Molten carbonate electrolysis modeling and optimization using feedforward neural networks

Molten Carbonate Electrolysis enables high-temperature co-electrolysis of CO 2 and H 2 O to syngas, offering an attractive route for carbon utilization and flexible power-to-X operation under variable renewable electricity. However, physics-based performance models often require extensive parameterization and are not readily applicable for rapid screening, control-oriented prediction, or operating-point optimization. In this work, compact feedforward artificial neural networks are developed as fast surrogate models to predict MCE cell voltage from operating conditions using literature datasets. Four single-hidden-layer ANN models were trained in MATLAB using Levenberg-Marquardt learning with Bayesian regularization and dropout: (i) a temperature-current density model with a 2-3-1 architecture, (ii) a fuel-side composition model with a 5-5-1 architecture, (iii) an oxidant-side composition model with a 4-4-1 architecture, and (iv) a combined thermal-flow model with a 9-5-1 architecture, where the notation denotes input-hidden-output neurons. The proposed models reproduce experimental polarization behavior with unseen test-set relative errors on the order of 0.3-0.4% and capture key nonlinear dependencies across the investigated operating ranges. Finally, the combined 9-5-1 ANN surrogate is embedded in a constrained optimization workflow to identify operating conditions that minimize cell voltage at fixed current density while improving electrochemical efficiency, yielding ANN-derived optimal thermal-flow parameters.