News
Efficient Higher-Order Derivatives for Machine Learning Potentials
Nils' work on getting higher-order derivatives of machine learning potentials (namely Hessians) via automatic differentiation was just published in JCTC! The new method is both more accurate and more efficient than the established finite-difference approach. The paper also shows that foundation models like MACE-MP-0 can be used for predicting vibrational properties (derived from our Hessians) like the heat capacities of porous materials.