Work

Current Progress on Semi-Supervised Learning for Data-Limited Chemistry and Material Discovery

Year: 2026
Type: preprint
Abstract: Machine Learning (ML) has become one of the most important tools in science and industry, yet chemistry and materials science have not benefited from it to the same extent as many other fields, because ML models depend on clean, large labeled datasets that are inherently difficult to obtain in these domains. Semi-supervised learning (SSL) offers a promising route, by leveraging small labeled sets alongside abundant unlabeled data, but the central question is not which SSL method to use; it is when unlabeled data can be trusted. Unlabeled data can just as easily hurt a model as help it when pseudo-labels are wrong or the data shifts, and this risk is not uniform. Reviewing work across molecular property prediction, synthesis, text mining, diffraction, microscopy, and materials screening, we find a consistent divide: SSL is reliably effective in classification and segmentation tasks where errors are bounded but remains structurally fragile in the continuous regression problems (yields, lifetimes, molecular activity) that dominate this field, where a single bad pseudo-label can propagate unchecked. We use this divide as a diagnostic lens throughout the review, and close by outlining what reliable SSL requires (uncertainty pseudo-labeling, principled evaluation, and a minimum reporting standard) and where the open challenges lie.
Source: ChemRxiv
Language: English
Open Access status: green
OpenAlex ID: W7204206144
Cites:
Cited by:
Related to: 10