Probabilistic and explainable data-driven modelling of Solid-oxide fuel cells

Solid-oxide cell based systems (fuel cells and electrolysers) are one of the most promising hydrogen technologies. This technology offers a unique way of using a single unit both for electricity and heat production i.e SOFC mode, and hydrogen production in SOEC mode. In comparison to other fuel cells technologies, which use platinum catalysts, solid-oxide systems are based on abundant and affordable raw materials, e.g. nickel, steel, and provide high fuel flexibility. Furthermore, among fuel cell technologies, solid-oxide systems have the highest conversion efficiency both in fuel cell as well as in electrolysis regime. As a result considerable efforts have been made in development and optimisation of solid-oxide systems. However, broad commercialisation of this technology is still an issue, with the main challenges being performance and morphology degradation, as well as scale-up. Therefore, the issues of performance optimisation are paramount. Since accurate models are prerequisite for (on-line) performance optimisation, modelling of solid-oxide systems dynamics and future behavior prediction is the main objective of this proposal. Machine learning models are gaining importance in various scientific fields that are predominantly using first principle models, solid-oxide technology being one of them. Such models are gaining traction addressing problems that are challenging, i.e. where first principle models require either significant sophisticated measurement equipment in order to estimate the model's parameters or the background knowledge is limited. Nowadays, we can safely claim that new approaches that can integrate background knowledge with state-of-the-art machine learning methods can provide new ways of addressing such problems. Since neither pure ML modelling nor solely first-principle models can be considered as sufficient for complex problems, the goal is to explore integrated approaches. The use of domain background knowledge is a completely new direction that can provide explainable data-driven models, whose “thirstiness” for data is supplemented with the expert's knowledge. Solid-oxide systems seem to be the perfect candidates for this. On one side, this is emerging and fast evolving technology. On the other hand, there are genuine time, financial and safety limitations that prevent exhaustive testing. As a result, we are confined to working only with limited data-sets. Therefore, applying the integrated data-driven approaches coupled with domain knowledge can have twofold benefits. First it can advance our understanding of solid-oxide systems. Second, it will prove that combining domain knowledge with ML methods is a viable direction.