U NI P OLYMER : A U NIFIED F RAMEWORK FOR P ROPERTY P REDICTION , S TRUCTURE R ECOMMENDATION , AND E VALUATION IN P OLYIMIDE D ESIGN
arXiv:2607.29256v1 [cs.LG] 31 Jul 2026
A P REPRINT Junquan Hu School of Software Dalian University of Technology Dalian, China [email protected]
Zhihui Wang School of Software Dalian University of Technology Dalian, China [email protected]
Xinru Guo School of Software Dalian University of Technology Dalian, China [email protected]
Xintong Li School of Software Dalian University of Technology Dalian, China [email protected]
Peng Xu School of Software Dalian University of Technology Dalian, China [email protected] Kun Lu School of Software Dalian University of Technology Dalian, China [email protected]
Ben Fei∗ Department of Information Engineering The Chinese University of Hong Kong Hong Kong SAR, China [email protected]
A BSTRACT Designing polyimide structures with specific glass transition temperatures (Tg) is highly challenging. Existing methods primarily focus on target-conditioned generation, lacking an assessment of the consistency between the generated structure and the target properties. This leads to low-quality candidates deviating from the design objective entering subsequent processes, increasing invalid experiments and prolonging the development cycle. To address this issue, we propose UniPolymer, a unified framework for property prediction, target-conditioned generation, candidate evaluation, and structure recommendation in polyimide design and a dataset containing 10066 deduplicated polyimide repeating units with Tg tags (PITg-Curated) was constructed. To improve the consistency between generated candidate structures and the target Tg, UniPolymer first establishes a reliable structure-property relationship mapping through self-supervised chemical semantic learning, structural consistency enhancement, and multi-scale information fusion. Subsequently, the model employs a continuous-discrete joint Tg representation to guide the autoregressive generation of SELFIES. The generated candidate structures are further evaluated using a frozen property predictor and polyimidespecific structural constraints, and ranked according to their deviation from the target Tg, thereby preventing structures deviating from the target from entering the subsequent validation stage. Experimental results show that UniPolymer achieved a property prediction accuracy of R2 = 0.93 and a candidate structure evaluation pass rate of 73.79%, which are 2% and 1.21% higher than the best baseline, respectively. Meanwhile, the predicted Tg values of the recommended candidates are in high agreement with the results of molecular dynamics simulations, thereby reducing the number of candidates that enter the high-cost experimental stage.
∗
Corresponding author.
UniPolymer
A P REPRINT
Keywords Polyimide Design · Glass Transition Temperature · Property Prediction · Target-Conditioned Generation · Structure Recommendation Code: https://github.com/hjunquan1100/UniPolymer
Datasets: https://github.com/hjunquan1100/UniPolymer
Introduction Polyimide (PI) exhibits excellent thermal stability, mechanical properties, and chemical resistance [Yi et al., 2020], and is widely used in aerospace, flexible electronics, and microelectronic packaging. Their glass transition temperatures govern polymer-chain segmental mobility and the applicable operating-temperature range, therefore, precise control of Tg is crucial for application-specific material design [Liu et al., 2020]. Traditional development relies on experience-driven structural combinations and iterative synthesis and testing, requiring screening across a vast chemical space, resulting in high costs and long development cycles [Tran et al., 2024]. Therefore, efficient design necessitates prioritizing candidate structures associated with a given Tg. However, the Tg of polyimides is jointly determined by the local chemical environment, segmental conformation, and other factors [Ma et al., 2019], which are difficult to fully capture using molecular descriptors, structural fingerprints, or a single sequence representation [Huang et al., 2025]. Existing methods mainly use target properties to control the generated distribution but lack verification of the actual properties of generated candidates. Consequently, many low-relevance structures whose properties deviate from the target may enter subsequent stages, resulting in unnecessary experiments and prolonged material development cycles. To address the aforementioned limitations, we propose UniPolymer, a unified framework integrating property prediction, target condition generation, candidate evaluation, and structure recommendation. To address the difficulty in maintaining consistency between generated structures and target properties, we employ self-supervised chemical semantic pretraining on large-scale unlabeled general polymer data (Supplementary Material S1) during the prediction phase to obtain polymer structure representation. This representation is then transferred to the PITg-Curated (Supplementary Material S2) dataset to establish a reliable polyimide structure-Tg mapping. To enable the model to more accurately learn the polyimide structure distribution under Tg constraints, we employ a continuous-discrete joint Tg conditional pointing to guide the autoregressive generation of SELFIES [Sahu et al., 2026], reducing the risk of generating invalid molecular sequences. We then use a predictive model to estimate the Tg of candidate structures and rank them based on target bias and physical constraint rules. Finally, molecular dynamics simulations (Supplementary Material S3) are used to further validate the reliability of the recommended structures, thereby reducing the number of low-relevance candidates entering the experimental validation process. The main contributions of this paper are as follows: • We propose UniPolymer, a unified framework for polyimide Tg prediction, target-conditioned generation, candidate evaluation, and structure recommendation. • We construct the PITg-Curated dataset containing 10066 deduplicated polyimide repeating units, and achieve discriminative and transferable structural representations from limited labeled material data through polymer semantic pre-training, structural consistency enhancement, and multi-scale information fusion. • We propose a generation strategy for the continuous-discrete joint Tg condition and jointly recommend and rank candidates based on predicted Tg deviations and polyimide-specific structural constraints.
Related Work Polymer Property Prediction. With the rapid development of machine learning and generative artificial intelligence in materials science, data-driven methods have gradually become important tools for understanding polymer structure-property relationships and accelerating candidate material design [Patra, 2021]. Early polymer property prediction typically relied on molecular descriptors, structural fingerprints, and quantitative structure-property relationship models, establishing a mapping between polymer structure and target properties through artificially defined structural features [Zhao et al., 2023]. In polyimide Tg prediction, the combination of molecular dynamics simulations and machine learning has been used to analyze the relationship between structure and thermal properties [Wen et al., 2020]. Pre-training of graph convolutional networks based on large-scale combinatorial structures and fine-tuning with experimental data has alleviated the problem of limited labeled Tg data [Volgin et al., 2022]. Building on this, graph neural networks further enhance the modeling ability of structure-Tg relationships and support the screening of target candidates from a large-scale combinatorial space [Qiu et al., 2024]. Polymer Representation Learning. To reduce reliance on artificial features, self-supervised learning and chemical language models have begun to be used in polymer representation learning. Transformer-based masked language modeling can extract transferable structural representations from large-scale unlabeled polymer sequences, while 2
UniPolymer
A P REPRINT
modeling polymer structures as a chemical language enables the learning of dense representations applicable to various property prediction tasks [Kuenneth and Ramprasad, 2023]. These representation learning methods improve the accuracy of property prediction and provide more reliable structure-property relationship knowledge for propertyoriented structural design. Search-Based Polymer Design. Based on predictive models, genetic algorithms can iteratively search for polymers that satisfy target properties through crossover and mutation of structural fragments [Kim et al., 2021], and large-scale virtual screening can also identify target Tg candidates from combinatorially generated polyimide structures [Qiu et al., 2023]. However, the exploration scope of search and screening methods is usually limited by predefined structural fragments, combinatorial rules, or candidate databases [Du et al., 2024]. Despite advancements in polymer property prediction, representation learning, and property-guided design, existing methods often lack unified modeling for Tg bias in candidate generation, polyimide-specific structural constraints, and candidate recommendation ranking. UniPolymer addresses this by integrating structure-Tg relationship learning, continuous-discrete condition generation, candidate evaluation, and molecular dynamics validation. It constructs a complete polyimide design pipeline guided by the target Tg, encompassing structure-property relationship learning, target condition generation, candidate screening, and physical validation.
Methodology Overall Framework UniPolymer first learns the mapping between structure and Tg from polyimide structure-property relationship data. After completing the property model training, it freezes the parameters of the structure encoder and Tg regressor. Then, it explores the candidate structure space of polyimide with the target Tg as a condition [Liu et al., 2023], and uses the frozen property prediction model to estimate the generated candidate Tg. Combining the predicted Tg bias and polyimide structure constraints, it gives the screening and ranking of high-quality candidates (Figure 1). Finally, it further verifies the reliability of the recommendation results through molecular dynamics simulation calculations [Li et al., 2009]. The prediction and generation stages employ different molecular representations and encoding modules. In the prediction stage, each polyimide structure is represented as a SMILES sequence and encoded by the structure encoder Eϕ . The resulting sequence representation is fused with RDKit molecular descriptors and Morgan fingerprints, and the fused representation is mapped to a Tg estimate by the regressor Pω . In the generation stage, the target Tg is encoded by continuous and discrete condition paths and further transformed by the Tg condition encoder EρTg into conditional memory. Conditioned on this memory, the Transformer decoder Dθ autoregressively generates SELFIES sequences, which are then decoded into candidate molecular structures. Each candidate is subsequently converted into the inputs required by the frozen property predictor and re-encoded for Tg estimation. The relevant formulas will be given in subsequent sections. M pre Given an unlabeled polymer pretraining dataset Dpre = {spre m }m=1 , where sm denotes the SMILES sequence of the m-th polymer structure, and M is the number of pretraining samples. UniPolymer first performs self-supervised chemical-semantic pretraining on Dpre to obtain the pre-training parameters of the polymer structure encoder. Then, the pre-trained encoder is transferred to the polyimide structure–property dataset A = {(xi , Ti )}N i=1 , where xi and Ti represent the i-th polyimide repeating unit and its glass transition temperature, respectively, and N is the number of samples. UniPolymer learns the Tg-aware structure representation and structure-property relationship mapping from A, T̂i = Pω (hi ), where T̂i represents the model’s predicted Tg value for structure xi , Pω represents the Tg regressor, and hi is the multi-scale structure representation, the specific construction process of which will be introduced later. Given a target glass transition temperature T ∗ , the conditional generative model generates a candidate set C(T ∗ ) = {e xj }Jj=1 , where x̃j represents the j-th candidate structure, and J is the number of candidates. Subsequently, UniPolymer performs Tg estimation and structural verification on each candidate, and prioritizes recommending polyimide structures that are more consistent with the target properties. The property predictor remains frozen throughout candidate generation and evaluation, ensuring that all generated structures are assessed using the same structure–Tg mapping learned from experimental data. This separation prevents the evaluation criterion from changing with the generative model and provides a consistent basis for comparing candidates generated under different target conditions.
Tg-Aware Structure–Property Learning UniPolymer progressively transfers from general-polymer pretraining [Martin and Audus, 2023] to the polyimide domain through three stages: chemical-semantic pretraining on unlabeled polymers, structure-consistency adaptation on 3
UniPolymer
A P REPRINT
Figure 1: Overview of UniPolymer. By jointly screening based on property consistency and structural rationality, a unified design process is achieved from target Tg input to the recommendation of highly target-relevant structures. PITg-Curated, and multi-scale supervised Tg regression. This strategy combines transferable polymer knowledge with polyimide-specific structural and property supervision. In the prediction model stage, we employ a progressive Tg-aware representation learning strategy that transfers from general polymer pre-training [Martin and Audus, 2023] to the polyimide domain. This process includes three stages: general polymer chemical semantic pre-training, polyimide structural consistency enhancement, and multi-scale Tg regression. First, the model learns transferable chemical semantic representations through self-supervised chemical semantic pre-training on large-scale unlabeled polymer structures. Subsequently, structural consistency learning and supervised property regression are performed on the PITg-Curated dataset to construct an efficient model of the polyimide structure-Tg relationship. Chemical Semantic Pretraining. For the m-th structure in the general polymer pre-training dataset, it is first converted into a tokenized SMILES sequence pre pre pre pre spre m = sm,1 , sm,2 , . . . , sm,Lm , sm,l represents the l-th token in the sequence, and Lm represents the length of the sequence. Then, UniPolymer performs masked language modeling for self-supervised pre-training [Irwin et al., 2022]. For spre m , the mask position set Mm is randomly sampled, and the corresponding tokens are replaced with a mask token, resulting in the perturbation sequence smask . The encoder learns the contextual chemical semantics in the polymer m sequence by recovering the original tokens at the masked positions [Ross et al., 2022], and its training objective loss is defined as:
LMLM = −
M X X
mask log pϕpre spre | s . m m,l
(1)
m=1 l∈Mm
Here, pϕpre is the token distribution predicted by the Transformer encoder Eϕpre . Recovering the masked tokens enables the encoder to learn functional-group combinations and chain-connectivity patterns, thereby providing transferable chemical priors for Tg modeling with limited labeled data [Li and Jiang, 2021]. Structure-Consistent Representation Enhancement. The pretrained encoder is transferred to PITg-Curated for structure-consistency learning. Because one polyimide structure can have multiple equivalent SMILES representations [Arús-Pous et al., 2019], differences in sequence (1) notation may cause representation shifts. UniPolymer therefore constructs two randomized SMILES views, si and 4
UniPolymer
A P REPRINT
(2)
si , for each structure and processes them using a shared encoder Eϕ and projection head gψ : (a) gψ Eϕ (si ) (a) a ∈ {1, 2}, zi = (a) gψ Eϕ (si )
(2)
2
(a) zi represents the normalized representation corresponding to the a-th equivalent view. For a training batch containing
B samples, a symmetric contrast objective is used: # " (2) (1) (1) (2) B exp sim(zi , zi )/τ exp sim(zi , zi )/τ 1 X + log P . Lcon = − log P (1) (2) (2) (1) B B 2B i=1 j=1 exp sim(zi , zj )/τ j=1 exp sim(zi , zj )/τ
(3)
Here, sim(·, ·) denotes cosine similarity and τ is the temperature coefficient. The objective pulls equivalent SMILES representations together while separating different structures in the batch, transferring pretrained chemical semantics to the polyimide domain and reducing sensitivity to SMILES notation [Lee et al., 2025]. Multi-Scale Tg Regression. To compensate for the insufficient coverage of local and global structural information by single sequence representations, UniPolymer combines the Transformer sequence representation hseq = Eϕ (si )[CLS] , RDKit molecular descriptors i fp des hi = Fdes (xi ), and Morgan fingerprints hi = Ffp (xi ) [Rogers and Hahn, 2010]. These features encode sequence context, global molecular properties, and local substructure patterns, respectively. The three types of structural information are then fused by vector concatenation along the feature dimension: fp des hi = hseq i ; hi ; hi .
(4)
Subsequently, the regressor Pω gives the predicted Tg of the structure xi , and the property regression stage uses a sample-wise weighted Huber loss: B 1 X LTg = wi ℓHub Tbi , Ti . (5) B i=1 Here, wi is the weight of the i-th sample, and ℓHub () is the Huber error function. Sequence semantics, overall molecular properties, and local substructure information work together to form a Tg-aware representation and establish a structureTg mapping [Shen et al., 2024]. After training, the encoder Eϕ , feature fusion module, and regressor Pω are all frozen and used only to generate candidate Tg estimates. Target-Conditioned Polyimide Generation After establishing the structure-Tg mapping, UniPolymer explores the polyimide structure space conditioned on the target Tg [Jiang et al., 2024]. To balance continuous numerical relationships with temperature-level differences, we construct a continuous-discrete cooperative Tg conditional representation and use it for SELFIES autoregressive generation. For the Tg value Ti , it is first standardized using the mean µT and standard deviation σT of the training set: T T i = Tiσ−µ . Then, the continuous path obtains the continuous condition vector hnum = Fnum (T i ) via the non-linear i T mapping Fnum , where hnum represents the continuous Tg condition vector. Meanwhile, the Tg range is divided into i KT temperature intervals, and the corresponding discrete temperature interval index is obtained via Bin(): bi = Bin(Ti ),
bi ∈ {0, 1, . . . , KT − 1}.
(6)
The discrete path embeds the temperature-bin index as hbin using Ebin . The continuous and discrete condition vectors i are then fused in a shared latent space [Pinheiro et al., 2022]: num hTg + hbin , i i = LN hi
(7)
LN() represents the layer normalization operation, and hTg i represents the fused Tg conditional representation. Finally, the conditional encoder EρTg further encodes it into conditional memory: ri = EρTg hTg . (8) i Here, continuous paths preserve the fine-grained numerical relationship of Tg [Born and Manica, 2023], while discrete paths characterize its temperature range information, together, they provide complementary target conditions. 5
UniPolymer
A P REPRINT
Conditional SELFIES Autoregressive Generation. To improve the chemical robustness of the generated sequences, the polyimide structure is first converted into SELFIES [Krenn et al., 2020, Nigam et al., 2021], the polyimide structure is first converted into SELFIES. The i-th structure is represented as ui = (ui,1 , ui,2 , . . . , ui,Si ), where Si is the sequence length. Given the Tg conditional memory ri , the conditional Transformer Decoder Dθ performs autoregressive modeling on the structure sequence: pθ (ui | Ti ) =
Si Y
pθ (ui,t | ui<t , ri ) ,
(9)
t=1
where pθ is the conditional probability distribution with parameter θ, and ui<t represents the SELFIES tokens before position t. Teacher forcing is used during the training phase [Williams and Zipser, 1989]. Let V denote the SELFIES vocabulary and Ω the set of all non-padding positions in a batch. The generation loss is then defined as: Lgen = −
1 X X qi,t,v log pθ (v | ui<t , ri ) , |Ω|
(10)
(i,t)∈Ω v∈V
where v represents any token in the vocabulary, and qi,t,v represents the label smoothing supervision probability: 1 − ε, v = ui,t ε , (11) qi,t,v = v ̸= ui,t |V| − 1 ε represents the label smoothing coefficient. During the inference phase, the target Tg undergoes continuous-discrete conditional encoding to obtain ri . The model progressively generates SELFIES from the initial token and uses nucleus sampling to obtain candidates that possess both target relevance and structural diversity [Holtzman et al., 2019], ultimately decoding them into polyimide structures. Recommendation and Evaluation To ensure consistency between the generated candidates and the target Tg, UniPolymer uses a frozen prediction model to evaluate the properties of the candidates x̃j and defines the target bias as dj = Tbj − T ∗ , where a smaller dj indicates a greater consistency between the candidate and the target Tg. Simultaneously, structural verification is performed on the candidates: χ(e xj ) = χparse (e xj )χimide (e xj )χmol (e xj ).
(12)
Here, χ(e xj ) ∈ {0, 1} indicates whether the candidate has passed structural verification, χparse indicates the RDKit parsability check, χimide represents the imide skeleton check, χmol represents the molecular property check, the judgment process is as follows: χmol (x) = I (150 ≤ MW (x) ≤ 2000) . × I (10 ≤ nH (x) ≤ 150) I (U (x) = 0)
(13)
Among them, MW (x), nH (x), and U (x) represent the detection results of molecular weight, number of heavy atoms, and unstable groups, respectively. The target-related effective candidate set is defined as: Cvalid (T ∗ ) = {e xj ∈ C | dj ≤ δT , χ(e xj ) = 1} .
(14)
Here, δT = 20 ◦ C is the maximum allowable absolute deviation from the target Tg. UniPolymer sorts the candidates in Cvalid (T ∗ ) in ascending order of dj and returns the top R recommended structures: RR (T ∗ ) = arg topR (−dj ) .
(15)
x ej ∈Cvalid (T ∗ )
RR (T ∗ ) represents the recommended candidate set corresponding to the target T ∗ . Through this procedure, UniPolymer performs candidate-property evaluation, structural screening, and target-relevant structure recommendation. 6
UniPolymer
A P REPRINT
Experiments and Results Datasets and Settings This paper constructs a polyimide structure-glass transition temperature dataset, PITg-Curated, containing 10066 deduplicated polyimide repeating units and their Tg labels. The cleaned Tg ranges from 19 to 730 °C. The dataset is divided into training, validation, and test sets at a ratio of 8:1:1 and represented using SMILES and SELFIES, respectively, for Tg-aware representation learning and target condition generation. In the property prediction task, MAE, RMSE, and R2 are used for evaluation [Chicco et al., 2021]. The generation task is evaluated from two perspectives: structural validity and target-property consistency. Structural validity includes RDKit parsability and compliance with polyimide-specific rules covering the imide skeleton, molecular weight, heavy-atom count, and potentially unstable groups. Target-property consistency is evaluated using the frozen Tg predictor. The results were evaluated using Target MAE and Hit@20. The proportion of candidates that simultaneously satisfy both structural constraints and |Tb − T ∗ | ≤ 20◦ C was defined as the final recommendation pass rate (Valid Hit@20). Furthermore, the top 100 candidates under each target Tg condition were selected for molecular dynamics simulations, and the results were compared with model predictions to verify the reliability of the property assessment and recommendation results. We implement UniPolymer in PyTorch 2.5.1 with Python 3.12 and CUDA 12.4, and conduct all experiments on an NVIDIA A40 GPU, and the source code is publicly available. The model uses the AdamW optimizer with an initial learning rate of 1 × 10−4 , a weight decay of 1 × 10−2 , and a label smoothing coefficient of 0.1. Results and Analysis Tg Prediction Performance and Error Analysis. Model
R2 ↑ MAE (◦ C) ↓ RMSE (◦ C) ↓
DNN ANN CatBoost Importance-Transformer TransPolymer [email protected] LLaMA-3-8B UniPolymer
0.85 0.78 0.90 0.84 0.89 0.91 0.88 0.93
20.09 23.88 18.58 20.61 18.77 18.14 20.53 17.87
27.60 33.09 23.06 27.64 27.86 38.41 58.13 24.42
Table 1: Tg prediction performance on the PITg-Curated test set. On the test set (Figure 2(a)), UniPolymer achieves R2 = 0.93, MAE = 17.87 ◦ C, and RMSE = 24.42 ◦ C, providing a reliable structure–property mapping for candidate evaluation. Figure 2(b) further shows that 65.49% of the test samples fall within 20°C of the experimental Tg. As summarized in Table 1, UniPolymer outperforms the representative baselines in terms of R2 and MAE.
7
UniPolymer
A P REPRINT
Figure 2: UniPolymer’s prediction performance of polyimide Tg on the test set. The predicted values and experimental values generally show a high consistency. Chemical Space Generation and Structural Quality Analysis. This paper uses Morgan fingerprinting to perform UMAP dimensionality reduction analysis on the distribution relationship between generated candidate polyimides and real polyimides in chemical space.
Figure 3: Generated and real polyimide structures are compared in terms of chemical-space distributions and structural similarities. UniPolymer captures the principal structural characteristics of real polyimides while expanding the explored chemical space. Figure 3(a) shows that the generated structures largely overlap with real polyimides in the principal chemical-space regions, indicating that the model learns the dominant distribution of the training data rather than producing structurally isolated samples. Figure 3(b) provides a complementary structure-level analysis by reporting the maximum Tanimoto similarity [Bajusz et al., 2015] between each generated structure and its nearest neighbor in the real dataset. The median similarity of 0.42 suggests that the candidates retain recognizable polyimide characteristics without simply reproducing known structures. Taken together, the two analyses show that UniPolymer preserves the underlying chemical distribution while extending it toward less explored regions. Nevertheless, chemical-space coverage does not necessarily imply chemical validity or compliance with polyimide-specific requirements. We therefore evaluate all generation baselines on 10000 candidates using RDKit validity, imide-skeleton constraints, molecular-property ranges, structural stability, novelty, plus the overall pass rate. As shown in Table 2, UniPolymer reaches 100% RDKit validity, 8
UniPolymer
A P REPRINT
a 77.63% imide-skeleton pass rate, 65.54% novelty, with an overall pass rate of 73.79%. This final metric exceeds the strongest baseline by 1.21 percentage points, showing that the expansion of chemical space is achieved without sacrificing the structural quality of generated candidates. Model CharRNN GraphINVENT REINVENT-RL PolyTAO-style PoGE UniPolymer
Chemical Validity Structural Validity RDKit
Imide skeleton
69.57 79.86 94.62 100.00 96.75 100.00
58.43 64.37 73.84 75.96 68.74 77.63
Molecular Property Validity
Novelty Valid
Molecular weight range Heavy atom count Stable groups 93.27 95.14 97.67 99.93 96.28 99.91
91.60 93.82 89.93 98.17 95.47 100.00
84.72 87.63 92.46 94.92 83.69 94.38
38.96 47.54 61.27 63.86 57.38 65.54
29.38 40.16 61.02 72.58 53.97 73.79
Table 2: Structural and molecular-property validity of generated candidates (%). Evaluation of Target-Tg Conditional Response and Property Consistency. In addition to structural quality, this paper further evaluates the responsiveness of the generated candidate structures to the target Tg. Under four target Tg conditions (T ∗ = 100◦ C, 200◦ C, 300◦ C, 400◦ C), we generate 1000 candidates and report Target MAE, Hit@20, and the final recommendation pass rate. Model predictions are further compared with molecular-dynamics results for the top 100 candidates per target.
Figure 4: Target-Tg-driven polyimide structure exploration: (a) predicted Tg distributions; (b) Target MAE and Hit@20; (c) agreement between model predictions and MD estimates; (d) representative recommended structures. As shown in Figure 4(a), the predicted Tg distributions shift toward the corresponding temperature ranges as the target Tg increases and remain concentrated around the target values. The Target MAE values under the four target conditions are 15.63, 17.83, 23.59 and 28.76 ◦ C, with Hit@20 values of 54.67%, 58.92%, 50.23% and 42.36%, respectively (Figure 4(b)). To verify the reliability of the recommendation results, molecular dynamics simulations were performed for the top 100 candidates recommended in each target-temperature range (Figure 4(c)). The comparison between the model predictions and MD estimates yields average MAE and RMSE values of 1.95 ◦ C and 2.40 ◦ C, respectively. Most candidates have simulated Tg values close to the corresponding target temperatures, while Figure 4(d) presents representative candidate structures from each temperature range. 9
UniPolymer
A P REPRINT
Ablation Experiment We ablate the Tg-aware representation and target-condition generation modules separately. Method
R2 ↑ MAE (◦ C) ↓ RMSE (◦ C) ↓
w/o Chemical Semantic Pretraining w/o Structure Consistency w/o Molecular Descriptors w/o Morgan Fingerprints UniPolymer (pred)
0.86 0.89 0.91 0.84 0.93
21.57 19.96 18.12 23.42 17.87
26.18 25.72 25.21 27.14 24.42
Table 3: Ablation study of the Tg prediction module.
Target MAE (◦ C) ↓ Hit@20 (%) ↑ Valid Hit@20 (%) ↑
Method Continuous Condition Discrete Condition UniPolymer (gen)
24.87 26.16 21.46
49.76 48.61 51.55
46.37 42.92 50.84
Table 4: Ablation of the target-Tg condition representation. Table 3 shows that each component contributes to prediction performance. Removing Morgan fingerprints causes the largest degradation: R2 falls to 0.84, whereas MAE rises to 23.42 ◦ C, underscoring the importance of local substructure patterns for distinguishing polyimides with different Tg values. Excluding semantic pretraining or structure consistency raises MAE to 21.57 and 19.96 ◦ C, respectively. Together, these ablations show that self-supervised chemical knowledge, invariance across equivalent SMILES representations, plus multi-scale structural cues support reliable Tg modeling. Turning to conditional generation, Table 4 compares the complete condition representation against continuous-only and discrete-only variants, averaged over four target temperatures. Relative to these variants, UniPolymer lowers Target MAE by 3.41 and 4.70 ◦ C; Hit@20 increases by 1.79 and 2.94 percentage points; Valid Hit@20 rises by 4.47 and 7.92 percentage points, respectively. These gains suggest that continuous conditions preserve fine-grained numerical variation, whereas discrete conditions encode complementary temperature ranges. Their combination yields more accurate target-conditioned generation.
Conclusion This paper presents UniPolymer, a unified framework designed to maintain consistency between generated polyimide candidates and the target Tg through property prediction, conditional generation, candidate evaluation, and structure recommendation. It establishes a reliable structure-Tg mapping through self-supervised chemical semantic learning, structural consistency enhancement, and multi-scale information fusion, and guides SELFIES generation using continuous–discrete Tg conditions. Candidates are further ranked according to their absolute deviations from the target Tg and their compliance with structural constraints. UniPolymer achieves R2 = 0.93 and an MAE of 17.87 ◦ C for Tg prediction, together with a 73.79% structural pass rate. The close agreement between model predictions and MDestimated Tg values further supports the reliability of candidate evaluation. Future research will focus on improving the prediction and generation capabilities for high Tg regions and introducing synthetic feasibility, uncertainty assessment, and wet experimental validation of recommended structures to enhance the reliability and practical application value of recommended candidates.
Acknowledgments This work was supported by the National Science and Technology Major Project on New Generation Artificial Intelligence: Research on Intelligent Methods for Scientific Research Oriented to Materials Design (No. 2025ZD0121802), the Data Intelligence Innovation Base Project (No. FPF10120260003), and the Scientific Embodied Multi-Agent System Project (No. FPF10120250008). 10
UniPolymer
A P REPRINT
Appendix S1: General-Polymer Pretraining Corpus Data Source and Corpus Construction To obtain transferable chemical representations before adapting the structure encoder to the polyimide domain, we employed the unlabeled polymer corpus released with TransPolymer [Xu et al., 2023], which was constructed from the PI1M database. PI1M contains approximately one million hypothetical polymer repeating-unit structures generated from polymer structures collected from the PoLyInfo database. Each polymer repeating unit is represented using a polymer SMILES (P-SMILES) sequence, in which two wildcard atoms (*) indicate the polymerization sites connecting adjacent repeating units. Following the data-processing procedure of TransPolymer, canonicalization was removed, and non-canonical SMILES corresponding to the same polymer structure were generated for sequence augmentation. Each original polymer entry was augmented into five SMILES representations, producing approximately five million unlabeled polymer sequences. S2: Construction and Representation of the PITg-Curated Dataset Data Collection and Standardization The PITg-Curated dataset contains 10066 deduplicated polyimide repeating units and their glass transition temperatures. Of these, 2129 records (21.15%) originated from molecular dynamics simulations conducted in this paper, while 7,937 records (78.85%) were collected from publicly available polymer databases and peer-reviewed literature. The publicly available data were primarily collected from PoLyInfo [Ishii et al., 2024], the Polymer Property Predictor and Database, and related literature. The repeating-unit structures, Tg values, temperature units, and measurement methods were manually verified. Records with unclear structural or property information were removed. All structures were parsed and normalized using RDKit [RDKit Contributors, 2026] and converted into canonical SMILES. Tg values were uniformly expressed in degrees Celsius, and records with the same canonical SMILES were considered to represent the same polyimide structure. For structures with multiple valid Tg labels, the median was used as the final label to reduce the influence of differences among data sources, testing conditions, and outliers. Statistical Characterization of PITg-Curated The Tg values in PITg-Curated range from 19 ◦ C to 730 ◦ C, with a mean, median, and standard deviation of 334.74 ◦ C, 305.00 ◦ C, and 112.98 ◦ C, respectively. As shown in Figure S1(a), the data are mainly distributed within the 200–500 ◦ C range, accounting for 82.21% of all samples, with the 200–300 ◦ C range containing the largest number of samples. At the same time, the dataset also covers low- and high-Tg regions, forming a relatively broad temperature distribution and providing a data foundation for learning polyimide structure–property relationships across different temperature ranges. To further characterize the structural composition of the dataset, we statistically analyzed representative structural features, including ether bridges, fused aromatic rings, flexible alkyl linkers, fluorinated groups, alicyclic units, ketone/carbonyl bridges, and sulfone bridges, based on RDKit substructure matching. As shown in Figure S1(b), PITg-Curated contains both rigid structures represented by fused aromatic rings and flexible structures containing ether bridges, flexible alkyl linkers, and alicyclic units. It also contains various fluorinated and polar functional groups. These structural features indicate that the dataset exhibits structural diversity in terms of chain-segment rigidity, bridging mechanisms, and local chemical environments, providing a sufficient chemical basis for the model to learn the complex relationship between polyimide structure and Tg.
11
UniPolymer
A P REPRINT
Figure S1: Statistical characterization of the PITg-Curated dataset. (a) Distribution of polyimide structures across different Tg ranges. (b) Composition of representative structural groups in the dataset. S3: Molecular Dynamics Simulation and Glass Transition Temperature Calculation To supplement the experimental data available in publicly accessible databases and literature, we further employed allatom molecular dynamics simulations to calculate the glass transition temperatures of selected polyimide structures [Huo et al., 2025, Suter et al., 2025]. The complete calculation procedure included polymer-chain construction, amorphous periodic model generation, geometry optimization, cyclic annealing, multi-temperature kinetic equilibration, and density–temperature fitting. All simulations were performed using Materials Studio. The amorphous models were constructed using the Amorphous Cell module, while geometry optimization and molecular dynamics calculations were performed using the Forcite module. Interatomic interactions were described using the COMPASS force field, following established atomistic simulation procedures for polyimide Tg evaluation [Huo et al., 2025]. The key steps of the simulation procedure are described below. Construction of Polymer Chains and Amorphous Periodic Models For each polyimide structure, a corresponding three-dimensional molecular model was first established based on its repeating unit. By specifying the first and last atoms of the repeating unit, the polymerization connection sites between adjacent repeating units were determined. The repeating units were then sequentially connected along the polymerization direction to construct a linear homopolymer chain with a preset chain length. To reduce the influence of a single initial conformation on subsequent simulation results, the main-chain dihedral angles were randomly sampled during polymer-chain generation, allowing different chain segments to adopt diverse initial conformations, as illustrated in Figure S2(a)–(b). After the polymer chains were constructed, they were randomly packed into an amorphous simulation box with threedimensional periodic boundary conditions, as shown in Figure S2(c). Each simulation system contained 10 identical polymer chains with an initial density of 0.6 g/cm−3 . The build quality of the amorphous model was set to medium, and five candidate initial configurations were generated for each polyimide system. The relatively low initial packing density reduced severe atomic overlap during random packing and provided sufficient conformational space for subsequent chain rearrangement, system compression, and density equilibration. The COMPASS force field was used during model construction, together with the atomic charges and force-field types assigned to the current structure.
12
UniPolymer
A P REPRINT
Figure S2: Construction of the polyimide amorphous periodic model. (a) Definition of the polyimide repeating unit and the first and last polymerization sites. (b) Construction of a linear homopolymer chain from the repeating units and packing of the chains into a three-dimensional periodic simulation box. (c) Amorphous polymer model generated with an initial density of 0.6 g/cm−3 .
Geometry Optimization and Cyclic Annealing After the amorphous periodic model was constructed, geometry optimization and cyclic annealing were performed, as illustrated in Figure S3. Geometry optimization was performed using the Smart algorithm in the Forcite module. The maximum number of iterations was set to 50000, and the van der Waals interaction cutoff radius was set to 9.5 Å. During optimization, the atomic coordinates of the system were continuously adjusted to reduce the total potential energy contributed by bond stretching, bond-angle bending, dihedral torsion, and non-bonded interactions. This procedure eliminated unreasonable atomic contacts and local high-energy conformations generated during the random packing of polymer chains, resulting in a more physically reasonable initial amorphous structure. After geometry optimization, the system was further subjected to cyclic annealing. The annealing simulation was performed under the NVT ensemble for 10 heating–cooling cycles. In each cycle, the system was heated from 300 K to 800 K and subsequently cooled to the initial temperature. 13
UniPolymer
A P REPRINT
Figure S3: Structural optimization and annealing equilibration of the polyimide amorphous model. (a) Geometry optimization using the Smart algorithm. (b) Cyclic annealing performed over the temperature range of 300–800 K. (c) Stable amorphous structure obtained after geometry optimization and cyclic annealing. The heating process was divided into 25 temperature intervals, with 200 molecular dynamics steps performed at each interval. The integration time step was set to 1.0 fs.
Multi-Temperature NVT and NPT Equilibration After geometry optimization and cyclic annealing, multi-temperature molecular dynamics simulations were conducted within a preset temperature range to obtain the equilibrium density of the polyimide system and its temperaturedependent relationship [Martí et al., 2024, Suter et al., 2025]. The simulation temperature range was centered on the estimated Tg of the polymer and extended toward both higher and lower temperatures. Multiple discrete temperature points were established at intervals of 20 ◦ C. At each temperature point, kinetic equilibration was first performed under the NVT ensemble, allowing the system to reach a stable state near the target temperature while maintaining a fixed simulation-cell volume. The NVT simulation employed the COMPASS force field and a Nosé thermostat. A total of 300000 molecular dynamics steps were performed at each temperature, corresponding to a simulation time of 300 ps. During the NVT simulation, system equilibration was evaluated by monitoring changes in temperature and energy components over time. As shown in Figure S4(a), the representative system exhibited relatively large temperature fluctuations during the initial stage of the simulation. These fluctuations were associated with the initial velocity distribution and the rapid adjustment of polymer-chain segments. As the simulation progressed, the fluctuation amplitude gradually decreased, and the system temperature eventually stabilized near the target temperature. Correspondingly, the potential energy, kinetic energy, non-bonded energy, and total energy gradually changed from large initial fluctuations to stable fluctuations without persistent drift, as shown in Figure S4(b). These results indicate that the system reached a stable thermodynamic state at the current temperature. The amorphous periodic structure obtained after NVT equilibration is shown in Figure S4(c). This structure was subsequently used as the initial configuration for the NPT simulation at the same temperature. 14
UniPolymer
A P REPRINT
Figure S4: Representative NVT equilibration process before density extraction. (a) Temperature evolution during the NVT simulation. (b) Variations in potential, kinetic, non-bonded, and total energies. (c) Equilibrated amorphous periodic structure.
Following NVT equilibration, NPT molecular dynamics simulations were performed at each temperature point. While maintaining a constant target temperature, an external pressure of 0.0001 GPa was applied to the system, allowing the volume of the periodic simulation cell to adjust automatically according to the thermal motion of the polymer chains. Compared with fixed-volume NVT simulations, NPT simulations allow the system to undergo thermal expansion or contraction and therefore provide more physically reasonable equilibrium volumes and average densities at different temperatures. At high temperatures, polymer-chain motion is enhanced, the free volume increases, and the system density is relatively low. As the temperature decreases, chain motion becomes progressively restricted, interchain packing becomes denser, and the system density increases accordingly. For each temperature point, the average density was extracted only from the stable portion of the NPT trajectory. The resulting average densities were used to establish the density–temperature dataset required for subsequent Tg fitting. 15
UniPolymer
A P REPRINT
Determination of the Glass Transition Temperature Based on the Density–Temperature Relationship To determine the simulated glass transition temperature, the average density of the steady-state portion of the NPT trajectory was extracted at each simulated temperature, producing the following temperature–density dataset: N
T Dρ = {(Ti , ρi )}i=1 ,
(S1)
where Ti represents the i-th simulated temperature, ρi represents the average density obtained from the NPT simulation at that temperature, and NT represents the number of temperature points used for fitting. A continuous piecewise linear fit was subsequently performed using all temperature–density data to describe the different thermal expansion behaviors of the polymer in the low-temperature glassy state and the high-temperature rubbery state [Gudla and Zhang, 2024, Suter et al., 2025]. The corresponding density–temperature relationship is expressed as: ρ(T ) =
ag T + bg ar T + br
T ≤ TgMD T > TgMD ,
(S2)
where ag and ar are the slopes of the two fitted lines, while bg and br are the corresponding intercepts. To ensure that the density–temperature relationship remains continuous at the transition point, the two linear relationships satisfy: ag TgMD + bg = ar TgMD + br .
(S3)
Accordingly, the glass transition temperature obtained from the molecular dynamics simulations is calculated as: TgMD =
br − b g . ag − a r
(S4)
During the fitting process, the squared error between the observed densities at all temperature points and the densities predicted by the piecewise linear model was minimized. The fitting objective is expressed as: TbgMD = argmin Tg
NT X
2
[ρi − ρb (Ti ; Tg )] .
(S5)
i=1
As shown in Figure S5, the black triangles represent the average densities obtained from NPT simulations at different temperatures, while the red lines represent the continuous piecewise linear fitting results. As the temperature increases, the overall system density decreases. However, the rate of density change varies significantly near 260 ◦ C, reflecting the transition of the polymer from the low-temperature glassy state to the high-temperature rubbery state. The transition point obtained from the piecewise fitting is: TgMD = 259.8 ◦ C.
(S6)
Therefore, 259.8 ◦ C was determined as the molecular dynamics simulation Tg of this representative polyimide structure and was used as the simulation label of the corresponding structure in the PITg-Curated dataset.
16
UniPolymer
A P REPRINT
Figure S5: Determination of the simulated glass transition temperature by continuous piecewise linear fitting of the density–temperature relationship. The black triangles represent the average densities obtained from NPT simulations at different temperatures, and the red lines represent the fitted low-temperature glassy-state and high-temperature rubbery-state relationships.
References Chenghan Yi, Weimin Li, Sheng Shi, Ke He, Pengchang Ma, Ming Chen, and Chunlei Yang. High-temperature-resistant and colorless polyimide: Preparations, properties, and applications. Solar Energy, 195:340–354, 2020. Tuan-Qing Liu, Feng Zheng, Xiaoru Ma, Tong-Mei Ding, S. Chen, W. Jiang, S.-Y. Zhang, and Qinghua Lu. High heat-resistant polyimide films containing quinoxaline moiety for flexible substrate applications. Polymer, 209:122963, 2020. doi:10.1016/j.polymer.2020.122963. Huan Tran, Rishi Gurnani, Chiho Kim, Ghanshyam Pilania, Ha-Kyung Kwon, Ryan P Lively, and Rampi Ramprasad. Design of functional and sustainable polymers assisted by artificial intelligence. Nature Reviews Materials, 9(12): 866–886, 2024. Xiaoru Ma, Feng Zheng, Cornelia GCE van Sittert, and Qinghua Lu. Role of intrinsic factors of polyimides in glass transition temperature: An atomistic investigation. The Journal of Physical Chemistry B, 123(40):8569–8579, 2019. Qi Huang, Yedi Li, Lei Zhu, Qibin Zhao, and Wenjie Yu. Unified multimodal multidomain polymer representation for property prediction. npj Computational Materials, 11(1):153, 2025. Harikrishna Sahu, Wei Xiong, Anagha Savit, Shivank S Shukla, and Rampi Ramprasad. Polyt5: an encoder-decoder foundation chemical language model for generative polymer design. npj Artificial Intelligence, 2(1):30, 2026. Tarak K Patra. Data-driven methods for accelerating polymer design. ACS Polymers Au, 2(1):8–26, 2021. Yuankai Zhao, Roger J Mulder, Shadi Houshyar, and Tu C Le. A review on the application of molecular descriptors and machine learning in polymer design. Polymer Chemistry, 14(29):3325–3346, 2023. Chengyuan Wen, Binghan Liu, Josh Wolfgang, Timothy E Long, Roy Odle, and Shengfeng Cheng. Determination of glass transition temperature of polyimides from atomistic molecular dynamics simulations and machine-learning algorithms. Journal of Polymer Science, 58(11):1521–1534, 2020. Igor V Volgin, Pavel A Batyr, Andrey V Matseevich, Alexey Yu Dobrovskiy, Maria V Andreeva, Victor M Nazarychev, Sergey V Larin, Mikhail Ya Goikhman, Yury V Vizilter, Andrey A Askadskii, et al. Machine learning with enormous “synthetic” data sets: predicting glass transition temperature of polyimides using graph convolutional neural networks. ACS omega, 7(48):43678–43691, 2022. Haoke Qiu, Jingying Wang, Xuepeng Qiu, Xuemin Dai, and Zhao-Yan Sun. Heat-resistant polymer discovery by utilizing interpretable graph neural network with small data. Macromolecules, 57(8):3515–3528, 2024. doi:10.1021/acs.macromol.4c00508. Christopher Kuenneth and Rampi Ramprasad. polybert: a chemical language model to enable fully machine-driven ultrafast polymer informatics. Nature communications, 14(1):4099, 2023. 17
UniPolymer
A P REPRINT
Chiho Kim, Rohit Batra, Lihua Chen, Huan Tran, and Rampi Ramprasad. Polymer design using genetic algorithm and machine learning. Computational Materials Science, 186:110067, 2021. Haoke Qiu, Xuepeng Qiu, Xuemin Dai, and Zhao-Yan Sun. Design of polyimides with targeted glass transition temperature using a graph neural network. Journal of Materials Chemistry C, 11(8):2930–2940, 2023. Yuanqi Du, Arian R Jamasb, Jeff Guo, Tianfan Fu, Charles Harris, Yingheng Wang, Chenru Duan, Pietro Liò, Philippe Schwaller, and Tom L Blundell. Machine learning-aided generative molecular design. Nature Machine Intelligence, 6(6):589–604, 2024. Di-Fan Liu, Yong-Xin Zhang, Wen-Zhuo Dong, Qi-Kun Feng, Shao-Long Zhong, and Zhi-Min Dang. High-temperature polymer dielectrics designed using an invertible molecular graph generative model. Journal of Chemical Information and Modeling, 63(24):7669–7675, 2023. M Li, XY Liu, JQ Qin, and Y Gu. Molecular dynamics simulation on glass transition temperature of isomeric polyimide. Express Polym. Lett, 3(10):665–675, 2009. Tyler B Martin and Debra J Audus. Emerging trends in machine learning: a polymer perspective. ACS Polymers Au, 3 (3):239–258, 2023. Ross Irwin, Spyridon Dimitriadis, Jiazhen He, and Esben Jannik Bjerrum. Chemformer: a pre-trained transformer for computational chemistry. Machine Learning: Science and Technology, 3(1):015022, 2022. Jerret Ross, Brian Belgodere, Vijil Chenthamarakshan, Inkit Padhi, Youssef Mroueh, and Payel Das. Large-scale chemical language representations capture molecular structure and properties. Nature Machine Intelligence, 4(12): 1256–1264, 2022. Juncai Li and Xiaofei Jiang. Mol-bert: an effective molecular representation with bert for molecular property prediction. Wireless Communications and Mobile Computing, 2021(1):7181815, 2021. Josep Arús-Pous, Simon Viet Johansson, Oleksii Prykhodko, Esben Jannik Bjerrum, Christian Tyrchan, Jean-Louis Reymond, Hongming Chen, and Ola Engkvist. Randomized smiles strings improve the quality of molecular generative models. Journal of cheminformatics, 11(1):71, 2019. Chae Eun Lee, Jin Sob Kim, Jin Hong Min, and Sung Won Han. Simson: simple contrastive learning of smiles for molecular property prediction. Bioinformatics, 41(5):btaf275, 2025. David Rogers and Mathew Hahn. Extended-connectivity fingerprints. Journal of chemical information and modeling, 50(5):742–754, 2010. Ao Shen, Mingzhi Yuan, Yingfan Ma, Jie Du, and Manning Wang. Complementary multi-modality molecular selfsupervised learning via non-overlapping masking for property prediction. Briefings in Bioinformatics, 25(4):bbae256, 2024. Shengli Jiang, Adji Bousso Dieng, and Michael A Webb. Property-guided generation of complex polymer topologies using variational autoencoders. npj Computational Materials, 10(1):139, 2024. Gabriel A Pinheiro, Juarez LF Da Silva, and Marcos G Quiles. Smiclr: contrastive learning on multiple molecular representations for semisupervised and unsupervised representation learning. Journal of Chemical Information and Modeling, 62(17):3948–3960, 2022. Jannis Born and Matteo Manica. Regression transformer enables concurrent sequence regression and generation for molecular language modelling. Nature Machine Intelligence, 5(4):432–444, 2023. Mario Krenn, Florian Häse, AkshatKumar Nigam, Pascal Friederich, and Alan Aspuru-Guzik. Self-referencing embedded strings (selfies): A 100% robust molecular string representation. Machine Learning: Science and Technology, 1(4):045024, 2020. AkshatKumar Nigam, Robert Pollice, Mario Krenn, Gabriel dos Passos Gomes, and Alan Aspuru-Guzik. Beyond generative models: superfast traversal, optimization, novelty, exploration and discovery (stoned) algorithm for molecules using selfies. Chemical science, 12(20):7079–7090, 2021. Ronald J Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1(2):270–280, 1989. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. Davide Chicco, Matthijs J Warrens, and Giuseppe Jurman. The coefficient of determination r-squared is more informative than smape, mae, mape, mse and rmse in regression analysis evaluation. Peerj computer science, 7:e623, 2021. Dávid Bajusz, Anita Rácz, and Károly Héberger. Why is tanimoto index an appropriate choice for fingerprint-based similarity calculations? Journal of cheminformatics, 7(1):20, 2015. 18
UniPolymer
A P REPRINT
Changwen Xu, Yuyang Wang, and Amir Barati Farimani. Transpolymer: a transformer-based language model for polymer property predictions. npj Computational Materials, 9(1):64, 2023. Masashi Ishii, Takuro Ito, and Koichi Sakamoto. Nims polymer database polyinfo (ii): Machine-readable standardization of polymer knowledge expression. Science and Technology of Advanced Materials: Methods, 4(1):2354651, 2024. doi:10.1080/27660400.2024.2354651. RDKit Contributors. Rdkit: Open-source cheminformatics software. https://www.rdkit.org/, 2026. Accessed: 2026-0728. Wenjia Huo, Boyang Liang, Xiang Wu, Zhenchang Zhang, Weichao Zhou, Haihong Wang, Xupeng Ran, Yaoyao Bai, and Rongrong Zheng. Prediction and interpretability study of the glass transition temperature of polyimide based on machine learning and molecular dynamics simulations. Polymers, 17(15):2083, 2025. doi:10.3390/polym17152083. James L. Suter, Werner A. Müller, Maxime Vassaux, Alexandros Anastasiou, Martin Simmons, David Tilbrook, and Peter V. Coveney. Rapid, accurate and reproducible prediction of the glass transition temperature using ensemblebased molecular dynamics simulation. Journal of Chemical Theory and Computation, 21(3):1405–1421, 2025. doi:10.1021/acs.jctc.4c01364. Didac Martí, Rémi Pétuya, Emanuele Bosoni, Anne-Claude Dublanchet, Stephan Mohr, and Fabien Léonforte. Predicting the glass transition temperature of biopolymers via high-throughput molecular dynamics simulations and machine learning. ACS Applied Polymer Materials, 6(8):4449–4461, 2024. doi:10.1021/acsapm.3c03040. Harish Gudla and Chao Zhang. How to determine glass transition temperature of polymer electrolytes from molecular dynamics simulations. The Journal of Physical Chemistry B, 128(43):10537–10540, 2024. doi:10.1021/acs.jpcb.4c06018.
19