A Physics-Informed Operator Learning Benchmark for Thermal-Mechanical Surrogate Modeling in WA-DED

Operator-learning surrogates are emerging as a practical route to real-time simulation for wire-arc directed energy deposition (WA-DED). However, predictive accuracy is often degraded when jointly learning transient reversible quantities (e.g., elastic and thermal strain) and history-dependent responses (e.g., plastic strain and stress). To address the resulting feature interference, we establish a benchmark for evaluating Physics-Informed Operator Learning (PIOL) pathways via constitutive behavior, including FEM labeling, decoupled modeling, and PIOL evaluating. The PIOL framework uses a shared trunk to represent geometry-conditioned basis functions, while employing heterogeneous branches matched to constitutive behavior: a lightweight CNN-MLP branch for transient elastic response and a CNN-LSTM branch for history-dependent plasticity and stress. A constitutive constraint term further regularizes the thermal-strain prediction. Finally, through thermally calibrated FEM simulated dataset across multiple process conditions, the PIOL scheme evaluated on the benchmark demonstrate improved stress prediction accuracy relative to baseline variants, with millisecond-scale inference suitable for online digital-twin deployment.