arXiv:2604.06290v1 [cs.SE] 7 Apr 2026
All LCA models are wrong. Are some of them useful? Towards open computational LCA in ICT Vincent Corlay
David Bekri
Marie-Anne Lacroix
Mitsubishi Electric R&D Centre Europe Rennes, France [email protected]
iCoSys institute, HES-SO Fribourg, Switzerland [email protected]
Univ Rennes, CNRS, IRISA Rennes, France [email protected]
Maxime Pelcat
Maxime Péralta
Pierre-Yves Pichon
INSA Rennes Rennes, France [email protected]
Univ. Grenoble Alpes, CEA, List Grenoble, France [email protected]
Mitsubishi Electric R&D Centre Europe Rennes, France [email protected]
Leo Saillenfest
Olivier Weppe
Sébastien Rumley
Fraunhofer Institute for Reliability and Microintegration IZM Berlin, Germany [email protected]
INSA Rennes Rennes, France [email protected]
iCoSys institute, HES-SO Fribourg, Switzerland [email protected]
Abstract
Keywords
Life Cycle Assessment (LCA) is increasingly used to quantify and regulate the environmental impacts of Information and Communication Technology (ICT) systems. Since direct biosphere measurements are complicated to perform, we claim that the environmental impact assessment of ICT relies heavily on models. In this paper, we first revisit the fundamentals of LCA: we emphasize that ICT LCAs effectively form systems of models, and we argue that such systems require an “extra-high” level of carefulness in construction, calibration, integration, and interpretation. We then document how this level of rigor is challenging to achieve with current practices. This is illustrated with emblematic examples of model misuse and an analysis of structural challenges related to database choice, scope mismatches, opaque aggregation, and model integration. From this analysis, we derive four key requirements for credible ICT LCA: explicit model lineage, clearly defined model scope, end-to-end traceability, and managed non-obsolescence. Finally, we propose a framework that operationalizes these requirements using explicit dependency graphs, an open and versioned LCA-oriented model repository, automatic enforcement of integrity constraints, and a well-defined model taxonomy.
Life Cycle Assessment, Model-Based LCA, Model Lineage, Traceability, Model Scope, Non-Obsolescence, Open Data.
CCS Concepts • Computing methodologies → Modeling and simulation; • Information systems → Data management systems; • Software and its engineering → Software organization and properties; Software creation and management.
This work is licensed under a Creative Commons Attribution 4.0 International License. CF Companion ’26 , Catania, Italy © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2569-2/2026/05 https://doi.org/10.1145/3801488.3807497
ACM Reference Format: Vincent Corlay, David Bekri, Marie-Anne Lacroix, Maxime Pelcat, Maxime Péralta, Pierre-Yves Pichon, Leo Saillenfest, Olivier Weppe, and Sébastien Rumley. 2026. All LCA models are wrong. Are some of them useful? Towards open computational LCA in ICT. In Proceedings of the 23rd ACM International Conference on Computing Frontiers (CF Companion ’26 ), May 19–21, 2026, Catania, Italy. ACM, New York, NY, USA, 12 pages. https://doi.org/10.1145/ 3801488.3807497
1
Introduction
Information and Communication Technology (ICT) is clearly an industry with non-negligible environmental impacts. On one hand, ICT is a voracious energy consumer: the International Energy Agency (IEA) stated in 2025 that the ICT sector consumed 1000 TWh of electricity in 2023, about 4% of global electricity use [34, Sec. 2.4], and the equivalent to Japan’s yearly electricity demand [33]. On the other hand, fabricating ICT equipment undeniably pollutes the environment, for instance by emitting CO2 (hundreds of millions of tons of CO2, expected to be more than 1% of the total emissions by 2030 [25, p. 17]) and other greenhouse gases, or by depleting fresh water reserves (see [64] for wafer production). However, abandoning ICT altogether is neither realistic nor desirable, given the deep reliance of modern societies on information technologies. But individuals, organizations and societies have to begin to ask themselves “how much ICT is too much?". Or, expressed more technically, when do the environmental costs associated with ICT outweigh its net contribution to social welfare? This question lies at the heart of the notion of digital sufficiency [55], which frames the need to keep ICT uses within levels compatible with environmental constraints and societal goals (as illustrated by Kate
CF Companion ’26 , May 19–21, 2026, Catania, Italy
Corlay et al.
LCA curses
LCA requirements
Difficulty to list all biosphere flows Section 2.2
Modeling approach Section 2.4 Model lineage (R1)
Curse 1: Model validation is difficult. Section 2.4.5
LCA requirements
Model structure as a dependency graph (S1) Section 4.1 Model lineage (R1)
Scope of models (R2) Scope of models (R2) Traceability (R3)
Curse 2: Have to compose models. Section 2.5.2
Non-obsolescence (R4)
Figure 1: LCA curses and corresponding requirements (Section 2).
Raworth’s “doughnut” framework[52] and quantified by initiatives such as the SBTi targets [57]). To address this fundamental matter, individuals, organizations and societies need to be able to estimate the environmental cost of ICT (and also the social cost, but that is out of scope of this contribution). Because calculating the “net contribution to social welfare" is also challenging, and because not all entities are equally motivated to contribute to global social welfare, governments and regulators introduce limits, quotas and taxes to curb environmentally harmful industries [13, 62]. ICT will be no exception. However, in order to tell where a company should stop producing or using ICT equipment, to ensure quotas are respected, and to calculate taxes, governments and regulators also need to be able to estimate the environmental cost of ICT. Finally, companies are accountable to shareholders and customers, who increasingly demand measurable environmental progress. Organizations are also encouraged by insurers to reduce their environmental impact. Thus, an increasing number of companies publish environmental reports detailing their cumulative negative impacts and outlining the steps they are taking to mitigate them. In addition, they actively seek environmental labels and certifications to demonstrate their commitment and strengthen their credibility. But to award labels, to calculate relative environmental improvements, again, they need to be able to estimate the environmental cost of ICT. Environmental science experts have converged on standardized Life Cycle Assessment (LCA) methodologies to assess direct 1 environmental impacts. They aim to quantify environmental impacts across the entire life cycle of an ICT system, following a so-called cradle-to-grave approach, from raw material extraction and manufacturing, through operation, to end-of-life treatment. LCA methodologies aim to estimate environmental costs and impacts. But aiming is not succeeding; being the dominant approach does not always imply it is a valid approach. How certain are we, as a community, that LCA methodologies truly capture the impact 1 Indirect and systemic effects (for example rebound effects, induced demand, or long-
term socio-technical transformations) remain difficult to quantify in a robust and comparable way. Nevertheless, recent works, including a report by ADEME [2], try to assess both direct and indirect effects for concrete digital use cases, using consequence trees and design-oriented approaches.
LCA solutions
Traceability (R3)
Open and versioned repository Model maintenance (S2) Section 4.2 Automatic enforcement of integrity constraints (S3) Section 4.3
Non-obsolescence (R4) Well defined model taxonomy (S4) Section 4.4
Figure 2: LCA requirements and proposed solutions (Section 4).
they intend to measure? That the numbers they provide can be trusted to estimate the environmental costs of ICT, and this at the very moment we need them to calculate taxes or insurance fees, to introduce quotas and fix limits? In this paper, we aim to reassure the reader: we believe that LCA methods are fundamentally sound; they have been successfully applied and can produce reliable results (under strict conditions [43, 49]). However, they must be used appropriately and interpreted with full awareness of their limitations. One of the most significant limitations is that, when applied to complex systems, LCA theoretically requires a divide-and-conquer approach extending to the most elementary components of the system. In practice, this quickly becomes intractable (e.g., how many elementary components are there in a laptop, and what exactly qualifies as an elementary component?). Consequently, one must rely on models to make the analysis feasible. For instance, rather than explicitly modeling every via, trace, and solder joint on a printed circuit board (PCB) and associated production processes, one may rely on a parametric PCB model that estimates environmental impacts from a limited set of design parameters, such as board area, number of layers, substrate type, and surface-mount component density. However, the introduction of models inevitably brings additional uncertainty and potential error. By definition, models simplify reality and therefore cannot perfectly represent it. They rely on hypotheses and assumptions that shape, and sometimes constrain, the results. The difficulty is amplified in the case of ICT equipment, which constitutes highly complex systems in their own right. Conducting an LCA of a device like a smartphone therefore requires not a single model but an interconnected set of models: a genuine system of models. For such a system to yield meaningful and robust results, its components must be tightly coupled and its overall complexity carefully controlled. By resorting to modeling, we avoid tedious successive decompositions down to biosphere flows. Yet this simplification is not without trade-offs. Complexity does not disappear; it is displaced. The burden shifts from exhaustive structural
All LCA models are wrong. Are some of them useful? Towards open computational LCA in ICT
decomposition to the careful construction, calibration, integration, and interpretation of models. We co-authors have doubts that we have these systems of models under control. We doubt that we construct, calibrate, integrate and interpret models with the required care. If we do not, then our confidence in the correct application of LCA methods may be largely unwarranted. We are not alone in raising this concern. A recent review [38] by the IEA’s 4E Technology Collaboration Programme (4E TCP), which analyzes over 100 publications estimating data-center energy use, highlights two key observations: First, these estimates rely almost exclusively on models2 . Second, the resulting estimates exhibit very high variability across studies. George Box famously wrote in 1976 that all models are wrong, some of them are useful [8]. In doing so, he acknowledged not only the unavoidable inaccuracy of models, but also a more uncomfortable implication: usefulness is not automatic. Some models are simply not useful and shall not be used. This leads us to a deliberately provocative question: could it be that a substantial portion, perhaps even the majority, of our efforts to model the environmental impacts of ICT systems is not useful? And shall not be used? This paper is a call to reconsider how we are using and applying LCA methodologies and models in the ICT sector. If we, the sustainable-ICT research community, want to estimate correctly the environmental figures society awaits us to deliver, we must be more careful about how we use them (see Section 3 for examples). To the extent of our knowledge, this paper is the first to pinpoint and classify the requirements and computational limitations of ICT LCA as practiced today. We also propose solutions to overcome these limitations with a more structured computational-oriented approach. We believe such solutions are crucial to build a bridge between environmental sciences and ICT, and to lay solid foundations for digital systems ecodesign. This paper is organized in three parts. In Section 2, we go back to the fundamentals of LCA and good scientific practice. We begin by documenting how LCA has progressively become embedded in regulations and legislations (Subsection 2.1). In Subsection 2.2, we remind that LCA fundamentally consists of listing all biosphere flows, i.e., counting what each process takes from and returns to nature. We then discuss the difficulty of establishing this listing through direct measurements, notably because it means instrumenting factories, which is very intrusive (Subsection 2.3). Because of this difficulty, we posit that LCA in ICT has essentially become a modeling activity, and so we review the principles of modeling in Subsection 2.4. We exhibit in Subsection 2.5 that LCA in ICT suffers two curses that make it difficult to follow the canons of modeling: 1) models are intrinsically hard to validate, and 2) we have to compose models. In Section 2.6, we review methods that can quantify some categories of uncertainty in LCA, while noting that 2 They identify the following main categories of models used in the literature for energy
use estimation: (i) Bottom-up models: based on estimates of the installed server and IT equipment base, combined with equipment specifications (e.g., average server power consumption), equipment lifespans, and other energy-influencing attributes such as power-usage effectiveness. (ii) Aggregated totals: often described in the literature as top-down approaches, relying on national, regional, or organizational energy consumption data that are measured or estimated at an aggregate level. (iii) Temporal proxy extrapolation: starting from an initial base estimate obtained using one of the methods above, this approach combines high-level proxies and indicators (e.g., data traffic or energy-intensity assumptions) to extrapolate data-centre energy use under varying activity and efficiency-improvement scenarios.
CF Companion ’26 , May 19–21, 2026, Catania, Italy
such methods do not necessarily reduce uncertainty itself. Finally, we deduce from all previous subsections that LCA in ICT requires an extra-high level of carefulness in the way models are addressed. We propose a set of requirements (R1–R4) and list their benefits (B1–B8) in Subsection 2.7. In Section 3, we begin by listing (Subsection 3.1) several instances where this “extra-high level of carefulness” is absent, and we enumerate classes of challenges in current LCA practice in Subsection 3.2. We map each example and each structural challenge to the relevant requirements (R1–R4) and benefits (B1–B8). We then discuss the underlying causes, “the why” of natural model misuse (Subsection 3.3). Section 4 finally introduces a framework that operationalizes the requirements identified earlier by (i) organizing models as an explicit dependency graph (Section 4.1), (ii) embedding them in an open and versioned repository (Section 4.2), (iii) automatically enforcing integrity constraints (Section 4.3), and (iv) establishing a well-defined model taxonomy (Section 4.4). Figures 1 and 2 provide a structured overview of how current problems relate to the requirements and to the proposed solutions.
2
Model-based LCA: needs, definitions, design, curses, and requirements
LCA is the standard methodology used to assess the environmental impacts of products and services, as standardized by ISO 14040–44 [35, 36]. It can either follow a cradle-to-grave approach in order to cover the entire life cycle of a product from raw material extraction to end-of-life; or can focus on a part of the life cycle of a product such as the production phase, following for example a cradle-to-gate approach. LCA provides a comprehensive evaluation across multiple environmental impact categories, such as the 16 midpoint impact categories defined in the Environmental Footprint (EF) 3.1 method [3], such as climate change (GWP), water use, land use, and minerals and metals use (ADP).
2.1
LCA as a regulatory support tool
LCA is becoming increasingly critical because its use now extends beyond eco-design and internal decision-making to the regulatory domain. In addition to guiding designers toward more environmentally efficient solutions, LCA results are expected to support both reporting at the product leveland at the corporate-level - GHG Protocol standards (scopes 1/2/3) or the Corporate Sustainability Reporting Directive (CSRD, Directive (EU) 2022/2464). As a first step, Product Carbon Footprints (PCFs), built on top on ISO 14040/14044 LCA methodologies and focusing on Carbon Emissions only [1], are starting to be adopted by the industry, and can be reported in an Environmental Product Declaration (EPD). Product Environmental Footprints (PEFs), for their part, are similar to PCFs but focus on a broader set of environmental criteria beyond carbon emissions (typically EF 3.1 categories). To facilitate and harmonize reporting, technical committees are currently drafting several Product Category Rules (PCR) for PCFs (PCR for PCFs) and Product Environmental Footprint Category Rules (PEFCRs), i.e., LCA methodologies specifying the scope, boundary conditions, functional units and other requirements for certain product categories.
CF Companion ’26 , May 19–21, 2026, Catania, Italy
Corlay et al.
Biosphere in: Water Product A Product B Product C Process B
Product X Process A Product Y
Biosphere out: CO2 Figure 3: Illustration of an LCA process with input/output technosphere (products in the figure) and biosphere flows. While PCFs are a first step for using LCA methodologies for reporting, the EU Commission is favouring the multi-environmental -criteria solution (PEF), with the objective of supporting its regulatory framework, the Ecodesign for Sustainable Products Regulation (ESPR, Regulation (EU) 2024/1781). This framework will include the Digital Product Passport (DPP), an information system to disclose and manage product data over time. The DPP may for instance reference EPDs. Its technical implementation (data storage, archiving, persistence, reliability, integrity, interoperability, access rights management, APIs, etc.) is being supported by European standards developed by the CEN-CENELEC JTC24 in response to a standardisation request from the European Commission. As a result, the European Commission is expected to issue legislative acts requiring the publication of Product Environmental Footprints (PEF) based on LCA and setting minimum environmental criteria for ICT products placed on the EU market.
2.2
Theoretical LCA: listing all biosphere flows
A fundamental component of LCA is the Life Cycle Inventory (LCI). It consists of listing all processes involved in the system, each process taking as input technosphere and biosphere flows and producing technosphere and biosphere flows, as illustrated in Figure 3. Technosphere flows primarily act as proxies to connect processes within the product system, while the ultimate objective of the inventory is to account for all relevant biosphere flows. In the theoretical formulation of LCA, this would require an exhaustive and explicit representation of all such flows across the entire life cycle.
2.3
The difficulty of listing biosphere flows
Measurement issues in LCA differ fundamentally from those encountered in experimental sciences such as physics. 1. Environmental burdens cannot be inferred simply from inspection of the final product, as many impacts occur upstream throughout the production chain during manufacturing processes, which themselves often cause material losses. This makes impacts often far remote in space and time from the finished product. 2. Even when measurements are attempted during production, many relevant flows are intrinsically difficult to quantify, such as diffuse gaseous emissions or small material losses that are hard to capture with sufficient accuracy. 3. Meaningful measurements have to be performed directly within industrial facilities, which is operationally complex, costly, and rarely compatible with routine manufacturing constraints. Moreover, this raises the issue of allocating biosphere flows to the
specific product under study when production chains manufacture multiple devices. This challenge naturally leads to the introduction of models, which are discussed in the following subsection. 4. Access to such measurements is further limited by confidentiality and trade-secret concerns, which restrict the disclosure of detailed process data. 5. Even if every flow were perfectly measurable, the theoretical number of flows involved in an exhaustive life-cycle inventory would be intractable. For example to produce an electronic device, we should take into account the use of machines that enabled its manufacture, but also a certain proportion (again raising the allocation problem mentioned in (3)) of the process that enabled this machine manufacture, and so on. Therefore, an LCA always begins with the definition of system boundaries (which processes are included in the assessment or not). In addition, cut-off criteria are specified to justify the exclusion of processes or flows expected to be negligible, typically using thresholds defined by a norm or based on assessments. 6. Finally, in contrast to physics, where measurements are typically performed on closed or controlled systems and model validation can rely on repeatable experiments with clearly defined hypotheses, LCA operates on open, evolving systems whose conditions are hard to recreate ex post; hence, re-measurement and experimental replication are challenging, which further constrains model validation. As a consequence, direct measurement of biosphere flows is rarely feasible in practice, and environmental impacts must instead be inferred through models that combine heterogeneous data, assumptions, and proxies. As an illustration, the notion of simplified LCA is frequently adopted.
2.4
LCA as a modeling activity
2.4.1 Principle. When it is not possible to explicitly list all biosphere flows, practitioners instead measure observable quantities, such as the weight of a smartphone, its price, its RAM, the number of layers of a PCB [41], and use these measurements to infer, rather than directly measure, environmental impacts. This inference relies on formalized procedures that relate observable product characteristics to unobservable environmental metrics. These procedures are what we refer to as models. Technical characteristics are thus often used as proxies for environmental exchanges, either by being converted into quantities of biosphere flows or by serving as inputs to background impact models3 (see Section 4.1 for examples of model categories). This turns effectively LCA into a modeling activity. 2.4.2 From “wrong" to useful. As mentioned in the introduction, all models are wrong. Nevertheless, being “wrong” does not always imply being useless. Newtonian mechanics is formally incorrect 3We note that several notions commonly used in LCA, such as PCRs, foreground and
background systems, primary and secondary data, and models, refer to distinct but interrelated concepts. PCRs define how an assessment should be conducted for a given product category, but they neither provide data nor constitute models themselves. An LCA is implemented through models related to the foreground system (which are specific to and often controlled by the studied system) with models related to background systems (outside the direct control of the studies). Models in both foreground and background systems may be calibrated based on primary data (direct measurement for the system under study) or secondary data (external sources).
All LCA models are wrong. Are some of them useful? Towards open computational LCA in ICT
when compared to relativity or quantum mechanics, yet it remains sufficiently accurate within a well-defined domain of validity to support most engineering applications. Engineers continue to rely on such models because they are useful answers to specific questions, given acceptable error margins. What we need to know is “how wrong?", whether they are useful for the specific questions we seek to answer, and whether we can make them more useful through validation and transparency. 2.4.3 Model design. Two main approaches can be distinguished when designing LCA models. (i) Data-driven or empirical modeling relies on direct measurements of biosphere flows (primary data) obtained under specific experimental or industrial conditions, often costly, difficult to perform, and limited in scope as previously mentioned in Subsection 2.3. These measurements are then used to construct empirical relationships between observable product or process characteristics and environmental exchanges, for instance through linear or polynomial regressions, or more complex techniques such as neural networks. While such models can be effective within the range of observed data, they are exposed to well-known risks: they may fail when extrapolated beyond the calibration domain, they can overfit noisy or sparse data, and they often exhibit limited interpretability, making it difficult to assess their assumptions or domain of validity. (ii) Physics-based or “first-principles" modeling (axiomatic) instead derives environmental exchanges from descriptions grounded in physical, chemical, or thermodynamic laws. These models are typically more interpretable and better suited for extrapolation. The number of free parameters that must be specified or calibrated is usually reduced compared with purely empirical approaches. Nevertheless, this approach introduces the risk of model mismatch: if the assumed model structure, namely the set of governing equations, boundary conditions, and simplifying assumptions used to represent the system, is incomplete or inaccurate [58], the constrained parameterization may bias inferred environmental exchanges and impacts (see also Section 2.6). In practice, LCA studies often combine both approaches. 2.4.4 Model composition. Beyond empirical and physics-based designs, practitioners can create new models by composing existing ones: an effective divide-and-conquer for high-complexity systems. However, composition comes with strict guardrails: (i) constituent sub-models must be validated (see the following Subsection 2.4.5) within their respective domains; (ii) their assumptions must be mutually compatible (units, scopes, boundary conditions, allocation choices, operating regimes) and integration can induce second-order interactions (feedbacks, cross-terms, double counting) that were absent or negligible at the unit level. (iii) Crucially, invalidation must propagate: if any sub-model is later found erroneous (e.g., due to a defective probe or flawed calibration), the composite model inherits that invalidation, at least for the portions that depend on the faulty component. Ideally, composite models would themselves be validated as valid composites according to some rules. 2.4.5 Model validation is necessary. Regardless of the modeling approach adopted, validation is indispensable.
CF Companion ’26 , May 19–21, 2026, Catania, Italy
(i) In both empirical and physics-based approaches, a model must ultimately be confronted with observations (preferably not the ones used to train the models) to establish its credibility. This is a key principle of the scientific method. Even a theory as foundational as the one presented in Philosophiæ Naturalis Principia Mathematica by Isaac Newton derived its scientific legitimacy from systematic comparison between predictions and measured phenomena; anchoring a model to reality therefore inevitably requires measurement. (ii) In some cases, however, models are proposed before such validation can be fully achieved. These models must then be regarded as conjectural: they may provide useful insights or support exploratory analyses, but they remain exposed to potential invalidation by future experimental evidence.
2.5
The curses of model-based LCA
Summarizing Subsections 2.2, 2.3, and 2.4, we posit that LCA in ICT is subject to the two following curses. 2.5.1 Curse 1: Model validation is difficult. As discussed in Section 2.3, measurement in LCA is highly challenging. As a result, assessing the validity of an LCA model is both intrinsically difficult and very costly. Even for seemingly simple objects, such as a single screw or an electronic component, directly validating predicted environmental impacts through measurements is often impractical and this difficulty increases rapidly with system complexity. Because comprehensive validation is so resource-intensive, each direct measurement of biosphere flows (typically primary data) has an exceptionally high value and must therefore be produced, documented, and referenced with extreme care. 2.5.2 Curse 2: We have to compose models. Because of the aforementioned complexity of ICT systems, we cannot avoid composing models. Nevertheless, when composing models in LCA, it is even less likely to have access to measurements. Moreover, not all models evolve at the same pace in LCA: some are regularly updated to reflect new knowledge or technologies, while others remain unchanged for long periods of time. In certain cases, several competing models coexist (see examples in Section 3.2). As a result, the requirements of Section 2.4.4 are even more relevant. For instance, when performing an LCA with models based on secondary data, ensuring consistency with fundamental conservation principles, e.g., energy conservation (first law of thermodynamics) and mass balance, is challenging.
2.6
LCA uncertainty assessment
The observations outlined above make uncertainty an inherent feature of LCA results. Different categories of uncertainty have long been identified in the LCA literature [30]4 . For instance, one can distinguish parameter uncertainty, which relates to uncertainty in numerical input values, from model uncertainty (closely related to model mismatch as discussed in Section 2.4.3), which arises from methodological choices. Contrary to parameter uncertainty, model uncertainty cannot always be addressed through probabilistic parameter variation alone. 4 The authors also highlighted the limited consideration of uncertainty in many LCA
studies, noting that “it is amazing that this interest has not been natural since the development of LCA and the rise of its use”.
CF Companion ’26 , May 19–21, 2026, Catania, Italy
The uncertainty assessment methods recalled below, also relying on explicit modeling assumptions, aim to quantify uncertainty. The requirements proposed in the following Section 2.7 go beyond quantification: they also aim to mitigate uncertainty when possible and, more generally, to limit model misuse through the adoption of adapted practices. Uncertainty assessment methods Parameter uncertainty is usually modeled by assigning a probability distribution to a parameter. It may pertain either to the data itself or to the mapping of a secondary dataset (see “scope mismatches" in Section 3.2), itself having a potential uncertainty. The distributions are then combined to model the resulting uncertainty. When measurement data or statistical evidence are unavailable, the probability distribution may be inferred through a model. A widely used model is the PEDIGREE matrix, where an expert choses scores instead of an explicit distribution. Initially proposed by Weidema and Wesnæs [66], this method translates qualitative data quality indicators5 into quantitative uncertainty estimates, i.e., a probability distribution. The PEDIGREE approach is implemented in mainstream LCA databases and softwares [12, 65]. We emphasize that the PEDIGREE approach is itself a modeling layer and therefore entails assumptions, model error, and potential out-of-scope use. Ecoinvent acknowledges that this model requires continued improvement and validation, including comparisons with uncertainty estimated from empirical data [11, 65, Chap. 10]. Alternatively, Bayesian methods, unlike the PEDIGREE approach, allow uncertainty to be directly informed and updated using new measurement data or alternative data source [4].
2.7
Induced requirements and their benefits
We deduce that the curses of Section 2.5 induce the following requirements. Figure 1 illustrates the connections between the curses and the proposed requirements. For each requirement, we discuss the benefits associated with its satisfaction. Model lineage (R1). Each model should come with an explicit description of its ancestry (the sub-models it relies upon) and siblings (the models that use it) and dependencies. This includes both its links to empirical measurements (when they exist) and its links to upstream models from which it inherits assumptions or structure. In other words, we should be able to distinguish models that have “royal blood” (i.e. a clear chain back to observations) from models that rest on conjectural or weakly justified assumptions. Benefit: Making this lineage explicit allows practitioners to assess how much empirical support underpins a given result (B1) and to identify where uncertainty or speculation enters the chain (B2). It also helps prevent model washing (B3), whereby a weakly supported or poorly validated model is embedded within, and thereby legitimized by a more reputable model. Scope of models (R2). A model is never valid “in general”; it is valid only within a specific domain defined by hypotheses on units, system boundaries, operating conditions, and technological context. Composing models therefore requires checking that their 5 Inventory data are evaluated according to criteria such as reliability, completeness,
temporal correlation, geographical correlation, and technological correlation. Each criterion is scored, and the scores are mapped to uncertainty factors, which are subsequently combined to derive a probability distribution for each parameter, most often assumed to be lognormal.
Corlay et al.
scopes are compatible: a model that outputs a quantity in one unit cannot be used as if it directly provided another (e.g. treating 1 L of water as 1 kg at 1 bar and 4◦ C), and simplifying assumptions must not be silently violated in downstream uses. Benefit: By making scope definitions explicit, practitioners can identify incompatible assumptions early, preventing composite models from appearing coherent while resting on inconsistent premises (B4). It also prevents out-of-domain reuse (B5). Traceability (R3). When a model is used, it should be possible to explain what it does, how it was built, and how a given numerical result was obtained. In practice, this means that calculations can be “unwound” into a dependency tree that closely resembles the idealized process graph introduced earlier. For example, when reporting an aggregate figure such as “data centres consume on the order of 400 TWh of electricity per year” [18, 34], it should in principle be possible to recompute this value end-to-end from documented data, code, and modeling choices, ideally made available in a public, versioned repository. Note that several scientific papers do publish their inventories and models in an open-source manner; see, for instance, [19, 40, 42, 46–48, 69, 70]. Benefit: Traceability thus connects impact numbers back to models (B6) and, ultimately, to observations (also through lineage) (B1). It supports explanation, replication, and auditing of results (B7). Non-obsolescence (R4). Finally, models and databases evolve over time as new data, technologies, and methodological insights appear. A credible framework must therefore organize this evolution rather than ignore it. Benefit: Non-obsolescence preserves the interpretability of published results in light of current knowledge: past model versions stay archived and citable (B6), but it is also possible to identify when a result depends on outdated components and how it would change under updated models (B8, B4). In substance, requirements (R1–R4) articulate what good scientific practice demands.
3
Patterns and causes of reduced methodological rigor with respect to the requirements
This section illustrates how methodological rigor is often relaxed in practice when using models and why this happens, even in good faith. We first present a few emblematic examples of model misuse (both within and outside the field of LCA), then summarize recurring structural challenges in current LCA practice, and finally discuss why these misuses are “natural” in the absence of the aforementioned explicit safeguards: model lineage (R1), scope (R2), traceability (R3), and non-obsolescence (R4).
3.1
Illustrative example of model abuse
Reinhart and Rogoff. In [53], Reinhart and Rogoff reported that public debt levels above 90% of GDP are associated with strongly reduced or negative growth. This finding was quickly interpreted as a quasi-rule justifying strict austerity policies. Herndon et al. [31] later uncovered spreadsheet errors and questionable methodological choices that materially affected the results, but by then the original claim had already influenced policy. The failure did not stem solely from a flawed calculation; it came from elevating a
All LCA models are wrong. Are some of them useful? Towards open computational LCA in ICT
fragile empirical pattern to a law, ignoring uncertainty and context. Possible improvement: (B2) via (R1) – “Where does uncertainty/ speculation enter?”, (B7) via (R3) – “Supporting explanation, replication, and auditing”. Gaussian copulas. Gaussian copula models were widely used to price complex credit products such as CDOs. While mathematically coherent, they relied on assumptions of relatively stable correlations and limited tail dependence. In practice, they were applied beyond their validity domain, including for systemic risk assessment. When correlations surged during the 2007–2008 crisis, risks were underestimated: the issue was less the formula itself than the uncritical reuse of a convenient model outside the conditions under which it had been validated [44]. Possible improvement: (B2) via (R1) – “Where does uncertainty/ speculation enter?”, (B5) via (R2) – “Preventing out-of-domain reuse”. Footprint of an email. Popular figures such as “4–50 g CO2 per email” conflated marginal and average impacts by mixing the fixed energy consumption of devices, data centers, and networks with the incremental cost of sending a single message. Later clarifications emphasized that, in most realistic contexts, the marginal footprint of an attachment-free email is orders of magnitude smaller than these headline numbers [5, 6, 32, 60]. Mike Berners-Lee posted the following in 2020 on Twitter: “To clarify, following FT and BBC pieces, the carbon footprint of sending an email is trivial. Looks like UK gov has misused a press release from OVO that in turn used estimates from the 2010 version of my book ’How Bad Are Bananas?’ (now updated)." Here, a rough illustrative model was reinterpreted as a precise and context-independent fact, despite its correction. Possible improvement: (B2) via (R1) – “Where does uncertainty/ speculation enter?”, (B5) via (R2) – “Preventing out-of-domain reuse”. Ecologits. Ecologits [24] is an environmental footprint calculator for AI models. Even with a documented methodology, it is practically impossible to track all hardware actually used for inference. More critically, during the communication of results based on this tool by some authors of this paper [29], the calculator and its underlying database were updated, significantly changing the impacts. This situation raises several issues: (i) a lack of transparency regarding updates to the database and the parameters used in the calculations; (ii) the impossibility of accessing previous versions of the database in order to reproduce or compare past results; and (iii) the absence of traceability of dependencies and of the downstream effects that such updates may have on existing results or on other models relying on this calculator. Possible improvement: (B3) via (R1) – “Avoiding model washing” (by preventing potentially weakly supported sub-models used by those trusted databases from being legitimized simply because they are embedded), (B6) via (R3) – “Connecting numbers back to models and data” (B7) via (R3) – “Supporting explanation, replication, and auditing”, (B8) via (R4) – “Handling outdated components and evolution over time”. NegaOctet. The NegaOctet project [63] provides an illustration of traceability and maintenance issues in LCA data, used as background by many other studies. As a time-limited research project, it did not guarantee long-term data maintenance. According to [63],
CF Companion ’26 , May 19–21, 2026, Catania, Italy
the database is no longer commercialized by the consortium. Only a fraction of it has been publicly disclosed and/or transferred to other databases (such as the French ADEME’s Base Empreinte). Based on the publicly available information to date, the current state of access to the data and knowledge contained in NegaOctet, as well as whether this data can be maintained over time, remains unclear. This situation raises concerns regarding downstream studies that reference this database, thus without possible access to the underlying data by the reader (e.g., a study by Capgemini relying on NegaOctet to estimate GPU impacts [16], as well as the recent A100 GPU study [19], which specifies a database version of NegaOctet (2022) without indicating how the database can be accessed). Possible improvement: (B3) via (R1) – “Avoiding model washing” (as for Ecologits), (B6) via (R3) – “Connecting numbers back to models and data”, (B7) via (R3) – “Supporting explanation, replication, and auditing”, (B8) via (R4) – “Handling outdated components and evolution over time”. These examples mirror potential model misuse: fragile, contextdependent models are reused as if they were robust, general laws, with limited visibility on their lineage, assumptions, and updates.
3.2
Structural challenges in current LCA practice
Beyond isolated anecdotes, several recurring structural issues in LCA practice make reduced rigor likely. Some of them are direct instances of unmet requirements (e.g., R2 and the scope mismatches discussed below), which we illustrate below using concrete LCA references. Sensitivity to database choice. Impact results can substantially vary with the chosen background database, sometimes as much as with the impact assessment method itself. Recent work on GPUs [19] and comparative EPD studies [39] show that changing only the background data can significantly alter conclusions. Possible improvement: (B1) via (R1) - “How much empirical support?”. Scope mismatches. Secondary datasets often describe generic technologies, system boundaries, or use conditions that only partially match the studied product. Such mismatches are difficult to detect, yet they can affect the results when an “almost” relevant dataset is reused outside its original scope [7, 9, 54, 67]. The paper [49, Sec. 6] on the impacts of electric vehicles, carefully considering methodological flaws, also highlights the importance of time scope (related with the following “changing technologies" item). Possible improvement: (B4) via (R2, R4) – “Detecting incompatible assumptions”, (B5) via (R2) – “Preventing out-of-domain reuse”. Non-explainability of aggregated indicators. LCA typically aggregates thousands of inventory flows into a small number of midpoint indicators. This many-to-few mapping is fundamentally non-invertible: impact scores alone rarely allow experts to reconstruct underlying processes or assumptions. EPDs and background reports [50], as well as recent AI-related LCAs [17, 56], exemplify how impact results are often published without the inventory or models needed for meaningful interpretation or reuse. Possible improvement: (B6) via (R3) – “Connecting numbers back to models and data”, (B7) via (R3) – “Supporting explanation, replication, and auditing”.
CF Companion ’26 , May 19–21, 2026, Catania, Italy
Integration barriers for research models. Many academically proposed models are well documented but difficult to integrate into operational LCA workflows. They may be provided as stand-alone scripts or spreadsheets [26, 27, 42, 69], requiring substantial re-implementation and interpretation before they can be composed with other models. This (i) breaks the lineage or makes it very thin, (ii) exposes to errors in the re-implementation and interpretation process (iii) exposes one to "fantom obsolescence" if the original model reimplemented is found obsolete, among others. Possible improvement: (B2) via (R1) – “Where does uncertainty/ speculation enter?”, (B5) via (R2) – “Preventing out-of-domain reuse”. Problematic proxies and changing technologies. Simplifying proxies (e.g., scaling flash memory impacts by die area or capacity) may become invalid when technology changes, as with the transition from planar to 3D NAND [68]. See also [49, Sec. 6] (time scope) for electrical vehicles. Without explicit documentation of domains of validity, such proxies may continue to be applied long after their original assumptions are violated. Possible improvement: (B2) via (R1) – “Where does uncertainty/ speculation enter?”, (B4) via (R2, R4) – “Detecting incompatible assumptions”, (B8) via (R4) – “Handling outdated components and evolution over time”. Accounting conventions and truncation. Multiple coexisting conventions for Scope 2 and Scope 3 emissions, or spend-based versus activity-based accounting, create additional layers of modeling choices. These conventions are often treated as interchangeable, even though they embed different system boundaries and truncation patterns. Convention choices can impact the whole chain: scope 3 emissions rely, almost by definition, on the models of the suppliers. Possible improvement: (B4) via (R2, R4) – “Detecting incompatible assumptions”. Correct interpretation of the impact results. [37] highlights how sustainability indicators are often misinterpreted once detached from their methodological context, in particular through the frequent confusion between attributional and consequential (marginal) reasoning. Numbers derived from average, system-level assessments are commonly used as if they represented the marginal impact of an additional use, leading to erroneous conclusions. More generally, aggregated indicators are difficult for non-experts to interpret, especially when they obscure whether impacts are localized or globally distributed. In a similar vein, [21] points an emerging gap between analysis-LCA and message-LCA, where communicated conclusions are “partly in conflict with some key features and principles of LCA". In the terms of our framework, such gaps amount to breaking the chain between models: LCA results are reused without preserving the lineage, scope, and uncertainty information encoded in the underlying models, so that the “message” may no longer reflects the conditions under which the analysis is valid. Possible improvement: (B5) via (R2) – “Preventing out-of-domain reuse”.
3.3
Why misuse is “natural”
Model misuse rarely stems from bad intentions. Early models are often developed under conditions of genuine ignorance: data are
Corlay et al.
scarce, measurement is costly, and the main objective is to obtain an order-of-magnitude estimate that can guide first decisions. This type of ignorance is commonly referred to as epistemic uncertainty, which arises from incomplete knowledge and is, in principle, reducible, as opposed to aleatory uncertainty, which reflects inherent variability. Engineers routinely navigate a precision–cost trade-off: coarse models with few inputs are attractive when time or data are limited, whereas more detailed models demand extensive data collection and expertise. Over time, however, the assumptions and scope of these initial models tend to be forgotten, especially when they are reused by people who were not involved in their design. In LCA, this effect is amplified by the strong dependence on secondary data and by the diversity of goals and functional units. The same physical system may be modeled differently for policy design, regulatory compliance, or product comparison, with different system boundaries and data requirements. When results and models are transferred across these contexts without explicit scope and lineage information, methodological rigor degrades almost inevitably. As highlighted by the benefits listed in Section 2.7, the proposed requirements and associated benefits can be read as a direct response to this natural drift and to the misuses illustrated above. The next section presents practical solutions to meet these requirements more consistently.
4
Proposed framework
The previous sections have identified four key requirements for more rigorous, explainable, and maintainable model-based LCA: model lineage (R1), scope of models (R2), traceability (R3), and non-obsolescence (R4). We now outline a computational-oriented framework that operationalizes these requirements by (i) structuring models as an explicit dependency graph (S1), (ii) embedding them in an open, versioned repository with software-like governance (S2), (iii) automatic enforcement of integrity constraints (S3), and (iv) establishing a well-defined model taxonomy (S4). Figure 2 illustrates the connections between the requirements and the proposed solutions. These principles are intended to serve as design guidelines of a concrete database implementation, such as ElecImpact [23], a project of a new open and collaborative LCA database for electronics developed by the GDR DEFIE working group [22].
4.1
Model structure as a dependency graph
A substantially more explicit model dependency graph is required than what is typically available today. Such a graph should formally and explicitely represent models (and sub-models) as nodes, and their dependencies as edges, making it possible to identify which results must be re-evaluated when an upstream component is corrected, updated, or very importantly invalidated. Hence, systematic propagation of invalidation along this graph is essential for maintaining coherence across complex, compositional model ecosystems. This would mirror the logic of Common Vulnerabilities and Exposures (CVEs) used in software development. CVEs is a standardized system for identifying and cataloguing publicly known cybersecurity vulnerabilities [45]. Ideally, a fully validated LCA model 𝑀 would be confronted directly with measurements. Since this is rarely feasible, a principled
All LCA models are wrong. Are some of them useful? Towards open computational LCA in ICT
alternative is to require that the sub-models used to construct 𝑀 be individually validated within documented domains of validity and with quantified uncertainty. The composite model 𝑀 must then (i) operate strictly within these domains, its context, and (ii) explicitly inherit and encode the assumptions and applicability criteria of its sub-models. This naturally leads to a hierarchy of models, where models can inherit from and be composed with one another. More diffuse relationships, where a model is considered validated by analogy with another model, must also be made explicit. Cycles in this dependency structure may reveal “incestuous” validation patterns with no solid empirical foundation; at minimum, such cycles should be identified and documented. Concretely, each proposed model should document at least: on which other models it depends, and whether these underlying models are validated and within which domains; how its own domain of validity is defined (functional unit, system boundaries, technological scope, temporal and geographical coverage); how it can be validated or recalibrated if new data become available. Proprietary or paid models that hide their internal structure and ancestry run counter to this vision, as they break dependency links and prevent tracing results back to empirical observations. This model structure supports requirements: (R1) model lineage, (R2) scope of models, (R3) traceability, and (R4) non-obsolescence.
4.2
Open and versioned repository
Implementing such a dependency graph requires infrastructure similar to that used in modern software engineering. To take another software development analogy: An LCA database can be compared to a compiled Docker image or to a compiled binary: a ready-to-use artifact encapsulating a complex model. However, without the equivalent of a Dockerfile, or of the source code, that it, without an explicit recipe describing the model’s structure, assumptions, data sources, and parameterization, the model is not fully reproducible, inspectable, or explainable. For instance, large software organizations increasingly rely on a monorepo approach, where source code, dependencies, and build rules are maintained in a single, versioned repository [51]. Package managers (e.g. those used in Python, Linux distributions, or container ecosystems) then resolve dependencies, enforce compatibility constraints, and make it possible to reconstruct precise environments. Another useful point of comparison can be found in open-source software engineering practices, where the management of complex dependency structures has long been addressed through shared repositories, package managers, and strict versioning conventions. In the Python ecosystem, for instance, models and libraries are commonly developed in public version-controlled repositories (e.g., GitHub) and distributed through centralized registries such as the Python Package Index (PyPI), with tools like pip enforcing explicit dependency declarations and compatibility constraints. Similar infrastructures exist in other communities, such as Maven Central for Java, where artifact versioning and dependency resolution are treated as first-class concerns. Beyond their technical role, these ecosystems embody a culture of openness, reproducibility, and compatibility management, in which models are expected to declare
CF Companion ’26 , May 19–21, 2026, Catania, Italy
their interfaces, assumptions, and supported versions. And as mentioned above, these ecosystems have in the last 10 years integrated mechanism to track the CVEs across the web of dependencies. LCA modeling should move toward a similar paradigm: a model repository in which every model is versioned, its ancestry is recorded, and its build instructions are part of the public record. Such a repository, equipped with explicit versioning and dependency tracking, is the natural place to enforce the non-obsolescence requirement (R4) via efficient model maintenance: deprecated models remain archived and citable, but new versions can be linked, compared, and substituted in a controlled way. This repository approach supports requirements: (R1) model lineage, (R3) traceability, and (R4) non-obsolescence.
4.3
Automatic enforcement of integrity constraints
Integrity constraints, both across models and for compliance with mandatory methodology, should be enforced automatically. It mirrors integrity constraints in entity–relationship database schemas, and the way modern dependency managers verify and reconcile constraints to resolve dependency conflicts before producing a consistent build. Typical checks could include the presence of mandatory product and process model parameters (see next subsection), adherence to allowed parent–child structures, conformance with relevant PCR, and schema-level assertions such as unit consistency, valid ranges, and type safety. A pragmatic two-pass pipeline may be effective in practice. Pass 1 loads sources, instantiates models and data classes, populates a local database. Pass 2 aggregates and propagates constraints across the model graph, checks integrity constraints, and, on success, records standardized model validation. This automatic verification supports requirements: (R3) traceability and (R4) non-obsolescence.
4.4
Well-defined model taxonomy
In model-based LCA, different types of models play distinct roles and may be categorized accordingly. At a minimum, we think that the following categories are useful. Product and process models, which specify typical technical parameters for a product or process category (e.g. a server product, PCB product, chip product, or assembly process, use phase process). Some instances of typed nodes may claim average or generic configurations, while others may describe more specific cases. Similarly, defining subtypes of nodes enables finer granularity in the representation of products and processes while establishing a clear hierarchy. As illustrated in Figure 3, products and processes can be represented as connected nodes, with the LCI forming a graph. Some nodes may be reused across multiple trees (i.e., across multiple LCIs). Impact models, which compute environmental impacts for a product or process directly from product or process parameters, which may themselves be obtained through parameter conversion models. As a result, these models may bypass the intermediate step of explicitly enumerating biosphere flows for the product under consideration. We distinguish three sub-classes, depending on whether parameters are mapped directly to environmental impact
CF Companion ’26 , May 19–21, 2026, Catania, Italy
indicators (e.g., midpoint indicators), to biosphere-flow quantities, or to operational performance–resource proxies: • Midpoint impact models, which map product or process parameters directly to midpoint impact indicators, thereby bypassing explicit biosphere-flow modeling. • Product-flow models, which infer reference-flow quantities (and, more generally, biosphere-flow quantities) directly from technical parameters (e.g., mass from dimensions, number of wafers from die count). • Computational handprint–footprint models, which rely on operational proxy metrics, for instance, computational performance or complexity measures (e.g., GFLOPs) and energy-use indicators (e.g., Joules), to approximate, respectively, the service provided and its associated resource use. Parameter conversion models, which map technical characteristics of a product to other characteristics. These mappings may describe relationships between parent and child products, infer intra-product characteristics (i.e., relationships between different parameters of the same product), or operate more generally between products and processes. Allocation models, which distribute impacts or flows among co-products, recycled materials, or services according to specified rules (e.g., module D-type allocation, cut-off or avoided burden rules, market-based allocations for Scope 3). Uncertainty models, which infer probability distributions from input parameters to support stochastic analyses (e.g., Monte Carlo simulation). For example, the PEDIGREE method is an uncertainty model (see Section 2.6). Making these categories explicit helps practitioners understand what each model does, how models can be composed, and where modeling choices enter the calculation. For instance, a given product node in a tree graph can often be modeled in two complementary ways. On one hand, it may be linked to an impact model that uses a small set of high-level product parameters; on the other hand, it can be decomposed into a more detailed sub-tree of sub-products and processes, each with their own parameters and models. Within the same dependency graph, an analyst can either use the impact model or traverse the detailed sub-tree to recompute impacts bottom-up. By making these alternative paths explicit, the framework exposes and documents the different approximations and assumptions. This taxonomy supports requirements: (R1) model lineage and (R2) Scope of models.
5
Conclusion
A direct critique of current ICT service LCAs [14] highlights the prevalence of simplistic assumptions, inadequate methodologies, and unsupported extrapolations. In this paper, we have complemented this perspective by arguing that LCA for ICT systems has effectively become a modeling exercise, in which complex systems of models replace direct observation of biosphere flows. Because these models are composite, hierarchical, and only partially validated, ICT LCA requires an “extra-high” level of carefulness in the way models are constructed, calibrated, integrated, and interpreted. Our review of current practices suggests that this level of rigor is not always attained: assumptions and scopes may be
Corlay et al.
under-documented, dependencies may be opaque, and updates or corrections difficult to trace, so that apparently precise impact figures may rest on fragile foundations. From emblematic examples and structural challenges, we have distilled four key requirements for credible model-based LCA in ICT: explicit model lineage, clearly defined model scope, end-to-end traceability, and managed non-obsolescence. We then outlined a framework that operationalizes these requirements by representing models as explicit dependency graphs, embedding them in an open, versioned LCA model repository, automatically enforcing integrity constraints, and establishing a well-defined model taxonomy. More broadly, verification tools are also being developed in other scientific fields, for instance, to detect problematic publications and mitigate their downstream effects (e.g., biases in meta-analyses) through living systematic reviews [10, 15, 28], as well as to enable computer-assisted formal verification via proof assistants such as Lean or Rocq, which have already revealed flaws in widely cited results [59, 61]. In this respect, adopting a skeptical attitude of the kind advocated by Richard Feynman [20], critical questioning, transparency, and a constant effort to distinguish robust results from convenient but fragile claims, are concrete steps toward reliable ICT LCAs.
References [1] 2018. Greenhouse gases — Carbon footprint of products — Requirements and guidelines for quantification. https://www.iso.org/standard/71206.html 2nd edition. [2] ADEME. 2025. Environmental Assessment of the Direct and Indirect Effects of Digital Technology on Use Cases. Study / Research Report. Agence de la transition écologique (ADEME). https://librairie.ademe.fr/societe-etpolitiques-publiques/8073-9720-environmental-assessment-of-the-direct-andindirect-effects-of-digital-technology-on-use-cases.html Models direct and indirect environmental effects of digital solutions using consequence trees and design-oriented approaches for concrete use cases. [3] Susanna Andreasi Bassi, F Biganzoli, N Ferrara, A Amadei, A Valente, S Sala, and F Ardente. 2023. Updated characterisation and normalisation factors for the Environmental Footprint 3.1 method. Publications Office of the European Union: Luxembourg (2023). [4] II Ascough, James C., Holger R. Maier, Jarin K. Ravalico, and Marcus W. Strudley. 2008. Combining Bayesian methods and Monte Carlo simulation for analysis of uncertainty in life cycle assessment. Environmental Modelling & Software 23, 10–11 (2008), 1308–1319. doi:10.1016/j.envsoft.2008.01.002 [5] Mike Berners-Lee. 2010. How Bad Are Bananas? The Carbon Footprint of Everything (1st ed.). Profile Books, London. Later revised and expanded editions published in 2015 and 2020. [6] Mike Berners-Lee. 2020. How Bad Are Bananas? The Carbon Footprint of Everything (updated edition ed.). Profile Books, London. [7] Mathilde Billaud, Christian Clemm, David Sánchez, Marina Proske, Marvin Jügel, Lutz Stobbe, Nils F Nissen, and Martin Schneider-Ramelow. 2023. ICs as drivers of ICT carbon footprint: an approach to more accurate die size assessment. In Going Green CARE INNOVATION 2023. [8] George E. P. Box. 1976. Science and Statistics. J. Amer. Statist. Assoc. 71, 356 (1976), 791–799. doi:10.1080/01621459.1976.10480949 [9] Andreas Busa and Malcolm Hegeman. 2019. Life Cycle Assessment of Dell R740. https://www.delltechnologies.com/asset/en-us/products/servers/technicalsupport/Full_LCA_Dell_R740.pdf. Accessed: 2025-02-19. [10] Guillaume Cabanac, Cyril Labbé, and Alexander Magazinov. 2022. The “Problematic Paper Screener” Automatically Selects Suspect Publications for PostPublication (Re)Assessment. arXiv preprint arXiv:2210.04895 (Oct. 2022). doi:10. 48550/arXiv.2210.04895 [11] Andreas Ciroth. 2012. Refining the Pedigree Matrix Approach in ecoinvent: Towards Empirical Uncertainty Factors. Technical Report. GreenDelta, Berlin, Germany. https://lca-net.com/files/Pedigree_report_final_May2012.pdf Version 7.1. [12] Andreas Ciroth, Stéphanie Muller, Bo Pedersen Weidema, and Pascal Lesage. 2016. Empirically based uncertainty factors for the pedigree matrix in ecoinvent. The International Journal of Life Cycle Assessment 21, 9 (2016), 1338–1348. doi:10.1007/ s11367-013-0670-5 Published online 20 December 2013; part of the ecoinvent database v3 documentation.
All LCA models are wrong. Are some of them useful? Towards open computational LCA in ICT [13] European Commission. 2025. 2025 Environmental Implementation Review. Technical Report. [14] Vlad C. Coroamă, Pernilla Bergmark, Mattias Höjer, and Jens Malmodin. 2020. A Methodology for Assessing the Environmental Effects Induced by ICT Services. Part I: Single Services. In Proceedings of the 7th International Conference on ICT for Sustainability (ICT4S 2020). Association for Computing Machinery, New York, NY, USA, 36–45. doi:10.1145/3401335.3401716 [15] COVID-NMA Initiative. 2020. COVID-NMA: A Living Systematic Review of COVID19 Evidence. https://covid-nma.com/ Accessed: 2026-03-31. [16] Clément Desroches, Martin Chauvin, Louis Ladan, Caroline Vateau, Simon Gosset, and Philippe Cordier. 2025. Exploring the Sustainable Scaling of AI Dilemma: A Projective Study of Corporations’ AI Environmental Impacts. arXiv preprint arXiv:2501.14334 (2025). doi:10.48550/arXiv.2501.14334 [17] Cooper Elsworth, Keguo Huang, David Patterson, Ian Schneider, Robert Sedivy, Savannah Goodman, Ben Townsend, Parthasarathy Ranganathan, Jeff Dean, Amin Vahdat, Ben Gomes, and James Manyika. 2025. Measuring the environmental impact of delivering AI at Google Scale. arXiv preprint arXiv:2508.15734 (2025). doi:10.48550/arXiv.2508.15734 [18] European Commission. 2025. In focus: Data centres – an energy-hungry challenge. https://energy.ec.europa.eu/news/focus-data-centres-energy-hungrychallenge-2025-11-17_en. Reports that data centres consume about 1.5% of global electricity, approximately 415 TWh per year. [19] Sophia Falk, David Ekchajzer, Thibault Pirson, Etienne Lees-Perasso, Augustin Wattiez, Lisa Biber-Freudenberger, Sasha Luccioni, and Aimee van Wynsberghe. 2025. More than Carbon: Cradle-to-Grave Environmental Impacts of GenAI Training on the NVIDIA A100 GPU. arXiv preprint arXiv:2509.00093 (2025). doi:10.48550/arXiv.2509.00093 [20] Richard P. Feynman. 1985. Surely You’re Joking, Mr. Feynman! W. W. Norton & Company, New York. [21] Matthias Finkbeiner, Lindsey Roche, and Peter Holzapfel. 2025. From analysisLCA to message-LCA: a lost cause? The International Journal of Life Cycle Assessment 30 (April 2025), 803–810. doi:10.1007/s11367-025-02460-9 [22] GDR DEFIE. 2024. GDR DEFIE: Défis de l’Électronique pour l’Écoconception. https://defie-cnrs.fr/. CNRS research network on sustainable electronics and ecodesign; accessed: 2026-02-20. [23] GDR DEFIE. 2026. ElecImpact: An Open Life Cycle Assessment Database for Electronics. https://framagit.org/defie/elecimpact. Open LCA database for electronic systems developed by the GDR DEFIE working group; accessed: 202602-20. [24] GenAI Impact. 2024. EcoLogits: Estimating the Environmental Footprint of Generative AI Models. https://ecologits.ai/latest/ Open-source software based on life-cycle assessment principles for estimating the environmental impacts of generative AI inference. [25] Global Enabling Sustainability Initiative (GeSI). 2024. Digital with Purpose: Delivering a SMARTer2030. Technical Report. Global Enabling Sustainability Initiative. https://www.gesi.org/wp-content/uploads/2024/10/GeSI-Digital-WithPurpose-Full-report.pdf Full report. [26] Louis Golard, Youssef Agram, François Rottenberg, François Quitin, David Bol, and Jérôme Louveaux. 2024. A Parametric Power Model of Multi-Band Sub-6 GHz Cellular Base Stations Using On-Site Measurements. In IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC). doi:10. 1109/PIMRC59610.2024.10817448 [27] Louis Golard, Robin Dethienne, Jérôme Louveaux, and David Bol. 2024. A Parametric Life-Cycle Model for Assessing Environmental Impacts of 4G and 5G Cellular Base Stations. The International Journal of Life Cycle Assessment (2024). [28] Carolina Graña Possamai, Guillaume Cabanac, Emilie Perrodeau, Laure Ghosn, Philippe Ravaud, and Isabelle Boutron. 2025. Inclusion of Retracted Studies in Systematic Reviews and Meta-Analyses of Interventions: A Systematic Review and Meta-Analysis. JAMA Internal Medicine (31 March 2025). doi:10.1001/ jamainternmed.2025.0256 Published online March 31, 2025. [29] Loïc Guibert, David Bekri, Louise Aubet, Steve Berberat, and Sébastien Rumley. 2026. Why Measuring AI Environmental Impact of Organisations is Non-Trivial?. In Proceedings of Machine Learning Research - Proceedings of the Fourth Swiss AI Days, Vol. 309. PMLR. [30] Reinout Heijungs and Mark A. J. Huijbregts. 2004. A review of approaches to treat uncertainty in LCA. International Journal of Life Cycle Assessment 9, 2 (2004), 127–137. doi:10.1007/BF02978428 [31] Thomas Herndon, Michael Ash, and Robert Pollin. 2014. Does High Public Debt Consistently Stifle Economic Growth? A Critique of Reinhart and Rogoff. Cambridge Journal of Economics 38, 2 (2014), 257–279. doi:10.1093/cje/bet075 [32] Gabrielle Huston. 2023. Your emails aren’t killing the planet: debunking the myth of emails’ carbon footprint. https://mobilesyrup.com/2023/12/01/emailscarbon-footprint-not-killing-planet-earth-debunking-myth/ Summarizes expert critiques and clarifications by Mike Berners-Lee; accessed 2026-02-20. [33] International Energy Agency. 2021. Japan 2021: Energy Policy Review. doi:10. 1787/72bb987a-en
CF Companion ’26 , May 19–21, 2026, Catania, Italy [34] International Energy Agency. 2025. Energy and AI. https://www.iea.org/reports/ energy-and-ai [35] ISO. 2006. ISO 14040:2006 - Life cycle assessment - Principles and framework. Norm. International Organization for Standardization. [36] ISO. 2006. ISO 14044:2006 - Life cycle assessment - Requirements and guidelines. Norm. International Organization for Standardization. [37] Romain Jacob. 2025. What Does That Really Tell Us? Interpreting Numbers in Sustainability Reports. ETH Zurich Research Collection. https://nsg.ethz.ch/publications/2025-09-26-what-does-that-really-tell-usinterpreting-numbers-in-sustainability-reports-20-500-11850-784322/ Creative Commons Attribution 4.0 International (CC BY 4.0). [38] George Kamiya and Vlad C. Coroamă. 2025. Data Centre Energy Use: Critical Review of Models and Results. https://www.iea-4e.org/wp-content/uploads/2025/ 05/Data-Centre-Energy-Use-Critical-Review-of-Models-and-Results.pdf Prepared for the IEA 4E TCP. [39] Freja Konradsen, Kristine Sofie Holse Hansen, Agneta Ghose, and Massimo Pizzol. 2024. Same product, different score: how methodological differences affect EPD results. The International Journal of Life Cycle Assessment 29, 2 (2024), 291–307. doi:10.1007/s11367-023-02246-x Open access. [40] Nikhil Krishnan, Sarah Boyd, Ajay Somani, Sébastien Raoux, Daniel Clark, and David Dornfeld. 2008. A Hybrid Life Cycle Inventory of Nano-Scale Semiconductor Manufacturing. Open-access PDF available via eScholarship. Environmental Science & Technology 42, 8 (2008), 3069–3075. doi:10.1021/es071174k Supplementary information provided in PDF; gate-to-gate and hybrid LCI details.. [41] Pierre Le Gargasson, Maxime Pelcat, Nicolas Beuve, Muriel Pressigout, Luce Morin, and Daniel Menard. 2025. PCBnCO: A Carbon Intensity Model of FR-4 Printed Circuit Boards Based on Company Data. In 2025 IEEE Conference on Technologies for Sustainability (SusTech). IEEE, Los Angeles, CA, USA. doi:10. 1109/SusTech63138.2025.11025732 [42] Philippe Loubet, Adrien Vincent, Annabelle Collin, Corinne Dejous, Anthony Ghiotto, and Christophe Jego. 2023. Life cycle assessment of ICT in higher education: a comparison between desktop and single-board computers. The International Journal of Life Cycle Assessment 28 (2023), 255–273. doi:10.1007/ s11367-022-02131-z [43] Adrian Lubecki, Jakub Szczurowski, and Katarzyna Zarębska. 2025. The importance of uncertainty sources in LCA for the reliability of environmental comparisons: A case study on public bus fleet electrification. Applied Energy 377 (2025), 124593. doi:10.1016/j.apenergy.2024.124593 [44] Donald MacKenzie and Taylor Spears. 2014. “The Formula That Killed Wall Street”: The Gaussian Copula and Modelling Practices in Investment Banking. Social Studies of Science 44, 3 (2014), 393–417. doi:10.1177/0306312713517157 [45] David E Mann and Steven M Christey. 1999. Towards a common enumeration of vulnerabilities. In 2nd Workshop on Research with Security Vulnerability Databases, Purdue University, West Lafayette, Indiana, Vol. 9. [46] Anders Nordelöf. 2019. A scalable life cycle inventory of an automotive power electronic inverter unit—Part II: manufacturing processes. The International Journal of Life Cycle Assessment 24, 4 (2019), 694–711. doi:10.1007/s11367-0181491-3 Manufacturing datasets; links to ecoinvent; complements the Part I parametric model.. [47] Anders Nordelöf, Mikael Alatalo, and Maria Ljunggren Söderman. 2019. A scalable life cycle inventory of an automotive power electronic inverter unit—Part I: design and composition. The International Journal of Life Cycle Assessment 24 (2019), 78–92. doi:10.1007/s11367-018-1503-3 Parametric LCI model; Excel dataset available via Swedish Life Cycle Center (SPINE@CPM).. [48] Anders Nordelöf, Emma Grunditz, Anne-Marie Tillman, Torbjörn Thiringer, and Mikael Alatalo. 2018. A scalable life cycle inventory of an electrical automotive traction machine—Part I: design and composition. The International Journal of Life Cycle Assessment 23, 1 (2018), 55–69. doi:10.1007/s11367-017-1308-9 Parametric LCI model with downloadable Excel (SPINE@CPM / Swedish Life Cycle Center).. [49] Anders Nordelöf, Maarten Messagie, Anne-Marie Tillman, Maria Ljunggren Söderman, and Joeri Van Mierlo. 2014. Environmental impacts of hybrid, plug-in hybrid, and battery electric vehicles—what can we learn from life cycle assessment? The International Journal of Life Cycle Assessment 19 (2014), 1866–1890. doi:10.1007/s11367-014-0788-0 Open access; published online 21 Aug 2014; Accessed: 2026-03-31. [50] Varun Gowda Palahalli Ramesh and Christina Lee. 2025. Interpretation of LCA Results and EPD Comparability. In Life Cycle Analysis Based on Nanoparticles Applied to the Construction Industry. Springer Nature, 147–161. doi:10.1007/9783-031-79115-4_9 Discusses that public EPDs present selected aggregated results while full LCI is in a non-public background report. [51] Rachel Potvin and Josh Levenberg. 2016. Why Google stores billions of lines of code in a single repository. Commun. ACM 59, 7 (2016), 78–87. [52] Kate Raworth. 2017. Doughnut Economics: Seven Ways to Think Like a 21st-Century Economist. Chelsea Green Publishing, White River Junction, VT. UK edition: Random House Business, London. [53] Carmen M. Reinhart and Kenneth S. Rogoff. 2010. Growth in a Time of Debt. American Economic Review 100, 2 (2010), 573–578. doi:10.1257/aer.100.2.573
CF Companion ’26 , May 19–21, 2026, Catania, Italy
[54] David Sánchez, Marina Proske, and Sarah-Jane Baur. 2022. Life Cycle Assessment of the Fairphone 4. https://www.fairphone.com/wp-content/uploads/2022/07/ Fairphone-4-Life-Cycle-Assessment-22.pdf. Accessed: 2025-01-30. [55] Tilman Santarius, Jan C. T. Bieser, Vivian Frick, Mattias Höjer, Maike Gossen, Lorenz M. Hilty, Eva Kern, Johanna Pohl, Friederike Rohde, and Steffen Lange. 2023. Digital sufficiency: conceptual considerations for ICTs on a finite planet. Annals of Telecommunications 78, 5–6 (2023), 277–295. doi:10.1007/s12243-02200914-x [56] Ian Schneider, Hui Xu, Stephan Benecke, David Patterson, Keguo Huang, Parthasarathy Ranganathan, and Cooper Elsworth. 2025. Life-Cycle Emissions of AI Hardware: A Cradle-To-Grave Approach and Generational Trends. arXiv preprint arXiv:2502.01671 (2025). doi:10.48550/arXiv.2502.01671 [57] Science Based Targets initiative. [n. d.]. Science Based Targets initiative. https: //sciencebasedtargets.org. Global initiative enabling companies to set sciencebased emissions reduction targets in line with climate science. [58] Nir Shlezinger and Yonina C. Eldar. 2023. Model-Based Deep Learning. arXiv (2023). arXiv:2306.04469 [eess.SP] doi:10.48550/arXiv.2306.04469 [59] Matthew Sparkes. 2026. Computer finds flaw in major physics paper for first time. New Scientist (26 March 2026). https://www.newscientist.com/article/2520546computer-finds-flaw-in-major-physics-paper-for-first-time/ [60] The Carbon Literacy Project. 2022. The Carbon Cost of an Email: Update! https://carbonliteracy.com/the-carbon-cost-of-an-email/. Accessed 2026-02-20. [61] Joseph Tooby-Smith. 2026. Formalizing the stability of the two Higgs doublet model potential into Lean: identifying an error in the literature. arXiv preprint arXiv:2603.08139 (2026). doi:10.48550/arXiv.2603.08139 [62] UNEP. 2019. Environmental Rule of Law: First Global Report. Technical Report. https://wedocs.unep.org/handle/20.500.11822/27279 [63] LCIE Bureau Véritas. 2025. Négaoctet. https://codde.fr/nos-marques/negaoctet. Accessed: 20/02/26.
Corlay et al.
[64] Qi Wang, Nan Huang, Zhuo Chen, Xiaowen Chen, Hanying Cai, and Yunpeng Wu. 2023. Environmental data and facts in the semiconductor manufacturing industry: An unexpected high water and energy consumption situation. Water Cycle 4 (2023), 47–54. [65] Bo P. Weidema, Christian Bauer, Roland Hischier, Christopher Mutel, Thomas Nemecek, Jürg Reinhard, Carl O. Vadenbo, and Gregor Wernet. 2013. Overview and Methodology: Data Quality Guideline for the ecoinvent Database Version 3. Technical Report 1. Swiss Centre for Life Cycle Inventories. [66] Bo P. Weidema and Mette S. Wesnæs. 1996. Data quality management for life cycle inventories—an example of using data quality indicators. Journal of Cleaner Production 4, 3–4 (1996), 167–174. doi:10.1016/S0959-6526(96)00043-1 [67] Olivier Weppe, Jérôme Chossat, Thibaut Marty, Jean-Christophe Prévotet, and Maxime Pelcat. 2024. Streamlined Models of CMOS Image Sensors Carbon Impacts. In 2024 27th Euromicro Conference on Digital System Design (DSD). IEEE, 250–257. [68] Olivier Weppe, Thibaut Marty, Sébastien Toussaint, Nicolas Brusselmans, JeanChristophe Prévotet, Jean-Pierre Raskin, and Maxime Pelcat. 2025. Embodied Carbon Footprint of 3D NAND Memories. In Proceedings of the 22nd ACM International Conference on Computing Frontiers: Workshops and Special Sessions (CF ’25 Companion). Association for Computing Machinery, New York, NY, USA, 108–116. doi:10.1145/3706594.3727962 [69] Cheng Zhang, Junfeng Jing, Liu Yun, Yu Zheng, and Haihong Huang. 2022. A cradle-to-grave life cycle assessment of high-voltage aluminum electrolytic capacitors in China. Journal of Cleaner Production 370 (2022), 133244. doi:10. 1016/j.jclepro.2022.133244 [70] Cheng Zhang, Yu Zheng, Haihong Huang, Zhifeng Liu, and Junfeng Jing. 2023. Environmental impact assessment of aluminum electrolytic capacitors in a product family from the manufacturer’s perspective. The International Journal of Life Cycle Assessment 28 (2023), 80–94. doi:10.1007/s11367-022-02117-x