Year: 2026
Type: article
Abstract: Crystal plasticity is an essential tool for uncovering structure–property relationships in polycrystalline materials, but its computational cost limits optimization-driven workflows. Existing full-field surrogates can reproduce overall stress distributions while inadequately resolving stress concentrations. Here, we employ a standard Fourier neural operator (FNO) to learn the nonlinear mapping from crystallographic orientation and grain morphology to full-field Cauchy stress. We assess global accuracy, hotspot fidelity, residual errors, microstructure distribution shifts, and cross-resolution behavior. A controlled comparison of orientation encodings shows that rotation matrices outperform quaternions, consistent with the continuity of the rotation matrix representation. The operator predicts hotspot locations and magnitudes and maintains hotspot fidelity under shifts in grain morphology and crystallographic texture. Spectral analysis shows that stress field energy is reproduced over a broad range of spatial frequencies, although residuals concentrate near grain boundaries and sharp stress gradients. Zero-shot super-resolution evaluation on finer grids reveals a spectral truncation trade-off: retaining higher Fourier modes improves native resolution hotspot fidelity but produces a discretization-specific artifact, whereas stronger truncation suppresses the artifact while attenuating hotspot magnitudes and grain boundary detail. Transfer learning adapts the pretrained operator to four additional loading conditions while largely preserving hotspot fidelity. Relative to the CP-FFT reference solver, inference is 126 × faster on one CPU core and approximately 80,000 × faster on an NVIDIA H100. These results provide an empirical basis for selecting and evaluating data-driven FNO surrogates for full-field crystal plasticity stress and hotspot prediction.
Institution: RWTH Aachen University
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,830
Funder: RWTH Aachen University
OpenAlex ID: W7212439684