Neural Operators Learn Material Behavior with Causal Attention.

Traditional methods for modeling material behavior, especially for materials that deform irreversibly, often rely on internal variables that are difficult to measure. This new approach bypasses that limitation by treating a deforming material as a functional mapping, directly learning its stress response from its entire strain history. The framework utilizes neural operators, which are trained on full loading paths to predict complete stress trajectories efficiently. A key innovation is the integration of a causally masked attention mechanism, ensuring that the model only considers past material states while maintaining computational parallelism. This allows for robust and accurate predictions of complex phenomena like plasticity and ductile damage, even without prior knowledge of internal material states.