Abstract Scope |
During the past few years, there has been an explosion of new ideas regarding features / descriptors, machine learning algorithms, and neural network architectures for predicting composition-property or structure-property relationships. Recently, we introduced a standard benchmark (Matbench) for measuring the performance of these algorithms through testing on a common data set, with initial conclusions showing "conventional" feature-based machine learning working well for smaller data sets and graph-based neural network methods working better for larger data sets. In this talk, I will first re-introduce the Matbench test set, which is a set of 13 supervised machine learning problems derived from 10 experimental and ab initio datasets and which range in size from 312 to 132,752 samples. I will next summarize our findings in using this data to benchmark state-of-the-art machine learning methods for property prediction including CGCNN, MEGNet, Automatminer, Roost, CRABNet, and MODNet, and others. |