Fig. 4
Generalizability of machine learning models. Generalization performances of the ML models using our interstice distribution features (filled violin plots) and those using symmetry functions (unfilled violin plots) on five generalization scenarios: (i) same composition (abbreviated as C) and same quenching rate (QR) but to new unseen configurations, (ii) same composition and different quenching rate; (iii) different composition and same quenching rate; (iv) different composition and different quenching rate; (v) different system. The performance is evaluated by the difference of the generalization score and the original, same-condition score. The height of the violin plot is proportional to number of tests that fall within a performance interval and the interior shows the data points.
