Simulation Informed Machine Learning Interpretation of Electrochemical Measurements

Abstract

This work resides at the intersection of rigorous physics-based simulation and machine learning. We seek to address problems that have complicated multiphysics and exist in vast parameter spaces. Where the data requirement for machine learning methods can make the experimental burden untenable and the time-sensitive nature or high computational cost limits the practicality of mechanistic physics model. Our overarching objective is to explore solutions to bridge these limitations in electrochemistry research through the integration of machine learning and mechanistic simulation. Failure detection in solid oxide fuel cell (SOFC) is complicated due to the need to disentangle the failure response from the effect of degradation - gradual change in performance with aging. We used physics models to simulate the behavior of SOFC under three failures that could occur during its operation: fuel maldistribution, delamination and oxidant gas crossover. These simulations revealed deviations in electrochemical impedance spectroscopy (EIS) from behavior of standard circuit elements under failures, underscoring the significance of physics-based modeling in SOFC diagnostics. Leveraging synthetic data of a 6-cell sub-stack, we trained a support vector machine to identify failure modes with a 90% accuracy across degradation effects and operating conditions, discerning imperceptible differences in stack-level EIS responses. Investigation of synthetic data offered insights to failure diagnosis with EIS in determining most responsive frequency range and the efficacy of different machine learning methods.

In the second project, we reexamined the utility of a reference electrode positioned outside the current path on a thin solid electrolyte. Extensive prior research with this design had demonstrated significant polarization shifts and half-cell EIS distortions from minor electrode misalignment, limiting its usefulness in quantitative assessment of two half-reactions. Employing a physics model to simulate these behaviors in a proton-exchange membrane electrolyzer, we trained a neural network model with simulated data to deconvolute edge effects and determine the true oxygen kinetic overpotential with high accuracy (r2-score ≥ 0.96). Validation with experimental data from electrolyzer cells of varying membrane thicknesses and misaligned electrodes confirmed the breakdown of oxygen and hydrogen evolution reaction losses to align with literature values. These findings unveil the potential use of this straightforward reference electrode design and intentional anode-cathode misalignment for evaluating individual electrode kinetics in performance or long-term degradation studies.