Anchor-Oracle Fourier Neural Operator Framework for 3D Membrane Field Prediction in PEM Fuel Cells: From Steady-State Operation to Transient Degradation
4 affiliations
- 1Hub de l’Energie, UMR CNRS, 15 rue Baudelocque, Amiens Cedex, France, 80039
- 2AVL List GmbH, Hans-List-Platz 1, Graz, Austria, 8020
- 3UMR CNRS, 15 rue Baudelocque, Amiens Cedex, France, 80039
- 4Institut Universitaire de France, 1 rue Descartes, Paris Cedex, France, 75231
Abstract
Membrane durability limits the service life of proton exchange membrane fuel cells, but the average degradation rate does not capture the degradation spatial heterogeneities. Existing approaches either resolve these spatial features at high computational cost (three-dimensional computational fluid dynamics modeling) or replace it with a scalar or curve-level surrogate; neither is fast enough to be queried across a wide spectrum of cell operation conditions. We develop an Anchor-Oracle Fourier Neural Operator, a three-stage neural operator that predicts the complete three-dimensional fields of membrane ionic conductivity and ionic current density from operating conditions and degradation time, achieving steady-state accuracy within one percent at four orders of magnitude lower cost than the reference solver. Using the surrogate as a full-field probe, we show that operating conditions reorganize the two fields' spatial heterogeneity along different axes, and that degradation rate and spatial heterogeneity vary non-monotonically with each other. We further find that current is progressively redistributed away from degraded regions over space and time. This redistribution is invisible to a single pooled correlation, and it advances together with the degradation front. We then use the surrogate to construct an inverse-design operating map, showing influences of different boundary conditions on power and durability.
Related
- Iridium Oxidation-State Dynamics during Transient Operation under PEM Water Electrolysis Single-Cell Conditions Revealed by Time-Resolved Operando XAS
- Region-Aware Degradation-Conditioned Vision-Language Learning for Blind Image-Quality Assessment of Power Field-Operation Imagery
- Zero-Gap Tandem Photoelectrochemical Cell Based on Nanostructured WO3 and CuFe2O4 Photoelectrodes and a Proton-Exchange Membrane: Bias-Free Photoelectrochemical Operation and Degradation Analysis
- A Shared-Task-State-Driven Multi-Agent Framework for Knowledge Graph Construction in Substation Operation and Maintenance
- Spatially resolved 3D aging model linking interdependent degradation mechanisms to knee and end-of-life prediction
BibTeX
@misc{zhang2026anchororacle,
title = {{Anchor-Oracle Fourier Neural Operator Framework for 3D Membrane Field Prediction in PEM Fuel Cells: From Steady-State Operation to Transient Degradation}},
author = {Xingyu Zhang and Robert Pöschl and Alejandro A. Franco},
howpublished = {ChemRxiv},
year = {2026},
month = aug,
doi = {10.26434/chemrxiv.15008100/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008100/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Anchor-Oracle Fourier Neural Operator Framework for 3D Membrane Field Prediction in PEM Fuel Cells: From Steady-State Operation to Transient Degradation
AU - Xingyu Zhang
AU - Robert Pöschl
AU - Alejandro A. Franco
PY - 2026
DA - 2026/08/31
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008100/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008100/v1
AB - Membrane durability limits the service life of proton exchange membrane fuel cells, but the average degradation rate does not capture the degradation spatial heterogeneities. Existing approaches either resolve these spatial features at high computational cost (three-dimensional computational fluid dynamics modeling) or replace it with a scalar or curve-level surrogate; neither is fast enough to be queried across a wide spectrum of cell operation conditions. We develop an Anchor-Oracle Fourier Neural Operator, a three-stage neural operator that predicts the complete three-dimensional fields of membrane ionic conductivity and ionic current density from operating conditions and degradation time, achieving steady-state accuracy within one percent at four orders of magnitude lower cost than the reference solver. Using the surrogate as a full-field probe, we show that operating conditions reorganize the two fields' spatial heterogeneity along different axes, and that degradation rate and spatial heterogeneity vary non-monotonically with each other. We further find that current is progressively redistributed away from degraded regions over space and time. This redistribution is invisible to a single pooled correlation, and it advances together with the degradation front. We then use the surrogate to construct an inverse-design operating map, showing influences of different boundary conditions on power and durability.
ER -