthier, Nitish Govindarajan, and Muhammed Shuaibi. The open catalyst 2025 (oc25) dataset and models for solid-liquid interfaces, 2025. URL https://arxiv. org/abs/2509.17862. Christoph Schran, Krystof Brezina, and Ondrej Marsalek. Committee neural network potentials control generalization errors and enable active learning. The Journal of Chemical Physics, 153 (10):104105, 2020. doi: 10.1063/5.0016004. Mathias Schreiner, Arghya Bhowmik, Tejs Vegge, Jonas Busk, and Ole Winther. Transition1x a dataset for building generalizable reactive machine learning potentials. Scientific Data, 9(1), December 2022. ISSN 2052-4463. doi: 10.1038/s41597-022-01870-w. URL http://dx.doi. org/10.1038/s41597-022-01870-w. Kristof T. Schütt, Huziel E. Sauceda, Pieter-Jan Kindermans, Alexandre Tkatchenko, and KlausRobert Müller. SchNet – a deep learning architecture for molecules and materials. The Journal of Chemical Physics, 148(24):241722, 2018. doi: 10.1063/1.5019779. Burr Settles. Active Learning. Springer International Publishing, 2012. ISBN 9783031015601. doi: 10.1007/978-3-031-01560-1. URL http://dx.doi.org/10.1007/978-3-031-01560-1. Justin S. Smith, Ben Nebgen, Nicholas Lubbers, Olexandr Isayev, and Adrian E. Roitberg. Less is more: Sampling chemical space with active learning. The Journal of Chemical Physics, 148(24): 241733, 2018. doi: 10.1063/1.5023802. Mohammadamin Tavakoli, Ryan J. Miller, Mirana Claire Angel, Michael A. Pfeiffer, Eugene S. Gutman, Aaron D. Mood, David Van Vranken, and Pierre Baldi. Pmechdb: A public database of elementary polar reaction steps. Journal of Chemical Information and Modeling, 64(6):1975–1983, March 2024. ISSN 1549-960X. doi: 10.1021/acs.jcim.3c01810. URL http://dx.doi.org/10. 1021/acs.jcim.3c01810. Jonathan Vandermause, Steven B. Torrisi, Simon Batzner, Yu Xie, Lixin Sun, Alexie M. Kolpak, and Boris Kozinsky. On-the-fly active learning of interpretable Bayesian force fields for atomistic rare events. npj Computational Materials, 6:20, 2020. doi: 10.1038/s41524-020-0283-z. 13

Eszter Varga-Umbrich, Shikha Surana, Paul Duckworth, Jules Tilly, Olivier Peltre, and Zachary Weller-Davies. Pretrained model representations as acquisition signals for active learning of mlips, 2026. URL https://arxiv.org/abs/2605.03964. Guanjie Wang, Changrui Wang, Xuanguang Zhang, Zefeng Li, Jian Zhou, and Zhimei Sun. Machine learning interatomic potential: Bridge the gap between small-scale models and realistic devicescale simulations. iScience, 27(5):109673, 2024. ISSN 2589-0042. doi: https://doi.org/10. 1016/j.isci.2024.109673. URL https://www.sciencedirect.com/science/article/pii/ S2589004224008952. Brandon M. Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John R. Kitchin, Daniel S. Levine, Kyle Michel, Anuroop Sriram, Taco Cohen, Abhishek Das, Ammar Rizvi, Sushree Jagriti Sahoo, Zachary W. Ulissi, and C. Lawrence Zitnick. Uma: A family of universal models for atoms, 2026. URL https://arxiv.org/abs/2506.23971. Viktor Zaverkin, David Holzmüller, Ingo Steinwart, and Johannes Kästner. Exploring chemical and conformational spaces by batch mode deep active learning. Digital Discovery, 1:605–620, 2022. doi: 10.1039/D2DD00034B. Viktor Zaverkin, David Holzmüller, Henrik Christiansen, Federico Errica, Francesco Alesiani, Makoto Takamoto, Mathias Niepert, and Johannes Kästner. Uncertainty-biased molecular dynamics for learning uniformly accurate interatomic potentials. npj Computational Materials, 10:83, 2024. doi: 10.1038/s41524-024-01254-1. Viktor Zaverkin, David Holzmüller, and Johannes Kästner. Active learning meets metadynamics: automated workflow for reactive machine learning interatomic potentials. Digital Discovery, 2026. doi: 10.1039/D5DD00261C. Linfeng Zhang, De-Ye Lin, Han Wang, Roberto Car, and Weinan E. Active learning of uniformly accurate interatomic potentials for materials simulation. Physical Review Materials, 3(2):023804, 2019. doi: 10.1103/PhysRevMaterials.3.023804. Qiyuan Zhao, Sai Mahit Vaddadi, Michael Woulfe, Lawal A. Ogunfowora, Sanjay S. Garimella, Olexandr Isayev, and Brett M. Savoie. Comprehensive exploration of graphically defined reaction spaces. Scientific Data, 10(1), March 2023. ISSN 2052-4463. doi: 10.1038/s41597-023-02043-z. URL http://dx.doi.org/10.1038/s41597-023-02043-z. Joanna Zou and Youssef Marzouk. Data curation for machine learning interatomic potentials by determinantal point processes, 2026. URL https://arxiv.org/abs/2603.22160.

14

A

Limitations

In this section, we outline further limitations of our work. Despite the scalability improvements introduced by the chunked feature-space acquisition pipeline, the cost of NTK-based acquisition remains sensitive to the feature dimension d, which restricts practical use to relatively lightweight parameter subsets. Furthermore, the overall runtime is still dominated by repeated feature extraction passes over the candidate pool, which may still become challenging at million-scale candidate sets. Our force-aware NTK construction also depends on specific aggregation and normalisation choices for atomic force Jacobians and the energy–force mixing parameters. The relative weighting between energy and force contributions remains heuristic and dataset-dependent. The kernel visualisations in Appendix G.3 further suggest that the force contribution can dominate the joint representation, indicating that more sophisticated normalisation or balancing strategies may improve uncertainty estimation. We focus on offline pool-based active learning and do not evaluate online active-learning workflows in which candidate generation and acquisition are coupled. In addition, all experiments are performed using MACE architectures, and it remains unclear how well the proposed force-aware NTK constructions transfer to other equivariant networks Schütt et al. [2018], Batzner et al. [2022], Wood et al. [2026], Neumann et al. [2025]. While we compare against standard committee-based uncertainty estimation, we do not evaluate several recently proposed committee variants designed specifically for pretrained atomistic foundation models Beck et al. [2025]. Additionally, the computational cost of the committee baselines could likely be reduced through more efficient training strategies such as sequential fine-tuning or parameter sharing across ensemble members.

B

MACE Architecture and Representation Extraction N (x)

We write a structure as x = (z, r), with atomic numbers z = (zi )i=1 and Cartesian coordinates N (x) r = (ri )i=1 . MACE Batatia et al. [2023b] constructs a radius-cutoff graph G(x): nodes are atoms, and directed edges connect neighbours within cutoff rmax . Node attributes are one-hot species vectors ai . For an edge j → i, define the relative vector rji = rj − ri , distance dji = ∥rji ∥, and direction r̂ji = rji /dji . We now outline the relevant parameter subsets of MACE that are used for the NTK and activation representations in the main text. B.1

Relevant MACE parameters

The embedding stage has two parts: an atomic (node) embedding block and a radial (edge) embedding block. Atomic embedding weights. The atomic embedding block maps the one-hot species attribute ai of each atom to an initial scalar node feature. Concretely, it is a learnable lookup table (0)

W emb ∈ RZmax ×c ,

hi,L=0 = Wzemb , i, :

(11)

where Zmax is the number of supported atomic species and c is the number of scalar channels. The initial node feature of atom i is therefore the row of W emb indexed by its atomic number zi : it depends only on the atomic species, not on the local environment, the geometry, or the rest of the structure. Two atoms of the same species in different configurations enter the network with identical initial features, and differences in their downstream representations arise entirely from the geometric information injected by the interaction blocks. Equivalently, the embedding weights act as a per-species learnable bias that carries the chemical identity of the atom into the rest of the model. Radial embedding. The radial embedding block maps each distance dji to edge features through a Bessel basis modulated by a smooth cutoff envelope. In the MACE checkpoints used in this work this basis is parameter-free, so the trainable parameters of the embedding stage are W emb . Angular 15

information is injected separately, and not at the embedding stage, through spherical harmonics of r̂ji . Interaction block parameters. The interaction blocks in MACE construct equivariant messages through Clebsch–Gordan tensor products between neighbour node features and spherical harmonics on the edge. Following Batatia et al. [2023b], messages along an edge j → i take the form X X X (ℓ) (ℓ) (ℓ) (ℓ) m3 m2 (12) mi,kl3 m3 = Cll13m R (r )Y (r̂ ) Wkk′ l2 hj,k′ l2 m2 . ji ji ,l m kl l l l 1 2 2 1 2 3 2 l1 ,m1 ,l2 ,m2

k′

j∈N (i) (t)

m3 where Cll13m are fixed Clebsch–Gordan coefficients, Rkl1 l2 l3 (rji ) are learned radial interaction 1 ,l2 m2 (ℓ)

weights constructed from the radial embedding, and Wkk′ l2 is a weight matrix. (t)

In practice, the interaction Rkl1 l2 l3 (rji ) are parameterized by MLPs acting on the radial basis embedding of the interatomic distance rji . The interaction parameter subset used in our NTK ablations consists of these learned radial tensor-product weights together with the surrounding equivariant linear maps. Interaction stack. Interaction blocks combine neighbour node features with radial and angular edge features to produce equivariant messages, which are then passed through symmetric tensor products aimed at increasing body order. After the interaction layer ℓ, the hidden state of atom i can be written as L max M (ℓ) (ℓ) hi = hi,L , (13) L=0

where L = 0 channels are invariant scalars and L > 0 channels transform as vectors or higher-order tensors under rotations. Readout parameters. After each interaction layer, MACE maps the invariant (L = 0) node features to per-site energy contributions through equivariant readout functions, the contribution from interaction layer t is P (t) (t) t < T,    k̃ Wreadout,k̃ hi,k̃00 , (t) (t) n o  = Ei = Rt hi (14) (t) MLP(t) h , t = T. readout i,k̃00 k̃

(t) Here hi,k̃00 denotes the invariant scalar channels of the equivariant node representation. The total

energy then has the form Eθ (x) =

N (x) T XX

(ℓ)

Ei (x),

(15)

i=1 ℓ=0

and forces are analytic coordinate gradients, Fi (x) = −∇ri Eθ (x).

(16) (t)

The readout parameters therefore consist of the linear projection weights Wreadout,k̃ used in intermediate layers together with the final MLP readout applied at the last interaction layer. Since the readout depends only on invariant channels, the resulting site energies are rotationally invariant. B.2

Representation extraction.

We extract two classes of representations from the pretrained MACE model: activation features and NTK features. For both force and energy NTK representations, we differentiate the predicted energy (or force) with respect to a chosen parameter subset θP . In this work, we consider three choices of θP : The embedding NTK uses θP = W emb , so its features measure the sensitivity of the energy to the learned species embeddings. Although these parameters are species-indexed, the gradients are 16

backpropagated through the full interaction stack and therefore depend on both composition and geometry. The interaction NTK uses the parameters of the equivariant interaction blocks, including the learned (ℓ) radial tensor-product weights Rkl1 l2 l3 and the associated equivariant linear maps. This subset captures sensitivity to the message-passing operations that couple local geometry, angular information, and neighbour features. The readout NTK uses the parameters of the scalar readout functions, including the linear readout (t) weights Wreadout,k̃ and the final readout MLP. This subset acts only on invariant scalar channels and is therefore lower-dimensional and cheaper to compute than the interaction NTK. Unless otherwise stated, the NTK representation in our work uses the embedding-parameter subset, while the cost-accuracy analysis additionally compares embedding, interaction, and readout subsets in Section 4.4. The activation representation instead pools the invariant scalar node features from each messagepassing layer and concatenates them, N (x)

ϕact (x) = concat({

X

(ℓ)

hi,L=0 }Tℓ=1 ),

(17)

i=1

This representation uses the hidden node features directly and does not involve differentiating with respect to model parameters.

C

Training Details

Active-learning rounds fine-tune pretrained MACE models. Model weights are optimised on the current labelled set without freezing any model parameters. At each round, the MLIP model is initialised from the pretrained weights and is fine-tuned on the augmented dataset. This is commonly done in active learning as fine-tuning on only the newly acquired data causes catastrophic forgetting of data acquired at previous rounds. The surrogate predicts total energies and atomic forces, and the training objective is a Huber loss on energy and force errors with equal energies and force weights λE = λF = 10.0. We use a variety of settings in order to obtain converged finetunings for the variety of dataset sizes and scenarios we train on. Reactivity datasets: For the T1x, RGD, PMechDB active learnings in Section 4.2, the batch size and learning rate are chosen as ( (4, 10−2 ), |T (t) | ≤ 250, (B, η) = (18) (16, 10−2 ), |T (t) | > 250, where |T (t) | denotes the size of the labelled training set at active-learning round t. The number of epochs is chosen as    4000 B E = max 50, . (19) |T (t) | Small T1x datasets: For the small T1x dataset trainings (the biasing in Section 4.3, and cost analysis 4.4), we start off with extremely small training sets. Each round therefore uses a dynamic schedule Varga-Umbrich et al. [2026]. If |T (t) | is the current labelled-set size, the batch size B and learning rate η are  −3 |T (t) | ≤ 20, (1, 10 ), −3 (B, η) = (2, 5 × 10 ), 20 < |T (t) | ≤ 100, (20)  (4, 5 × 10−3 ), |T (t) | > 100, and the number of epochs is    1000B . E = max 10, |T (t) | 17

(21)

OC20:

For OC20, we use a fixed batch size and learning rate, (B, η) = (4, 10−3 ),

together with a dynamically adjusted number of epochs    1000 × B , E = max 30, |T (t) | where |T (t) | denotes the size of the labelled training set at active-learning round t.

18

(22)

(23)

C.1

Pretrained models

The hyperparameters from the SPICE2 and MPTraj models can be found in Tables 2 and 3 Table 2: Hyperparameters for the SPICE 2 MACE model. Hyperparameter Value Number of interaction layers (Nlayers ) 2 Number of channels (Nchannels ) 128 Maximum angular momentum (ℓmax ) 3 Node symmetry 3 Correlation order 2 Readout irreducible representations {16 × 0e, 0e} Include pseudotensors False Number of Bessel basis functions 8 Activation function SiLU Radial envelope Polynomial envelope Cutoff distance (Å) 5.0 Number of species 15

Table 3: Hyperparameters of the MPTraj MACE model. Hyperparameter Value Number of interaction layers (Nlayers ) 2 Number of channels (Nchannels ) 128 Maximum angular momentum (ℓmax ) 2 Node symmetry 2 Correlation order 3 Readout irreducible representations {16 × 0e, 0e} Number of Bessel basis functions 10 Include pseudotensors False Activation function SiLU Radial envelope Polynomial envelope Cutoff distance (Å) 6.0 Number of species 88

D

Scalable feature-space acquisition

This section details the shortlist acquisition pipeline used for large candidate pools. The implementation operates entirely in feature space and scales linearly in the pool and train size due to the shortlisting procedure. D.1

Feature-space posterior variance

In this section, for completeness, we give a derivation that posterior variance acquisition is equivalent to scoring via the feature matrix MT Rasmussen and Williams [2005]. For kernels induced by a feature map k(x, x′ ) = ϕ(x)⊤ ϕ(x′ ) with ϕ(x) ∈ Rd , posterior variance can be computed without forming the kernel matrix. Let the training feature matrix be ΦT ∈ RnT ×d and the candidate pool size be nP . The Gaussian posterior variance for a candidate x is −1

σT2 (x) = k(x, x) − kT (x)⊤ (kT T + λI)

kT (x),

(24)

where kT T = ΦT Φ⊤ T and kT (x) = ΦT ϕ(x). Substituting gives h i
−1 ⊤ σT2 (x) = ϕ(x)⊤ I − Φ⊤ ΦT ϕ(x). T ΦT ΦT + λI

(25)

Applying the Woodbury identity Golub [2012], ⊤ −1 −1 I − Φ⊤ ΦT = λ(Φ⊤ , T (ΦT ΦT + λI) T ΦT + λI)

19

(26)

yields −1 σT2 (x) = λ ϕ(x)⊤ (Φ⊤ ϕ(x). T ΦT + λI)

(27)

−1 Defining the feature-space precision matrix MT = (Φ⊤ ∈ Rd×d , we obtain σT2 (x) = T ΦT + λI) ⊤ λ ϕ(x) MT ϕ(x). Since λ is constant, ranking by posterior variance is equivalent to scoring by s(x) = ϕ(x)⊤ MT ϕ(x).

To compute MT efficiently, we accumulate the Gram matrix Φ⊤ Specifically, for T ΦT in chunks. P chunks ΦC ∈ Rc×d of the training features, we maintain a running sum C Φ⊤ C ΦC after which the feature precision matrix can be formed. This requires materialising matrices of size at most O(max(cd, d2 )), making the computation linear in nT and memory-efficient in the number of training points. Candidate scores s(x) for x ∈ P are then also computed in chunks using the same strategy. During the scoring pass, we maintain a running top-K representation shortlist by caching the highest-scoring candidates, enabling efficient selection without evaluating or storing the full pool. D.2

LCMD

Posterior variance alone tends to favour isolated high-uncertainty points. We therefore use the posterior variance score only to shortlist a set of K candidates with the highest uncertainty. We then perform batch selection using Largest-Cluster Maximum-Distance (LCMD) Holzmüller et al. [2023] on the shortlisted set. LCMD modifies greedy farthest-point sampling to favour dense but under-covered regions. Let S denote the set of points already selected into the current batch. At each step, every remaining candidate is assigned to its nearest centre in S ∪ T (t) . For each resulting cluster Cc with centre zc , we compute the total squared distance mass X mc = d2k (x, zc ), (28) x∈Cc

where the kernel-induced squared distance is d2k (x, z) = ∥ϕ(x) − ϕ(z)∥22 .

(29)

LCMD then selects the cluster with the largest mass and adds the point within that cluster farthest from its centre: xnext = arg max d2k (x, zc⋆ ), c⋆ = arg max mc . (30) c

x∈Cc⋆

By using shortlisting, the cost of including diversity via LCMD is changed from O(nP nT ) to O(KnT ). Pseudo code for the full shortlisting is given in Algorithm 1 and detailed performance summaries can be found in Appendix I. Algorithm 1 Chunked feature-space PV with shortlist selection Require: Feature map ϕ, training set T , pool P, batch size B, shortlist size K, chunk size c, regularisation λ 1: Initialize GT ← 0d×d 2: for chunks CT ⊂ T of size ≤ c do 3: Compute features ΦCT = ϕ(CT ) ∈ R|CT |×d 4: Accumulate GT ← GT + Φ⊤ CT ΦCT 5: end for 6: Compute MT = (GT + λI)−1 7: Initialize empty top-K shortlist S 8: for chunks C ⊂ P of size ≤ c do 9: Compute features ΦC = ϕ(C) ∈ R|C|×d 10: Compute scores sC = diag(ΦC MT Φ⊤ C) 11: Update S with the top-K candidates seen so far 12: end for 13: Run LCMD or greedy PV on S to select batch A of size B 14: return A

20

E

Additional reactivity information

In this section, we give further plots and benchmarks for Section 4.2 of the main text. E.1

MAE metrics

Table 4 reports the MAE area under the learning curve together with the final energy and force MAE values. Compared to the RMSE results in Table 1 of the main text, the improvements under active learning are generally smaller for MAE than for RMSE. This is consistent with active learning preferentially reducing high-error regions of the distribution, since RMSE is more sensitive to rare but large errors. We additionally observe that committee-based methods exhibit a substantially larger gap between RMSE and MAE improvements than the representation-based methods. This suggests that a significant fraction of their RMSE gains may arise from preferential selection of outlier or high-error structures, potentially due to the lack of explicit conditioning on the training set or candidate-pool geometry. Table 4: Energy and force acquisition performance across reactive benchmarks using MAE metrics. Reported metrics are energy MAE area under the acquisition curve (AUC) and final-round MAE in meV, and force MAE AUC and final-round MAE in meV/Å. Lower is better; best values in each row are bolded. Percent differences are relative to Random. Dataset

Metric Random

RGD

E AUC F AUC Final E Final F

193.5 2170.0 6.5 85.5

189.0 (-2.3%) 202.5 (+4.7%) 210.0 (+8.5%) 192.5 (-0.5%) 189.0 (-2.3%) 191.5 (-1.0%) 2185.0 (+0.7%) 2237.0 (+3.1%) 2252.5 (+3.8%) 2176.0 (+0.3%) 2174.0 (+0.2%) 2171.5 (+0.1%) 6.0 (-7.7%) 6.5 (0.0%) 6.0 (-7.7%) 6.0 (-7.7%) 6.0 (-7.7%) 6.5 (0.0%) 85.0 (-0.6%) 85.5 (0.0%) 84.5 (-1.2%) 85.0 (-0.6%) 84.5 (-1.2%) 85.0 (-0.6%)

T1X

E AUC F AUC Final E Final F

90.5 843.0 2.0 20.5

70.5 (-22.1%) 155.0 (+71.3%) 143.0 (+58.0%) 73.0 (-19.3%) 75.5 (-16.6%) 75.0 (-17.1%) 700.0 (-16.9%) 1082.0 (+28.4%) 990.5 (+17.5%) 697.0 (-17.3%) 711.5 (-15.6%) 703.5 (-16.5%) 1.5 (-25.0%) 2.0 (0.0%) 2.0 (0.0%) 2.0 (0.0%) 2.0 (0.0%) 2.0 (0.0%) 19.0 (-7.3%) 20.5 (0.0%) 20.0 (-2.4%) 19.5 (-4.9%) 19.5 (-4.9%) 20.0 (-2.4%)

PMechDB

E MAE F MAE Final E Final F

316.0 1061.5 11.5 41.5

305.0 (-3.5%) 352.0 (+11.4%) 334.5 (+5.9%) 307.5 (-2.7%) 311.0 (-1.6%) 313.0 (-0.9%) 1074.5 (+1.2%) 1104.0 (+4.0%) 1116.5 (+5.2%) 1072.5 (+1.0%) 1064.0 (+0.2%) 1068.0 (+0.6%) 11.0 (-4.3%) 13.0 (+13.0%) 12.0 (+4.3%) 11.0 (-4.3%) 11.0 (-4.3%) 11.5 (0.0%) 41.5 (0.0%) 42.0 (+1.2%) 42.0 (+1.2%) 41.0 (-1.2%) 41.0 (-1.2%) 41.0 (-1.2%)

E.2

Activation

Committee-E

Committee-F

NTK-E

NTK-F

NTK-EF

Descriptor baselines

Table 5: RMSE-based comparison of Tanimoto, SOAP, and NTK-EF acquisition strategies across reactive benchmarks. Reported values include energy RMSE area under the acquisition curve (AUC) and final-round errors in meV, together with force RMSE AUC and final-round errors in meV/Å. Percent differences are relative to the Random baseline. Dataset

Metric

Tanimoto

SOAP

NTK-EF

RGD

E AUC F AUC Final E Final F

339.0 (-2.6%) 4735.5 (-0.6%) 12.0 (-7.7%) 188.0 (-0.3%)

338.5 (-2.7%) 4744.0 (-0.4%) 12.0 (-7.7%) 186.0 (-1.3%)

344.5 (-1.0%) 4684.0 (-1.7%) 12.0 (-7.7%) 186.0 (-1.3%)

T1X

E AUC 189.5 (-23.1%) 171.5 (-30.4%) 175.0 (-29.0%) F AUC 2008.5 (-12.7%) 1914.0 (-16.8%) 1796.5 (-21.9%) Final E 6.0 (-7.7%) 5.5 (-15.4%) 5.0 (-23.1%) Final F 61.5 (-0.8%) 58.0 (-6.5%) 60.5 (-2.4%)

PMechDB

E AUC F AUC Final E Final F

450.0 (-0.7%) 2739.5 (-4.1%) 17.0 (0.0%) 111.0 (-2.2%)

21

442.0 (-2.4%) 2993.5 (+4.8%) 16.5 (-2.9%) 119.5 (+5.3%)

431.0 (-4.9%) 2777.0 (-2.8%) 16.5 (-2.9%) 109.5 (-3.5%)

Table 6: MAE-based comparison of Tanimoto, SOAP, and NTK-EF acquisition strategies across reactive benchmarks. Reported values include energy MAE area under the acquisition curve (AUC) and final-round errors in meV, together with force MAE AUC and final-round errors in meV/Å. Percent differences are relative to the Random baseline. Dataset

Metric

Tanimoto

RGD

E AUC F AUC Final E Final F

188.5 (-2.6%) 2170.0 (0.0%) 6.5 (0.0%) 85.5 (0.0%)

192.0 (-0.8%) 191.5 (-1.0%) 2191.0 (+1.0%) 2171.5 (+0.1%) 6.5 (0.0%) 6.5 (0.0%) 85.0 (-0.6%) 85.0 (-0.6%)

T1X

E AUC F AUC Final E Final F

82.0 (-9.4%) 774.5 (-8.1%) 2.0 (0.0%) 21.0 (+2.4%)

74.0 (-18.2%) 735.0 (-12.8%) 1.5 (-25.0%) 20.0 (-2.4%)

PMechDB

E AUC 322.5 (+2.1%) 312.5 (-1.1%) 313.0 (-0.9%) F AUC 1044.5 (-1.6%) 1089.5 (+2.6%) 1068.0 (+0.6%) Final E 12.0 (+4.3%) 11.0 (-4.3%) 11.5 (0.0%) Final F 41.0 (-1.2%) 42.5 (+2.4%) 41.0 (-1.2%)

22

SOAP

NTK-EF

75.0 (-17.1%) 703.5 (-16.5%) 2.0 (0.0%) 20.0 (-2.4%)

F

Energy-force weighting sweep

Here we display results as the relative weighting wE :wF is varied. Across the reactive benchmarks, we generally observe small improvements when increasing the relative energy weighting, supporting the conclusion that energy-based acquisition already captures much of the informative acquisition signal for these datasets. Overall, however, the differences between weighting variants remain small and are often attributable to only a handful of selected structures across the active-learning trajectory. Although the underlying kernels are linear combinations of energy and force components, the downstream acquisition procedures are not, so even small changes in kernel geometry can lead to different candidate selections over successive AL rounds. Table 7: Comparison of NTK-based acquisition strategies with different energy-force weighting ratios across reactive benchmarks. Reported values include energy/force RMSE area under the acquisition curve (AUC) together with final-round errors. Lower is better; best values in each row are bolded. Units are in meV and meV/Å. Percent differences are relative to the Random baseline. Dataset

Metric

NTK-E

NTK-F

NTK-EF 1:1

NTK-EF 4:1

NTK-EF 32:1

NTK-EF 128:1

RGD

E AUC F AUC Final E Final F

339.0 (-2.6%) 4651.5 (-2.4%) 12.0 (-7.7%) 186.0 (-1.3%)

338.0 (-2.9%) 4665.5 (-2.1%) 12.0 (-7.7%) 184.0 (-2.4%)

344.5 (-1.0%) 4684.0 (-1.7%) 12.0 (-7.7%) 186.0 (-1.3%)

341.5 (-1.9%) 4656.0 (-2.3%) 12.0 (-7.7%) 186.0 (-1.3%)

330.0 (-5.2%) 4614.5 (-3.1%) 12.0 (-7.7%) 184.0 (-2.4%)

323.0 (-7.2%) 4638.0 (-2.6%) 12.0 (-7.7%) 185.5 (-1.6%)

T1X

E AUC 159.5 (-35.3%) 178.5 (-27.6%) 175.0 (-29.0%) 168.5 (-31.6%) 167.5 (-32.0%) 162.5 (-34.1%) F AUC 1754.5 (-23.7%) 1802.5 (-21.6%) 1796.5 (-21.9%) 1775.0 (-22.8%) 1757.5 (-23.6%) 1770.0 (-23.0%) Final E 5.0 (-23.1%) 5.5 (-15.4%) 5.0 (-23.1%) 5.5 (-15.4%) 5.0 (-23.1%) 5.5 (-15.4%) Final F 57.0 (-8.1%) 59.5 (-4.0%) 60.5 (-2.4%) 59.5 (-4.0%) 58.0 (-6.5%) 58.0 (-6.5%)

PMechDB

E AUC F AUC Final E Final F

426.5 (-5.9%) 2718.5 (-4.9%) 16.0 (-5.9%) 106.5 (-6.2%)

433.5 (-4.3%) 2780.5 (-2.7%) 16.0 (-5.9%) 111.5 (-1.8%)

431.0 (-4.9%) 2777.0 (-2.8%) 16.5 (-2.9%) 109.5 (-3.5%)

23

432.5 (-4.5%) 2739.0 (-4.1%) 16.5 (-2.9%) 109.0 (-4.0%)

427.0 (-5.7%) 2777.5 (-2.8%) 16.0 (-5.9%) 110.5 (-2.6%)

423.5 (-6.5%) 2777.5 (-2.8%) 16.0 (-5.9%) 109.5 (-3.5%)

F.1

Learning curves

This section displays the learning curves for energy and force RMSE/MAE across the reactive benchmarks. Curves are averaged over two seeds. Energy errors are reported in meV, while force errors are reported in meV/Å.

Energy RMSE/atom [meV]

RGD

T1x

PMechDB

30

30

20 20 10 15

20

5 0

2

4

6

8

Random

10

12

14

16

Activation

18

20

0

2

4

6

8

10

AL Round

Committee E

12

14

16

Committee F

18

20

0

2

NTK E

4

6

8

NTK F

10

12

14

16

18

20

NTK EF

Figure 3: Energy RMSE learning curves (meV) for the methods presented in the main text. RGD

T1x

PMechDB

Energy MAE/atom [meV]

30 20

20

15

10

20

5

10

15 2 0

2

4

6

8

Random

10

12

14

16

Activation

18

20

0

2

4

6

8

10

AL Round

Committee E

12

14

16

Committee F

18

20

0

2

NTK E

4

6

8

NTK F

10

12

14

16

18

20

18

20

NTK EF

Figure 4: Energy MAE learning curves (meV) for the methods presented in the main text.

Force RMSE [meV/A]

RGD

300

300

T1x

PMechDB 200

200 150 150 100

200

0

2

4

Random

6

8

10

12

14

Activation

16

18

20

0

2

Committee E

4

6

8

10

AL Round

12

14

Committee F

16

18

20

NTK E

0

2

4

6

NTK F

8

10

12

14

16

NTK EF

Figure 5: Force RMSE learning curves (meV/Å) for the methods presented in the main text.

24

RGD

T1x

Force MAE [meV/A]

150

PMechDB

100

50

50

100

20 0

2

4

Random

6

8

10

12

14

Activation

16

18

20

0

2

Committee E

4

6

8

10

AL Round

12

14

Committee F

16

18

20

NTK E

0

2

4

6

NTK F

8

10

12

14

16

18

NTK EF

Figure 6: Force MAE learning curves (meV/Å) for the methods presented in the main text.

25

20

G

Details of controlled biasing

T1x provides a controlled setting for studying offline active learning under candidate-pool bias. Each reaction family contains structures along NEB and climbing-image NEB paths, so the frame index gives a discrete proxy for position along the reaction coordinate. This lets us separate two practically important forms of pool shift: imbalance across reaction families and imbalance within a reaction pathway. Formally, for reaction family r, let Xr = {xr,i,f }, where i indexes reaction instances or optimization trajectories and f ∈ {1, . . . , Fr,i } indexes frames along the path. Each point xr,i,f is a molecular geometry with associated energy and force labels. Let Nr = |Xr | denote the number of available structures for reaction family r. Uniform sampling over all structures induces p(r) ∝ Nr , producing the empirical reaction-family imbalance that we refer to as natural bias. To construct controlled inter-reaction bias, we introduce reaction-level weights w = (w1 , . . . , wR ) and sample reaction families according to p(r) ∝ wr Nr , while sampling frames uniformly within each chosen reaction. To construct controlled intra-reaction bias, we fix a reaction family and alter Fr,i the frame distribution within that family. For a trajectory of length Fr,i , frame weights {πf }f =1 define p(f | r, i) ∝ πf , f ∈ {1, . . . , Fr,i }. In all settings, the test set is fixed and balanced, so differences in performance reflect robustness to the acquisition pool rather than changes in the evaluation distribution. All biasing experiments use the same offline active-learning protocol. We randomly sample 5 pathways from T1x, giving approximately 4.2k structures, and fix a balanced test/validation set of 175/50 structures. For inter-reaction bias, we up-weight individual reaction families by a factor of 5 relative to a uniform (1:1:1:1:1) baseline. For intra-reaction bias, we similarly upweight specific frames (0th, 5th, and 8th) by a factor of 5 relative to uniform sampling along the reaction coordinate. We perform 40 active-learning rounds, acquiring 5 points per round, starting from an initial seed set of 5 structures, and finetuning a SPICE-2-pretrained MACE model after each round. G.1

Inter-reaction bias

Figure 7 summarizes the inter-reaction bias results. The left panels report energy and force RMSE AUC, while the right panels show the cumulative selection fraction across reaction families. Across both metrics, model-based methods substantially improve over random selection, with Activation and NTK-based approaches achieving the lowest AUC. Force-aware variants (NTK-F, NTK-EF) remain competitive, indicating that incorporating force information does not degrade performance under this form of bias. The selection distributions highlight differences in behaviour. Random selection largely reflects the imposed reaction imbalance, while Committee-E shows unstable allocation across reaction families. In contrast, NTK and NTK-EF distribute selections more evenly, indicating reduced sensitivity to the skewed pool. Overall, learned representation-based methods maintain more stable behaviour under inter-reaction bias than committee-based disagreement.

26

600 500 400 300 200 do

n Ra

m

-E

ee

itt

m om

AP

SO

n

TK

N

Ac

C

F RMSE AUC (meV round/A)

io

at tiv

-F EF TK TKN

9000 8000 7000 6000 5000 4000 3000 om

nd

Ra

C

-E

ee

itt

m om

AP

SO

io

at iv

t Ac

n

TK

N

1.0

Random

-F EF TK TKN

N

Committee-E

0.8 0.6 0.4 5:1:1:1:1 rxn02253 rxn02485 rxn03838

0.2 0.0

10

N

Reaction-biased - F RMSE AUC

10000

Cumulative selection fraction

700

rxn07298 rxn09087

20

30

40

10

AL round

Cumulative selection fraction

E RMSE AUC (meV round/atom)

Reaction-biased - E RMSE AUC

1.0

20

30

40

30

40

AL round

NTK

NTK-EF

0.8 0.6 0.4 0.2 0.0

10

20

AL round

30

40

10

20

AL round

Figure 7: Inter-reaction bias. Left: energy RMSE AUC (meV·round/atom) and force RMSE AUC (meV·round/Å) under reaction-weighted candidate pools. Error bars denote the standard deviation across different biasing configurations. Right: cumulative selection fraction across reaction families over active learning rounds.

27

G.2

Intra-reaction bias

The intra-reaction setting is a finer-grained test of robustness. Instead of changing the reaction-family distribution, we fix the reaction family and skew the candidate pool along the reaction coordinate by changing the frame weights πf . This makes the benchmark harder: structures within a single reaction manifold can be highly similar, so acquisition must resolve subtle geometric differences rather than simply cover distinct reaction classes. In practice, we bias frames 1–8; frames 9 and 10 are excluded from the biasing sweep because they typically contain fewer than 10 structures, making them too sparse for a stable controlled-bias comparison. Figure 8 summarizes performance under frame-biased pools. The left panels report energy and force RMSE AUC, while the right panels show the cumulative selection fraction across reaction coordinates. Across both metrics, model-based methods improve substantially over random selection, with Activation and NTK-based approaches achieving the lowest AUC. Force-aware variants (NTK-F, NTK-EF) remain competitive, indicating that incorporating force information does not degrade performance in this setting.

700 600 500 400 300 200 do

n Ra

m

-E

ee

itt

m om

AP

io

at iv ct

SO

n

TK

N

A

C

-F EF TK TKN

F RMSE AUC (meV round/A)

8000 7000 6000 5000 4000

do

m

C

m om

-E

ee

itt

AP

SO

io

Ac

at tiv

n

TK

N

1.0

Random

-F EF TK TKN

N

Committee-E

0.8 0.6 0.4 0.2 0.0

fr01 x5 FR01 FR02 FR03 FR04

FR05 FR06 FR07 FR08

10

20

N

Frame-biased - F RMSE AUC

n Ra

Cumulative selection fraction

Frame-biased - E RMSE AUC 800

30

40

10

AL round

Cumulative selection fraction

E RMSE AUC (meV round/atom)

The selection distributions highlight differences in behaviour. Random selection tracks the imposed bias, while Committee-E exhibits unstable allocation across frames. In contrast, NTK and NTK-EF produce more balanced selections along the reaction coordinate, indicating that they are less sensitive to the skewed candidate distribution. Overall, this setting shows that when variation is confined to a single reaction manifold, learned representation-based methods provide more stable behaviour than committee-based disagreement.

1.0

20

30

40

30

40

AL round

NTK

NTK-EF

0.8 0.6 0.4 0.2 0.0

10

20

AL round

30

40

10

20

AL round

Figure 8: Intra-reaction (frame) bias. Left: energy RMSE AUC (meV·round/atom) and force RMSE AUC (meV·round/Å) under frame-skewed candidate pools. Error bars denote the standard deviation across different biasing configurations. Right: cumulative selection fraction across frames over active learning rounds.

28

Table 8 reports the average AUC on the test set and the maximum deviation across bias settings. Across both regimes, model-based methods outperform random selection, with Activation and NTK-based approaches achieving the lowest AUC, and force-aware variants (NTK-F, NTK-EF) remaining competitive. Descriptor-based methods provide moderate improvements but are generally less competitive than learned representations, while Committee-E shows higher AUC and larger deviations across bias settings, indicating greater sensitivity to candidate-pool shift. Table 8: Controlled-bias summary metrics, reported as mean [max deviation] across bias settings. Final columns are final-round RMSE in meV for energy and meV Å−1 for forces. AUC is the discrete sum of RMSE over acquisition steps. Best (lowest) values are in bold. All descriptor based methods use LCMD acquisition. Bias

Method

Energy AUC (↓)

Force AUC (↓)

Final E RMSE (↓)

Final F RMSE (↓)

Inter (Reaction)

Random Committee-E Tanimoto SOAP Activation NTK NTK-F NTK-EF

633.3 [129.7] 571.7 [108.7] 587.2 [168.8] 313.2 [72.8] 245.3 [35.3] 261.3 [22.3] 329.7 [41.3] 343.2 [71.2]

7865.2 [1807.8] 7165.5 [543.5] 6973.8 [1498.2] 4909.8 [816.2] 3916.8 [217.8] 3945.3 [393.3] 4395.5 [467.5] 4287.7 [352.3]

6.3 [3.7] 4.5 [2.5] 8.5 [3.5] 2.8 [0.8] 3.2 [0.8] 2.8 [1.2] 3.5 [0.5] 3.3 [1.3]

123.7 [65.3] 83.8 [12.8] 118.3 [25.7] 58.3 [11.3] 56.0 [11.0] 54.3 [11.3] 59.5 [13.5] 59.7 [13.7]

Intra (Frame)

Random Committee-E Tanimoto SOAP Activation NTK NTK-F NTK-EF

528.3 [164.8] 671.0 [210.0] 492.0 [139.0] 362.8 [14.8] 274.5 [30.5] 279.0 [22.0] 361.8 [56.3] 338.0 [41.0]

7117.8 [1018.3] 7690.5 [722.5] 6255.0 [623.0] 5541.3 [388.8] 4086.5 [318.5] 4269.8 [261.3] 4551.0 [578.0] 4466.3 [196.3]

6.5 [2.5] 5.0 [2.0] 7.0 [3.0] 3.5 [1.5] 3.5 [0.5] 4.0 [1.0] 3.0 [0.0] 3.3 [0.8]

112.0 [19.0] 90.5 [12.5] 112.5 [26.5] 67.5 [6.5] 60.5 [10.5] 62.3 [13.8] 59.0 [12.0] 60.0 [10.0]

29

G.3

NTK force kernel plots

In this section, we visualise the structure of the force-aware NTK by comparing it to the energy NTK on 5 randomly selected reactions from T1x. In Figure 9, the kernel matrix is ordered first by reaction ID and then by frame index along each reaction pathway. Both kernels exhibit structured block patterns; however, the force-aware kernel more clearly separates regions corresponding to reactant, transition-state, and product configurations, as reflected in the finer structure within each block. Figure 9 hints at the importance of how energy and force representations are combined in joint kernels. In particular, the combined kernels visually resemble the force NTK more strongly than the energy NTK, indicating that the force contribution may dominate the joint representation despite cosine normalization. This suggests that the relative weighting and normalization of energy and force features are important design choices. More sophisticated magnitude-alignment strategies such as standard normalization prior to kernel combination could thus potentially yield better-balanced joint representations and improve uncertainty estimation performance. rxn00722

rxn00722

rxn02072

rxn02072

0.9

rxn00722 0.9

rxn02072

0.8

0.8

0.6

0.2

0.7

Kernel value

0.4

rxn05576

Kernel value

Kernel value

0.7 0.8

rxn05576

rxn05576

0.6 0.5 0.4

0.0

(b) Force NTK

0.2

9

8

35

32

06

06

rxn

rxn

6 57 05

59

28

63

63 rxn 0

rxn 0

76 rxn 0

55

07 rxn 22 02 07 2

rxn 0

9

8

35

32

06

06

rxn

rxn

6 57 05 rxn

00 7 rxn 22 02 07 2

rxn

(a) Energy NTK

0.4

0.3

rxn06359

rxn

rxn06359

rxn06359

rxn06328

00 7 rxn 22 02 07 2

0.2

0.6

rxn

rxn06328

rxn06328

(c) Energy-Force NTK

Figure 9: Global kernel matrices on a five-reaction T1x subset. Structures are sorted by reaction family and frame index. The Energy-Force NTK is an equal weighting (wE = wF = 0.5).

H

Additional OC20 information

This appendix collects experimental details and additional results for the OC20 benchmark of Section 4.1. H.1

Learning curves

Figure 10 shows the AL learning curves on the 20,000-structure test set for the headline method subset (Random, Committee-E, Committee-F, Activation, NTK-E, NTK-F, NTK-EF), as a function of training-set size from the initial 2,000 to the final 9,500 labelled structures. The joint NTK-EF achieves the lowest test error from the first acquisition round onward across all four metrics (energy and force RMSE/MAE), and the gap relative to the energy-only and force-only NTK variants persists across the full schedule. Both committee variants stay above all NTK methods on every metric, with Committee-E showing the largest deviation on energy errors. Activation tracks the NTK family closely on energy metrics but trails on force errors. H.2

Per-split test errors at the final round

Figure 11 resolves the final-round test errors of Figure 1 into all four OC20 splits, rather than aggregating the three OOD splits into a single OOD column. The ranking of methods is consistent across splits: NTK-EF is best on every (metric, split) pair on energy errors and ties with NTK-F within seed noise on the force splits. The OOD splits amplify the spread between methods most strongly on the energy axis: Committee-E rises from ≈ 29 to ≈ 32 meV/atom on E RMSE between val_is and val_oos_ads_bulk, while NTK-EF stays close to 25 meV/atom across the four splits. The energy-only NTK degrades sharply on val_oos_ads, consistent with the discussion in Section 4.1 that energy features alone do not distinguish adsorbate chemistry.

30

4 × 10

1

31.5

3 × 10

1

Test F RMSE [eV/Å]

Test E RMSE / atom [meV]

5 × 10

Random Committee-E Committee-F Activation NTK-E NTK-F NTK-EF

1

2.5 × 10

−1

2.4 × 10

−1

2.3 × 10

−1

2.2 × 10

−1

2.1 × 10

−1

2 × 10

27.1 26.3 25.7 25 24

−1

1.9 × 10

−1

1.8 × 10

−1

1.7 × 10

−1

1.4 × 10

−1

0.19 0.187 0.184 0.181 0.18 0.177 0.172

Training set size

1.35 × 10 3 × 10

1

22.9

2 × 10

20.4

1

Test F MAE [eV/Å]

Test E MAE / atom [meV]

Training set size

1.3 × 10 1.25 × 10

1.15 × 10

6000

−1

0.119 0.116

−1

0.113

1.1 × 10

17

4000

−1

−1

1.2 × 10

18.7 18

2000

−1

8000

0.111 0.11 0.11 0.109

−1

2000

4000

Training set size

6000

8000

Training set size

0.24

F RMSE [eV/Å]

E RMSE / atom [meV]

Figure 10: OC20 test error vs. training-set size on the 20,000-structure aggregated test set. Columns: energy and force RMSE/MAE (energy reported per atom, in meV; forces in eV/Å). Shaded bands are across-seed RMS deviations; the value at the right margin annotates each method’s final-round error.

30

25

0.22 0.20 0.18 0.16

20

val_is

s lk s_bulk os_ad os_bu val_o os_ad val_o val_o

val_is

s lk s_bulk os_ad os_bu val_o os_ad val_o val_o

val_is

s lk s_bulk os_ad os_bu val_o os_ad val_o val_o

0.125

24

F MAE [eV/Å]

E MAE / atom [meV]

26

22 20 18 16

0.120 0.115 0.110 0.105

val_is

Random

val_o

s lk s_bulk os_ad os_bu os_ad val_o val_o

Committee-E

Committee-F

Activation

NTK-E

NTK-F

NTK-EF

Figure 11: OC20 final-round test errors resolved by validation split (val_is, val_oos_ads, val_oos_ads_bulk, val_oos_bulk). Bars are across-seed means; error bars are across-seed RMS deviations. Y-axes are zoomed to the data range so that method differences remain visible.

31

I

Computational considerations: runtime and memory

This appendix contains per-method runtime and memory measurements behind Figure 2, and parameter-subset analysis referenced in Section 4.4. I.1

OC20 per-method timings

Table 9 reports acquisition wall-clock, training wall-clock, and peak memory usage on the OC20 candidate pool (nP ≈ 200k). At this scale, direct candidate–candidate kernel materialisation is infeasible, so all NTK and activation-based methods use the chunked shortlist-based PV pipeline described in Section 3.2. Timings are reported as steady-state averages over active-learning rounds, which starts from round 2, since the initial round is initial training, and round 1 is jit-cold Bradbury et al. [2018]. Table 9: OC20 per-round acquisition and training cost (mean ± std across seeds). Timings are reported as steady-state averages over active-learning rounds 2-5, excluding the initial training round and the jit-cold first acquisition round. “Acq.” is the time spent inside the acquisition function itself. Method NTK-EF (embeddings) LCMD NTK-EF (readouts) LCMD NTK-F (embeddings) LCMD NTK-F (readouts) LCMD NTK-E (embeddings) LCMD NTK-E (embeddings) PV NTK-E (readouts) LCMD Activation LCMD Activation PV Committee-E (M =3) Committee-F∗ (M =3) Random

Acq. (s)

Train (s)

Total (s)

8317 ± 171 3845 ± 15 5890 ± 175 3025 ± 107 5188 ± 1541 4168 ± 1456 875 ± 30 613 ± 34 982 ± 316 4707† ± 474 – <1

1924 ± 27 1840 ± 11 1729 ± 27 1850 ± 148 1736 ± 79 1829 ± 5 1843 ± 108 1804 ± 150 1956 ± 104 8890 ± 947 – 2323 ± 134

10242 ± 170 5685 ± 4 7619 ± 148 4874 ± 40 6925 ± 1466 5997 ± 1456 2718 ± 77 2417 ± 164 2937 ± 276 13596 ± 1420 13119 ± 496 2323 ± 134

† For committee acquisition, each member rebuilds the candidate-pool graphs independently, and re-JITs the force field, and

the M =3 forward passes run sequentially. Graph caching and vmap-fused implementation could reduce this fairly significantly, and could push the acquisition time below that of NTK methods, but the training (and overall wall-time) we expect to remain larger even in a more efficient implementation. ∗ Committee-F shares Committee-E’s timings: the MACE forward pass returns energy and forces jointly, so the only difference between the two is whether the per-structure score aggregates energy disagreement or per-atom force disagreement; the cost of acquisition and training is identical.

We report both PV and LCMD-based acquisition. PV selects candidates directly according to their feature-space posterior variance, while LCMD first constructs a high-uncertainty shortlist using PV and then applies diversity-based selection on the shortlist. All methods operate on the same candidate pool using the chunked feature-space pipeline; the primary difference between activation and NTKbased methods is therefore the representation dimension (dAct = 256 versus dNTK = 11,264 for the embeddings subset). Both committees train M =3 ensemble members plus a 4th evaluation model and run the same forward passes on the pool, differing only in whether energy or force disagreement is aggregated. Since the evaluation model is not necessary, one could further divide the train times by a factor 4/3 without changing the qualitative results. Committee is still Pareto dominated on both fronts. When the experiments were launched, committee-F acquisition and train times were not logged, so the only time shown in Table 9 is the total time.. I.2

Memory footprint and scaling

Table 10 shows the breakdown of peak memory usage for the chunked vs. non-chunked PV implementation across pool sizes. 32

Table 10: Acquisition peak memory overhead relative to the pre-acquisition baseline for chunked and non-chunked feature-space PV at varying pool sizes nS = nP + nT , using the OC20 NTK-E (embeddings) representation. The non-chunked implementation materialises the full nS × nS kernel and exhausts host RAM beyond nS ≈ 6 × 104 , whereas the chunked implementation scales across the full tested range. “–” indicates configurations that were not measured, and “OOM (host)” indicates that the implementation exhausted host RAM before completing.

I.3

nS

chunked PV (GiB)

non-chunked PV (GiB)

500 1,000 5,000 10,000 25,000 40,000 50,000 60,000 200,000

0.35 0.48 2.72 3.29 – – 4.09 – 5.64

0.41 0.63 2.41 4.92 14.56 26.11 35.33 44.63 OOM (host)

Parameter-subset trade-off on T1x

Table 11 compares NTK parameter subsets on a 5-reaction T1x subset using full LCMD acquisition. We report energy and force RMSE together with per-pass NTK runtime. Table 11: NTK parameter-subset cost–accuracy trade-offs on T1x. Each variant restricts the NTK to gradients with respect to a single MACE parameter block of the SPICE-2 backbone, using full (non-shortlisted) LCMD acquisition on a balanced 5-reaction T1x subset (4,196 candidates). Per-pass NTK time reports the post-JIT runtime, excluding one-time JAX compilation overhead. Test errors are averaged over the final 5 active-learning rounds, with ± denoting the across-seed standard deviation. Energy errors are reported in meV/atom and force errors in eV/Å. Lowest values in each column are highlighted in bold. Parameter subset

Parameters

NTK pass [s]

RMSE E/atom [meV]

RMSE F [eV/Å]

readouts embeddings interactions last_layer

2,192 1,920 762,880 1,372,160

1.07 ± 0.06 3.97 ± 0.12 11.93 ± 0.64 12.07 ± 0.91

2.6 ± 0.4 2.4 ± 0.3 2.4 ± 0.3 2.4 ± 0.0

0.0432 ± 0.0004 0.0462 ± 0.0020 0.0445 ± 0.0032 0.0442 ± 0.0026

Runtime increases substantially with feature dimension, ranging from ∼ 1 s for readouts to ∼ 12 s for the largest parameter subsets. Despite this, all subsets achieve comparable energy accuracy within seed noise, while the lightweight readouts and embeddings subsets remain competitive on force accuracy. On this small study, larger subsets therefore provide little measurable accuracy benefit relative to their computational cost. In practice, chunked feature-space PV requires storing a d × d posterior-covariance matrix, making feature dimension the dominant scalability bottleneck. As a result, only lightweight parameter subsets such as embeddings and readouts remain practical at OC20 scale, while larger subsets would require additional dimensionality reduction. For reference, a float32 d × d matrix with d = 105 requires approximately 4 × 1010 bytes of memory (∼ 40 GB or 37.3 GiB), meaning that the posteriorcovariance matrix alone would occupy roughly half of an 80 GB GPU.

J

Broader Impact

This work develops scalable, force-aware active learning methods for machine-learning interatomic potentials (MLIPs), aimed at reducing the cost of curating training data for atomistic simulation. The positive societal impacts are primarily indirect: more data-efficient MLIPs lower the computational and financial cost of accurate atomistic modelling, which underpins applications such as catalyst discovery for clean energy and carbon capture, battery and electrolyte design, drug discovery, and the development of novel functional materials. By reducing the number of expensive DFT calculations 33

required to fine-tune foundation models for new chemical domains, our methods can also lower the energy footprint of large-scale atomistic dataset construction and broaden access to high-quality MLIPs for research groups without large compute budgets. We do not foresee direct negative societal impacts specific to this work, as it concerns a methodological improvement to data selection given already generated structures rather than the introduction of new capabilities or models.

34