The Glass Predictor

Predicting whether a liquid will form a glass when cooled — rather than crystallizing — is an old problem. The theoretical landscape is cluttered with competing models and ad hoc rules. Carvalho, Loponi, and Cassar train random forest classifiers on over 50,000 compositions and achieve strong prediction (ROC-AUC ~0.89).

The surprise is what matters and what doesn’t. The bandgap energy of the constituent chemical elements correlates positively with glass formation — wider gaps favor the amorphous state. But established glass stability parameters and Jezica metrics, the traditional tools for assessing glass-forming ability, provide no additional predictive power when the elemental features are already included. The expert knowledge is redundant.

The through-claim: the traditional metrics were proxies for information already contained in simpler elemental properties. When the model can access the underlying features directly, the derived indicators add nothing. The metrics weren’t wrong — they were summaries of something the raw data already said.

This inverts the usual relationship between theory and prediction. Normally, theoretical parameters explain why predictions work. Here, the predictions work without the theoretical parameters, suggesting the parameters were always downstream of more fundamental composition features. The glass doesn’t know about your stability criteria. It knows about its atoms.

Practical consequence for inverse design: screening new glass compositions requires only elemental properties, not the expensive characterization needed to compute stability parameters. The shortcut is also the correct route.