Physically motivated structuring and optimization of neural networks for multi-physics modelling of solid oxide fuel cells

Andreas Rauh (1, 2) , Julia Kersten (3) , Wiebke Frenkel (3) , Niklas Kruse (3) , Tom Schmidt (3)

1 ENSTA Bretagne - École Nationale Supérieure de Techniques Avancées Bretagne
2 Lab-STICC_ROBEX - Equipe ROBotics for EXploration
3 University of Rostock

Neural network models for complex dynamical systems typically do not explicitly account for structural engineering insight and mutual interrelations of various subprocesses that are related to the multiphysics nature of such systems. For that reason, they are commonly interpreted as a kind of data-driven, black box modelling option that is in opposition to a physically inspired equation-based system representation for which suitable parameters are subsequently identified in a grey box sense. To bridge the gap between datadriven and equation-based modelling paradigms, this paper proposes a novel approach for a physics-inspired structuring of neural networks. The derivation of this kind of structuring, an optimal choice of network inputs and numbers of neurons in a hidden layer as well as the achievable modelling accuracy are demonstrated for the thermal and electrochemical behaviour of hightemperature fuel cells. Finally, different network structures are compared against experimental data.

Mots clés: Fuel cell systems; neural networks; function approximation; nonlinear dynamical systems; singular value decomposition; robustness analysis

Domaines: Automatique; Réseau de neurones [cs.NE]; Energie électrique

Andreas Rauh, Julia Kersten, Wiebke Frenkel, Niklas Kruse, Tom Schmidt. Physically motivated structuring and optimization of neural networks for multi-physics modelling of solid oxide fuel cells. Mathematical and Computer Modelling of Dynamical Systems, 2021, ⟨10.1080/13873954.2021.1990966⟩. ⟨hal-03421430⟩