Non-Random Parameterized Networks for Cross-Scale Modeling of Compositional Interplay

Accurate prediction of molecular properties in complex chemical systems is crucial for accelerating materials discovery and chemical innovation. However, current computational methods often struggle to capture the intricate compositional interplay across multiple scales, from intramolecular bonds to intermolecular forces. Here, we introduce MesoNet, a novel non-random parameterized network specifically designed for cross-scale modeling. MesoNet's innovation lies in the construction of Non-Random Parameters (NRPs) – dynamically enriched atomic descriptors generated via Neural Circuit Policies (NCPs). NRPs uniquely capture both intrinsic atomic properties and their dynamic compositional context. Subsequently, these NRPs are processed through a hierarchical cross-scale message passing mechanism that explicitly integrates intra- and intermolecular interactions, essential for representing compositional interplay. Comprehensive evaluations across diverse datasets demonstrate that MesoNet achieves significantly superior predictive accuracy and enhanced chemical interpretability for molecular properties in both single and multi-component systems compared to existing methods. This work establishes a powerful and interpretable approach for cross-scale modeling of compositional complexity, aiming at advanced chemical simulations and design.