From Neural Surrogates to Language Agents: A Survey of Artificial Intelligence for Life Cycle Assessment

Two decades of AI for life cycle assessment, 428 works verified against their sources: most studies address tasks whose outputs a computation or a reference can check, and few propagate their prediction error into the assessment.

Abstract

Life cycle assessment (LCA) quantifies the potential environmental impacts of products, but assessments are slow, data-hungry and dependent on expert judgement. For two decades, artificial intelligence has been proposed to relieve these bottlenecks: first neural and knowledge-based approximations, then supervised learning on inventory databases and semantic matching, and now pretrained language models and agents. This survey reviews 428 works; every published work was verified against Crossref, arXiv or the issuing body. They span LCA methodology and data, the AI methods used in LCA, 127 core AI-for-LCA studies, AI for climate disclosure, and the environmental footprint of AI. Every work was placed by technique and LCA phase twice: by a first rater (an AI assistant working for the author) and by a typed classifier. Disagreements on the questions that bear on the analysis were settled by a blinded language-model adjudicator, and the main results are reported under all three label sets. We trace four generations and five lineages, from learning-system approximations and the FineChem models to emission-factor recommenders and multimodal LCA agents. We define the verifiability level of an LCA task by the evidence that can check an automated output: a computation, reference labels, constraints only, or expert judgement. Depending on the label set, the primary AI-for-LCA studies number 84–88, and 74–83% of them produce outputs that a computation or a reference can check. In inventory, impact assessment and surrogate prediction, the phases with most AI work, 67–100% of studies address checkable tasks. The phases dominated by methodological choices and review attract few studies. Studies checkable only by judgement or against constraints are few in every period, and their rise since 2023 is not statistically significant. Of the 69 primary studies with abstracts, only 18 report a comparison with held-out references, and only 1 propagates its prediction uncertainty into the assessment result. A targeted search without citation limits finds the same pattern. We call it a verifiability gradient and weigh it against five alternative explanations. We derive six open problems with testable hypotheses: building references, propagating the error of learned values, evaluation without leakage, semantic interoperability, controlled agentic workflows, and the footprint of AI-assisted assessment.

This is a working paper. The full text is in preparation for publication; the abstract above reports its current results.

@techreport{guo2026ai,
  title       = {From Neural Surrogates to Language Agents: A Survey of Artificial Intelligence for Life Cycle Assessment},
  author      = {Guo, Yanming},
  institution = {AIDC Research},
  year        = {2026},
  type        = {Survey},
  url         = {https://www.ai-dc.ai/research/ai-lca-survey}
}