Exploring elastic properties of molecular crystals with universal machine learning interatomic potentials

We benchmarked existing and newly trained universal machine learning interatomic potentials for modelling molecular crystals, particularly their elastic properties. We found that potentials trained on the SPICE dataset provide reasonable predictions of the elastic properties of molecular crystals that are as good as predictions made using density functional theory-based methods. Still, the uncertainty of predictions and difference to experimental values is relatively high (larger than 5 GPa for Young's modulus). We have performed a high-throughput study of the elastic properties of molecular crystals. We have found that some of the molecular crystals show negative linear compressibility and validated our results using density functional theory. • Universal machine learning interatomic potentials are tested for predictions of elastic properties of molecular crystals. • Accuracy of best machine learning interatomic potentials is similar to density functional theory. • The high-throughput search identifies molecular crystals with interesting properties such as negative linear compressibility.