Hybrid machine learning/physics-based approach for predicting oxide glass-forming ability

Summary¶

Glass-forming ability (GFA) is commonly operationalized as the critical cooling rate required to bypass crystallization on quench from the melt. Industrial oxide glass development still relies heavily on empirical composition maps and heuristics, while physics-based models struggle to deliver quantitative, composition-specific GFA predictions across large formulation spaces. This Acta Materialia article proposes a hybrid approach: a deliberately simplified toy potential energy landscape captures essential liquid thermodynamics and kinetic accessibility features associated with vitrification, and machine learning maps those structured descriptors to critical cooling rates or closely related GFA surrogates for oxide compositions. The intent is to combine interpretable physics with data-driven flexibility rather than treating ML as a black-box fit to composition vectors alone.

Methods¶

Reviews / non-atomistic methods (per AGENTS block 4). This is not a LAMMPS production paper. The authors couple a toy landscape model of liquid thermodynamics and kinetics (to access vitrification and critical quench rate / GFA behavior) with machine learning to predict critical quench rates from composition (abstract: “toy landscape model … combined with machine learning … calculate the critical quench rate”). The landscape is meant to access the physics controlling glass-forming ability by simulating liquid thermodynamics and kinetics in a reduced model, then feeding results into ML; training uses published GFA and related data (see main text and supplementary material for sources and train/test design). 1 — MD application: N/A — explicit atomistic MD is not the core engine; the toy landscape replaces bulk melt MD in this workflow. 2 — ReaxFF / classical FF training: N/A — not a ReaxFF or EAM parameterization study. 3 — Static QM: N/A as the primary result. 4 — Experiments: N/A for new lab work—experimental GFA labels come from the literature compilation used for training.

Findings¶

The paper positions the toy landscape + machine learning combination as a way to estimate critical quench rates / glass-forming ability (GFA) with an explicit liquid thermodynamics and kinetics story (abstract: “toy landscape model … combined with machine learning”), rather than a black-box map from composition alone. Outcomes in the main text are the trained models’ performance on published GFA / r_{crit} labels and the interpretation of which landscape-derived descriptors matter—N/A in this note to reproduce every benchmark table; see Acta Materialia and SI. Comparisons include discussion of prior empirical and physics-based GFA routes (Zachariasen-style rules, constraint theories, calorimetric parameters) in the introduction; hybrid scores should be read against those baselines in the paper. Sensitivity to training splits and feature sets is a methodological theme for ML; detailed sweeps are in the article. Limitations and outlook follow the authors’ discussion of industrial impact, physical insight, and what the glass community would need for better GFA prediction (abstract and conclusions).

Limitations¶

This is not a ReaxFF or atomistic MD paper; reactive pathways, phase separation, and heterogeneous nucleation on experimental timescales are outside the model class unless added as extensions.

Reproducibility notes¶

Hybrid GFA models should be evaluated with train/validation splits that respect chemically related compositions; otherwise ML can memorize families rather than learn transferable landscape features. When integrating with atomistic pipelines, treat the toy landscape outputs as priors for composition screening rather than replacements for nucleation measurements.

For industrial readers, the key reproducibility artifact is the toy landscape parameterization itself: any change to basin depths or barrier heights shifts predicted critical cooling rates, so sensitivity analysis should accompany single-point GFA predictions. Where available, compare predicted rankings against experimental critical cooling measurements or surrogate metrics from dilatometry rather than trusting cross-validated scores alone.

(No dedicated corpus hub page yet for toy-landscape GFA.)