Fast prediction of microstructural responses based on realistic material topology is vital for linking process,structure,and properties.This work presents a digital framework for metallic materials using microscale fe...Fast prediction of microstructural responses based on realistic material topology is vital for linking process,structure,and properties.This work presents a digital framework for metallic materials using microscale features.We explore deep learning for two primary goals:(1)segmenting experimental images to extract microstructural topology,translated into spatial property distributions;and(2)learning mappings from digital microstructures to mechanical fields using physics-informed operator learning.Loss functions are formulated using discretized weak or strong forms,and boundary conditions-Dirichlet and periodic-are embedded in the network.Input space is reduced to focus on key features of 2D and 3D materials,and generalization to varying loads and input topologies are demonstrated.Compared to FEM and FFT solvers,our models yield errors under 1–5%for averaged quantities and are over 1000×faster during 3D inference.展开更多
基金the funding support provided to develop the present work in the project Cluster of Excellence“Internet of Production”(project:390621612).
文摘Fast prediction of microstructural responses based on realistic material topology is vital for linking process,structure,and properties.This work presents a digital framework for metallic materials using microscale features.We explore deep learning for two primary goals:(1)segmenting experimental images to extract microstructural topology,translated into spatial property distributions;and(2)learning mappings from digital microstructures to mechanical fields using physics-informed operator learning.Loss functions are formulated using discretized weak or strong forms,and boundary conditions-Dirichlet and periodic-are embedded in the network.Input space is reduced to focus on key features of 2D and 3D materials,and generalization to varying loads and input topologies are demonstrated.Compared to FEM and FFT solvers,our models yield errors under 1–5%for averaged quantities and are over 1000×faster during 3D inference.
