Bayesian-optimization-assisted efficient operation for direct ammonia solid oxide fuel cells

The direct ammonia solid oxide fuel cell (DA-SOFC) is a promising energy conversion device that directly utilizes the ammonia as a fuel through the ammonia cracking inside the fuel electrode. However, because given that various operating parameters related to the electrochemical performance and ammonia cracking reaction, which significantly affects the efficiency and performance of DA-SOFCs, have multi-dimensional and complex correlations under multi-physics situations, considerable time, cost, and manpower investments are required to find the optimum operating conditions. Here, we demonstrate the effectiveness of Bayesian optimization (BO), an iterative response-model-based global optimization algorithm, when used for the rapid determination of the optimum operating conditions of DA-SOFCs. First, we compared nine BO-based models with 1,140 collected data sets for the ground truth, with four-dimensional variable conditions, i.e., the temperature, voltage, fuel flow rate, and the ammonia mole fraction. To select the appropriate BO model, we evaluated the performance of each model, which have the surrogate models (Matérn kernel 3/2 and 5/2, and radial basis function) and the acquisition functions (Probability of improvement (PI), Expected improvement, and Lower confidence bound). Among them, the surrogate model consisting of a Matérn kernel 3/2 and the acquisition function of PI exhibits the highest performance with high reliability when tasked with finding the global optimum with ground-truth data. Thereafter, this model is introduced to find the optimum operating conditions without the ground truth. With only four experimental trials, it finds a better operating condition that shows 10–20 % higher current density compared to that under conventionally used conditions.