A hybrid Fourier neural operator–convolutional surrogate for predicting geometrically necessary dislocation density in polycrystalline aluminum
Abstract
Geometrically necessary dislocation (GND) density (ρGND), which arise from the accommodation of plastic strain gradients, play an important role in strain hardening and damage initiation in polycrystals, but the computational cost of full-field crystal plasticity simulations can limit their applicability to design-scale applications. We train deep-learning surrogates on 1394 crystal plasticity fast Fourier transform (CPFFT) simulations of aluminum microstructures to predict local ρGND density from crystal orientation, dislocation density, and strain. On 140 held-out test structures, a baseline convolutional neural network (CNN) produces spatially flat predictions and fails to reproduce grain-boundary pile-ups at any strain (R^2GB = −1.71 ± 0.24, worse than predicting the mean). A Fourier neural operator (FNO) recovers the correct GND network topology but systematically under-predicts its magnitude (R^2GB = 0.85±0.05). A hybrid FNO–CNN architecture, coupling spectral mixing with a local convolutional branch, achieves the best fidelity (R^2GB = 0.88 ± 0.02), cutting grain-boundary RMSE (GB-RMSE) by 32% and 11% relative to the CNN and FNO. These results demonstrate that spectral–local coupling is key to reproducing the localized GND pile-ups driving strain hardening and damage initiation at grain boundaries. Accurate, low-cost ρGND predictions could ultimately enable rapid determination of internal stress and strain-gradient fields governing local yielding at grain boundaries.
Related
- Quantum convolutional neural network for predicting nuclear charge radii
- Predicting plasmonic pole representations of complex dielectric spectra using convolutional neural networks
- Predicting spatiotemporal coke growth in an industrial ethane cracker with a hybrid mechanistic–neural surrogate of its plant-tuned SPYRO model
- Predicting Weight Windows for Duct-Streaming Shielding Geometries with a Convolutional Neural Network
- Charge-to-spin conversion in epitaxial and polycrystalline Bi and Bi/Ag layers
BibTeX
@misc{woguem2026hybrid,
title = {{A hybrid Fourier neural operator–convolutional surrogate for predicting geometrically necessary dislocation density in polycrystalline aluminum}},
author = {Yen Fred Woguem and Brieuc Conan-Guez and Frédéric Pennerath and Lydia Boudjeloud-Assala and Vincent Taupin and Antoine Guitton},
howpublished = {HAL},
year = {2026},
month = oct,
url = {https://hal.univ-lorraine.fr/hal-05757043v1},
note = {CC BY-ND 4.0},
}
RIS
TY - UNPB
TI - A hybrid Fourier neural operator–convolutional surrogate for predicting geometrically necessary dislocation density in polycrystalline aluminum
AU - Yen Fred Woguem
AU - Brieuc Conan-Guez
AU - Frédéric Pennerath
AU - Lydia Boudjeloud-Assala
AU - Vincent Taupin
AU - Antoine Guitton
PY - 2026
DA - 2026/10/04
PB - HAL
UR - https://hal.univ-lorraine.fr/hal-05757043v1
AB - Geometrically necessary dislocation (GND) density (ρGND), which arise from the accommodation of plastic strain gradients, play an important role in strain hardening and damage initiation in polycrystals, but the computational cost of full-field crystal plasticity simulations can limit their applicability to design-scale applications. We train deep-learning surrogates on 1394 crystal plasticity fast Fourier transform (CPFFT) simulations of aluminum microstructures to predict local ρGND density from crystal orientation, dislocation density, and strain. On 140 held-out test structures, a baseline convolutional neural network (CNN) produces spatially flat predictions and fails to reproduce grain-boundary pile-ups at any strain (R^2GB = −1.71 ± 0.24, worse than predicting the mean). A Fourier neural operator (FNO) recovers the correct GND network topology but systematically under-predicts its magnitude (R^2GB = 0.85±0.05). A hybrid FNO–CNN architecture, coupling spectral mixing with a local convolutional branch, achieves the best fidelity (R^2GB = 0.88 ± 0.02), cutting grain-boundary RMSE (GB-RMSE) by 32% and 11% relative to the CNN and FNO. These results demonstrate that spectral–local coupling is key to reproducing the localized GND pile-ups driving strain hardening and damage initiation at grain boundaries. Accurate, low-cost ρGND predictions could ultimately enable rapid determination of internal stress and strain-gradient fields governing local yielding at grain boundaries.
ER -