Deep energy method based on physics-guided spatial attention mechanism
Аннотация
When applying the deep energy method (DEM) to highly heterogeneous multi-material systems, the inherent ‘spectral bias’ of neural networks leads to a preference for low-frequency signals, causing a spectral aliasing problem that results in excessive smoothing of the displacement field and making it difficult to resolve abrupt changes in interfacial stress. To address this, this paper proposes a physics-driven attention-enhanced DEM algorithm. This method prevents spectral aliasing through decoupled mapping of geometric and physical features and introduces stress-driven attention gating to enhance feature representation in regions with constitutive discontinuities and high-stress concentrations. A validation model of an aircraft tire, featuring 17 material layers with extreme stiffness contrasts and undergoing large hyperelastic deformation, is established using experimentally fitted data. Experiments demonstrate that the algorithm accurately reconstructs the mechanical transfer mechanisms at complex interfaces, achieving high-fidelity meshless solutions while matching the accuracy of commercial finite element software. Owing to its mesh-free nature, the DEM proposed in this paper possesses a natural advantage over mesh-based methods in circumventing convergence issues caused by severe mesh distortion. It provides a reliable new paradigm for high-fidelity mechanical simulation of complex composite structures and advances the implementation of deep learning-based partial differential equation solvers in industrial-scale applications.