Microstructure-Centric Inverse Design of PEMFC Catalyst Layers via a Hybrid Generative–Bayesian ML Pipeline

Prediction and inverse design of the polarization curve of a membrane electrode assembly (MEA) in a proton exchange membrane fuel cell (PEMFC) remain fundamentally limited by the absence of an analytical relationship linking catalyst ink properties and deposition parameters to the resulting catalyst-layer (CL) microstructure and, ultimately, electrochemical performance [1]. The traditional empirical trial-and-error approach is described in the literature as a “serious limitation to efficiency” [1]. This work proposes a hybrid ML pipeline that circumvents this limitation by using CL microstructure as an ontologically appropriate intermediate link—the only modality that is both observable (via electron microscopy) and causally related to the polarization curve. The feature space is divided into nine fabrication and design parameters and a separate operating-condition vector . The pipeline combines four models: (1) a conditional diffusion model that generates ensembles of synthetic microstructures (up to 16 million images) using an Adaptive Parameter-Space Refinement (APSR) mechanism with an initial 3-bit discretization; (2) a predictive model that selects candidates based on the target polarization curve and reproducibility; (3) a Bayesian inverse-inference model that identifies multimodal distributions of equifinal recipes; and (4) a fast surrogate that distills the entire cycle for continuous optimization. The loop is closed through real-world experiments and fine-tuning of all models. We show that quantitative literature data on the effects of water saturation and ionomer-film thickness on polarization curves [3,4] support the choice of microstructure as the target modality, while recent studies on generative models for electrode microstructures and Bayesian inverse inference confirm the technical feasibility of each component [5–8]. The main contribution of this work is the systematic integration of disparate state-of-the-art methods into a closed loop in which equifinality is treated not as an undesirable ambiguity but as a controllable resource for process flexibility.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24351509
Primary Topic
Fuel Cells and Related Materials
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Microstructure-Centric Inverse Design of PEMFC Catalyst Layers via a Hybrid Generative–Bayesian ML Pipeline

Artur F Khairullin
2 links
Zenodo (CERN European Organization for Nuclear Research)
Fuel Cells and Related Materials
preprint

Microstructure-Centric Inverse Design of PEMFC Catalyst Layers via a Hybrid Generative–Bayesian ML Pipeline

Artur F Khairullin
preprint en

Abstract

Prediction and inverse design of the polarization curve of a membrane electrode assembly (MEA) in a proton exchange membrane fuel cell (PEMFC) remain fundamentally limited by the absence of an analytical relationship linking catalyst ink properties and deposition parameters to the resulting catalyst-layer (CL) microstructure and, ultimately, electrochemical performance [1]. The traditional empirical trial-and-error approach is described in the literature as a “serious limitation to efficiency” [1]. This work proposes a hybrid ML pipeline that circumvents this limitation by using CL microstructure as an ontologically appropriate intermediate link—the only modality that is both observable (via electron microscopy) and causally related to the polarization curve. The feature space is divided into nine fabrication and design parameters and a separate operating-condition vector . The pipeline combines four models: (1) a conditional diffusion model that generates ensembles of synthetic microstructures (up to 16 million images) using an Adaptive Parameter-Space Refinement (APSR) mechanism with an initial 3-bit discretization; (2) a predictive model that selects candidates based on the target polarization curve and reproducibility; (3) a Bayesian inverse-inference model that identifies multimodal distributions of equifinal recipes; and (4) a fast surrogate that distills the entire cycle for continuous optimization. The loop is closed through real-world experiments and fine-tuning of all models. We show that quantitative literature data on the effects of water saturation and ionomer-film thickness on polarization curves [3,4] support the choice of microstructure as the target modality, while recent studies on generative models for electrode microstructures and Bayesian inverse inference confirm the technical feasibility of each component [5–8]. The main contribution of this work is the systematic integration of disparate state-of-the-art methods into a closed loop in which equifinality is treated not as an undesirable ambiguity but as a controllable resource for process flexibility.

Zenodo (CERN European Organization for Nuclear Research)
Affordable and clean energy
Fuel Cells and Related Materials
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon