Fig. 1: Leave-one-element-out performance on Materials Project.
From: Probing out-of-distribution generalization in machine learning for materials
a Leave-one-element-out performance across the periodic table for the XGB and ALIGNN models. Despite the presence of negative R2 for H and F, the colorbar range is bounded between 0 and 1 for better visibility. The OOD tests are performed only for elements with more than 200 data. Results for other models and datasets are provided in Supplementary Figs. 2–12. The periodic tables are plotted using pymatviz59. b Violin plots of SHAP contributions from compositional and structural features for selected left-out elements. c Number of leave-one-element-out tasks with R2 higher than a given threshold as a function of threshold. LLM-Prop is not included due to the large number of tasks and high training cost. d Comparison of performance between ML models for selected leave-one-element-out tasks.
