Assessing Quality of Machine-Learning Models and Underlying Data

The power of artificial intelligence has allowed us to reach a new level of scientific approaches with predictive power. On the one hand, machine learning is used to explore trends in material properties. On the other hand, one may aim at highly accurate modeling. How accurate finally predictions are for a given problem at hand, depends on various factors. With a range of very different examples, we assess what Big Data means in the context of typical machine-learning problems in materials science [1]. This concerns data volume, data quality and veracity, but also infrastructure issues. We also show how machine learning, in turn, can be used for error quantification and data augmentation [2,3].<br/><br/>[1] D. Speckhard, T. Bechtel, L. M. Ghiringhelli, M. Kuban, S. Rigamonti, and C. Draxl, Faraday Discussion, https://doi.org/10.1039/D4FD00102H<br/>[2] D. Speckhard et. al., https://arxiv.org/abs/2303.14760<br/>[3] M. Kuban, S. Rigamonti, and C. Draxl, https://arxiv.org/abs/2403.10470