Electrical conductivity of porous layers in reversible fuel cells based on transfer learning
Chinese Science Bulletin · 01.03.2026
DOI: 10.1360/csb-2026-0141 ↗Kurz gefasst
Fachartikel in der Zeitschrift Chinese Science Bulletin, online erschienen am 22.04.2026. Verfasst von Yiquan Rui und 4 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben.
Abstract (Originalsprache)
: Driven by global initiatives toward a green and low-carbon energy transition, reversible proton exchange membrane fuel cells (RPEMFCs) have become an attractive technology for energy storage and conversion due to their dual-function capability in both power generation and hydrogen production. The porous transport layer (PTL) is the core component of RPEMFCs, and its physical properties are important to the gas, water, heat and electricity transport processes. Traditional methods to obtain these properties, such as experiments and numerical simulations, are time-consuming and high-cost, limiting the efficient optimization of PTLs. To address this challenge, this study proposes a data-driven approach combining machine learning (ML) and transfer learning to predict the through-plane (TP) effective conductivity of PTLs. To obtain the data set, microstructural models of two typical PTL materials (carbon paper and titanium felt) were constructed using the random reconstruction method. Numerical simulations were performed to generate a standardized dataset: 105 groups for carbon paper (low computational cost, easy data acquisition) and 36 groups for titanium felt (high computational cost, scarce data). Four ML models—support vector regression (SVR), eXtreme Gradient Boosting (XGBoost), kernel ridge regression (KRR), and multilayer perceptron (MLP)—were established, with hyperparameters optimized via Bayesian optimization (BO) to enhance model performance. For the small-sample prediction of titanium felt conductivity, a transfer learning strategy was adopted: a pre-trained BO-MLP model from the carbon paper dataset (source task) was fine-tuned on the titanium felt dataset (target task) to leverage shared knowledge between the two tasks. Additionally, SHAP (SHapley Additive exPlanations) analysis was employed to interpret the model’s decision-making mechanism and identify key influencing factors. The results demonstrate that BO significantly improves the performance of all ML models. The optimized BO-MLP model achieves a coefficient of determination ( R 2 ) of 0.9981 and a root mean square error (RMSE) of 0.0015 for carbon paper conductivity prediction, outperforming other models. For titanium felt, the transfer learning model yields an R 2 of 0.9979 and an RMSE of 0.0015, which is 64% lower than the model trained directly on the small titanium felt dataset ( R 2 =0.9844, RMSE=0.0042). SHAP analysis reveals that porosity is the dominant factor regulating the conductivity of both materials, exerting a nonlinear effect: lower porosity enhances conductivity by promoting fiber contact, and its influence is remarkably more sensitive in the low-porosity range—small variations in porosity within this range induce more significant changes in conductivity compared to equivalent variations in the high-porosity range. For carbon paper, thickness has a slightly greater impact than fiber diameter, while for titanium felt, fiber diameter (with a wider variation range of 20–40 μm) outweighs thickness due to its stronger regulation of electron transport paths. Its influence is remarkably more sensitive in the low-porosity range—small variations in porosity within this range induce more significant changes in conductivity compared to equivalent variations in the high-porosity range This study provides a novel paradigm for low-cost and high-precision prediction of PTL transport properties. The combination of Bayesian optimization and transfer learning effectively addresses the challenges of low hyperparameter tuning efficiency and small-sample scarcity in ML-based predictions. The interpretability provided by SHAP analysis, particularly regarding the nonlinear regulation of porosity and its enhanced sensitivity in the low-porosity range, also offers theoretical guidance for the structural optimization of PTLs in RPEMFCs.
Zitieren
APA
Rui, Y., Wang, B., Du, Q., Liu, Z., & Jiao, K. (2026). Electrical conductivity of porous layers in reversible fuel cells based on transfer learning. Chinese Science Bulletin. https://doi.org/10.1360/csb-2026-0141
Vancouver
Rui Y, Wang B, Du Q, Liu Z, Jiao K. Electrical conductivity of porous layers in reversible fuel cells based on transfer learning. Chinese Science Bulletin. 2026. doi:10.1360/csb-2026-0141
BibTeX
@article{Rui2026,
author = {Rui, Yiquan and Wang, Bowen and Du, Qing and Liu, Zhi and Jiao, Kui},
title = {Electrical conductivity of porous layers in reversible fuel cells based on transfer learning},
journal = {Chinese Science Bulletin},
year = {2026},
doi = {10.1360/csb-2026-0141}
}
EUVIMED