Nudged-elastic band calculations of polymorph transitions and solid-state reactions in molecular crystals

Abstract

The modeling of solid-state transformations, such as polymorphic transitions and chemical reactions in molecular crystals, is vital for many applications, including drug design and the development of new synthesis methods. However, a description via nudged-elastic band (NEB) calculations faces several crucial challenges. First, the automatic initial pathway generation based on a linear interpolation often fails for periodic systems, leading to unrealistic geometries and atomic collisions. Second, the necessary system sizes are typically beyond the scope of density functional theory (DFT) calculations in terms of computational cost, but the associated accuracy is vitally needed. To address these issues, we introduce a hybrid interpolation method that combines linear interpolation for cell parameters with spherical linear interpolation for molecular structures or intramolecular fragments, ensuring smooth and realistic transitions. Moreover, we train and benchmark machine-learned force fields (MLFFs) based on the SO3krates equivariant neural network architecture to accelerate NEB calculations while retaining near-DFT accuracy. We apply our approach to two polymorph transitions and a solid-state Diels–Alder reaction and show that our new interpolation method reliably produces viable initial pathways. The MLFFs are trained on PBE+MBD reference data and reproduce DFT lattice energies with a mean absolute error of 0.4 kJ mol−1 along the minimum-energy paths. These results highlight the potential of combining advanced interpolation techniques with MLFFs to enable automated, accurate, and efficient exploration of solid-state transformations in molecular crystals.