Enabling discovery of materials through enhanced generalisability of deep learning models

The road towards the discovery of new, useful materials is full of imperfections. Machine learning, which is currently the power horse of material discovery when it works in concert with density functional theory, has been able to accelerate the discovery of new materials for various applications by learning the properties of known, stable materials to infer the properties of unknown, deformed materials. Physics-informed machine learning (PIML) is particularly believed to bridge the gap between known materials and the virtually infinite space of crystal structures with imperfections such as defects, grain boundaries, composition disorders and others. State of the art PIML models struggle to bridge this gap, however. In this work we show that a critical correction of the physics underpinning a PIML, our direct integration of external potential (DIEP) method, which mimics the integration of the external potential term in typical density functional theory calculations, improves the generalisability of the model to a range of imperfect structures, including diamond defects. DIEP has the potential of enhancing the accuracy of potential energy surface prediction, which can accelarate structure discovery.