Potenciación de la Predicción de Propiedades Cristalinas mediante la Actualización Dinámica de Características y...

Graph neural networks (GNNs) inherently excel at representing atomistic material, therefore enabling predictions of properties with high accuracy with embedded descriptors. Approaching to accurate prediction of a variety of physical properties of diverse materials, particularly those with multicomponents, remains challenging due to the complexity of many-body atomic interactions and the reliance on handcrafted symmetry descriptors. Here, we introduce the Wavelet Atomic Neighborhood Network (WANN), a novel framework that implicitly captures many-body interactions through an iterative subembedding module for updating atomic features, thereby eliminating the integration of the predefined feature engineering. A wavelet-based regression component is built up through multiscale feature analysis. WANN demonstrates superior accuracy, achieving small mean absolute error (MAE) in predicting a variety of physical and chemical quantities, including thermodynamical, dielectric, piezoelectric, magnetic, and thermoelectric properties. For instance, it could significantly reduce MAE maximum by 61.64% in comparison to Matformer for shear moduli and by 90.84% in comparison to ALIGNN for the exfoliation energy. This scalable and preprocessing-light deep-learning framework provides an efficient toolkit for facilitating high-throughput screening of emergent materials, like high-entropy alloys and ceramics.