Machine learning approaches have become increasingly important in engineering research and development in recent years, and the first steps in this field can also be found at proton exchange membrane water electrolysis (PEMWE) development [1]. Well-known conventional simulation models for engineers are usually based on physical and empirical equations, with different approaches depending on the desired model depth. However, especially for the simulation of complex processes, model formulation and parametrization with sufficient accuracy are often challenging, with these types of models reaching their limits. In contrast, data-driven models use machine learning approaches to reproduce the desired processes based on large data sets and do not require physical or empirical model formulations [2]. For decision making on the deployment of machine learning methods, the effort required for familiarization, conceptual design, and implementation is the crucial aspect to answering the question of beneficial gain. Our conference contribution aims to answer these questions regarding invested effort and resulting benefit through a self-experiment using the example of artificial neural networks (ANN) in PEMWE research. It is intended to guide engineers interested in machine learning methods without prior knowledge, addressing their questions and ultimately assisting them in decision-making.
Our proposed research is shown in Figure 1 and consists of conceptualizing and programming ANN for two application cases, one for high-frequency resistance (HFR) regression and the other for PEMWE single-cell aging prediction. Here, the processes of familiarization time, conceptualization, and implementation of ANN are transparently presented from an engineer's perspective. The modeling results are evaluated in terms of their utility to the user. The data sets for training and validating the ANN to calculate HFR are generated using a physics-based simulation model. For aging prediction, the datasets are based on experimental data obtained from accelerated stress tests (AST) of a PEMWE single cell.
Our contribution finalizes with an assessment of the profit growth of engineering applications for pioneers in the field of machine learning methods.
Figure 1. Scheme for the evaluation of PEMWE single cells by artificial neural networks from an engineering perspective.
References:
[1] Bensmann, B., et al. "An engineering perspective on the future role of modelling in proton exchange membrane water electrolysis development." Current Opinion in Chemical Engineering (2022) (accepted)
[2] Mistry, A., et al. "How machine learning will revolutionize electrochemical sciences." ACS energy letters 6.4 (2021): 1422-1431.
