A General Machine Learning Force field for Structure Prediction of 2D Organic-Inorganic Perovskites

Two dimensional hybrid organic-inorganic perovskites (2D HOIPs) represent an extremely promising class of electronically active materials for both light absorption and emission. However, the design space of 2D HOIPs is large, because of the diversity of organics molecules that can be combined with various inorganic lattices to tune the material electronic and mechanical properties. This large design space necessitates the development of new tools for the high throughout analysis of candidates in silico. In this work, we present an accurate, efficient and widely applicable machine learning interatomic potential (MLIP) for predicting the structures of new 2D HOIPs. Using the MACE framework, we train an MLIP to the currently available experimental databse of 2D HOIP structures, with PbX₄ (X: Cl, Br, I) and 98 organic cations with +1 and +2 charge states.  This generalized potential is accurate enough that is capable of optimizing the geometry of totally unseen HOIPs to within the accuracy of the density functional theory (DFT) optimizations. The main advantage over DFT optimization is the computational efficency of the MLIP which can get to the global miminina with a fraction of the cost of DFT.

To further demonstrate the accuracy of our potential, we present a random structure searching algorithm where given an arbitrary 2D HOIPS, made of any PbX₄ and organic units, one can search the energy landscape with the MLIP and find the global minima structure that matches the experimental predictions. Our study shows that we can predict new HOIPs that have not been synthesized before and DFT calculations validates the stability of our predictions.