Sustainability Objectives in Polymer Machine Learning: A Critical Review of Integration, Reliability, and Evidence Gaps

Source abstract

Machine learning (ML) is increasingly applied to polymers described as sustainable, yet sustainability may enter a workflow as material identity rather than as an explicit computational objective. This critical review combines narrative interpretation with a structured evidence analysis of a fixed analytical sample of 49 studies published from 2022 to 2026. It is a non-systematic review and does not claim exhaustive coverage; all numerical frequencies refer only to the analysed sample. Across four objective families—functional/material performance, end-of-life, hazard/ecotoxicity, and life-cycle impact—performance was modelled in-loop in 43/49 studies, compared with 14/49 for end-of-life and 3/49 for life-cycle impact; none of the 49 sampled studies modelled hazard/ecotoxicity in-loop. Ten studies modelled at least two objective families, and 8/10 used formal cross-family joint optimisation. Formal multi-objective optimisation appeared in 15 studies, but none of these 15 reported qualifying out-of-distribution (OOD) or grouped testing, explicit extrapolation testing, external-dataset validation, or applicability-domain analysis. Within the analytical sample, 5 of 15 possible non-empty objective combinations were occupied. These sample-level patterns suggest that optimisation sophistication has advanced faster than multidimensional sustainability integration and explicit validity assessment. Future polymer-ML workflows should state which sustainability endpoints are computationally represented, integrate validated non-performance objectives where data permit, and accompany candidate-space expansion with explicit evidence on uncertainty, generalisation, or applicability domains.