luMOD: A Lightweight Universal Midrange Atomic Orientational Descriptor

Extracting structural information in atomistic simulations is crucial for applying machine learning to understand structure-property relationships and accelerate materials exploration. Existing descriptors can encode beyond-local environments and address multispecies complexities, but often suffer from high dimensionality and limited flexibility. In this work, we develop luMOD, a lightweight universal Midrange atomic Orientational Descriptor composed of only 24 features. Despite its compact size, it integrates direct and convoluted bond orientational order parameters with neighbor species distributions, enabling efficient description of heterogeneous atomic environments in several nearest-neighbor layers. luMOD demonstrates strong versatility across both single- and multispecies systems, including complex crystal structure interfaces such as perovskites and olivines. In addition to the well-maintained accuracy and applicability of luMOD, its fixed short length generally helps reduce the training time of downstream machine learning models, particularly in multispecies systems. We believe this development offers a promising step toward more efficient and scalable structural descriptors, helping to accelerate data-driven materials discovery and design.