A reproducibility study of atomistic line graph neural networks for materials property prediction†

Abstract

Use of machine learning has been increasingly popular in materials science as data-driven materials discovery is becoming the new paradigm. Reproducibility of findings is paramount for promoting transparency and accountability in research and building trust in the scientific community. Here we conduct a reproducibility analysis of the work by K. Choudhary and B. Brian [npj Comput. Mater., 7, 2021, 185], in which a new graph neural network architecture was developed with improved performance on multiple atomistic prediction tasks. We examine the reproducibility for the model performance on 29 regression tasks and for an ablation analysis of the graph neural network layers. We find that the reproduced results generally exhibit a good quantitative agreement with the initial study, despite minor disparities in model performance and training efficiency that may be resulting from factors such as hardware difference and stochasticity involved in model training and data splits. The ease of conducting these reproducibility experiments confirms the great benefits of open data and code practices to which the initial work adhered. We also discuss some further enhancements in reproducible practices such as code and data archiving and providing data identifiers used in dataset splits.