MatSciBench: A flexible benchmarking platform for machine learning in condensed matter science

With the increasing application of deep learning and big data technologies in materials science, researchers face challenges such as fragmented private datasets, inconsistent evaluation standards, and difficulties in reproducing results. To address these issues, this paper presents MatSciBench: an open and extensible machine learning benchmark platform for materials science. The platform integrates multi-source data from materials databases such as the Materials Project and Materiae, as well as open experimental datasets, to provide standardized datasets. It also establishes a multi-dimensional evaluation system (assessing accuracy, validity, RMSE, etc.) covering multiple task types: property prediction, structure generation, object detection, and scientific reasoning. Through online model submission and leaderboard functionalities, MatSciBench enables an integrated workflow for data acquisition, model evaluation, and result sharing. Case studies on topological materials classification, crystal structure generation, and XRD structure resolution demonstrate the platform's effectiveness in model performance comparison, data distribution visualization, and result reproducibility. The open-access and iterative nature of MatSciBench provides a fair, transparent, and efficient evaluation environment for machine learning research in condensed matter physics and materials science.