luMOD: A lightweight universal  Mid-range atomic Orientational Descriptor

Extracting structural information in atomistic simulations is crucial for linking the behaviors of individual atoms with collective properties of materials. Various modern descriptors have been developed to encode beyond-local environments and address multi-species complexities, promoting developments of machine learning algorithms for understanding structure-property relationship and accelerating materials exploration. However, these constructions often come at the cost of high dimensionality and compromised flexibility. In this work, we aim to develop an efficient descriptor that is capable of handling heterogeneities for arbitrary material systems in atomistic simulations. A lightweight universal Mid-range atomic Orientational Descriptor, luMOD, concatenating twenty-four engineered features, was designed based on direct and convoluted bond orientational order parameters as well as distributions of neighbor species. The versatility of luMOD was demonstrated across various single- and multi-species systems, including successful descriptions of interfaces composed of complex crystal structures such as perovskites and olivines. While the accuracy and applicability of our proposed descriptor are well maintained, its fixed short length generally helps reduce the training time for machine learning models, which would be seen as more significant when a high number of species are involved in the system. We believe this development holds promise for facilitating effective materials structure representation and computational data-driven materials design.