Machine Learning-Based Simulation and Optimization of an Abiotic Alkaline Glucose Fuel Cell

초록

Glucose-based fuel cells offer a sustainable route for low-temperature power generation due to the abundance, low cost, nontoxicity, and renewable origin of glucose. This study develops a machine-learning (ML)-based framework to predict and optimize the performance of an abiotic alkaline glucose fuel cell (AAGFC). A total of 29 ML models were developed using 539 literature-derived experimental data points digitized from a single previously published AAGFC study, without conducting additional experiments, to predict the output voltage, power density (PD), and substrate-to-energy conversion efficiency (SECE) under varying glucose concentration, potassium hydroxide (KOH) concentration, feeding flow rates, temperature, and current density. Among these models, the Matern 3/2 Gaussian process regression (GPR) model exhibited the best predictive performance under conventional random point-wise splitting, achieving near-perfect interpolation performance. Additional leave-one-polarization-curve-out cross-validation (LOPOCV) indicated moderate grouped generalization, with reduced reliability under boundary-condition groups. Sensitivity analyses revealed that temperature and KOH concentration positively influenced performance, whereas excessive glucose concentration and high anode flow rate had detrimental effects. Particle swarm optimization (PSO) and paretosearch were used for single- and multiobjective optimization. Under the computationally optimized conditions, the model predicted a maximum PD (MPD) of similar to 38 mW/cm(2) for the investigated AAGFC configuration, approximately threefold higher than that under the reference conditions, though the corresponding energy efficiency remained limited (similar to 4%). These findings demonstrate the potential of ML-assisted modeling and optimization for AAGFCs while highlighting the need for further improvements in catalyst, electrode, and cell design. Expanding the current single-source dataset with more diverse data is essential to enhance model transferability and practical applicability.

키워드