Electronegativity-Inspired Multidimensional and Complex Representations for Modeling Chemical Environments in Materials

Abstract

Materials discovery is increasingly facilitated by data-driven approaches, which necessitate accurate descriptors of chemical bonding. Conventional scalar electronegativity, one of the essential descriptors, cannot capture the complex electronic environments in materials. Here, we introduce multidimensional electronegativity-inspired representations for most elements and extend them to complex-valued forms to more effectively describe the local chemical environments. Starting from Pauling’s definition based on bond dissociation energy, we optimize these vectors using the energy stabilization associated with material formation. Inspired by Mulliken’s definition, we further incorporate ionization energy and electron affinity into the real and imaginary components of the vectors. Using crystal graph convolutional neural networks (CGCNNs) and a complex-valued variant, we demonstrate that these descriptors more accurately capture atomic interactions in local environments, leading to improved convergence behavior. Our results highlight that multidimensional and complex-valued representations are effective descriptors for modeling chemical environments in materials and provide a promising approach to materials discovery.