A multi-physics modeling framework, which includes electrochemical and chemical reactions, mass transfer, and energy balance, has been developed and validated against experiment measurements to investigate the performance of solid oxide CO2-steam co-electrolysis (SOEC) under various operating conditions and cell designs. However, multi-physics modeling for complex SOEC systems can be computationally expensive and not tractable for general system design and optimization. Reduced-order models (ROMs) have been proven as a powerful tool for reducing computational costs, closely mimic the high-fidelity SOEC stack models and identify the operational condition parameters to which the SOEC stack is sensitive and needs appropriate controls for stable operation. In this study, the deep neural networks (DNN) algorithm is employed to construct ROMs according to multi-physics simulations for SOECs to systematically investigate the SOECs’ electrochemical performance and rank the contributions of operational condition parameters for both 2 cm2 button cell and 300 cm2 planar cells. Sensitivity analysis reveals that cell voltage, fuel composition (denoted by CO2/H2O ratio), and the operating temperature impact the cell performance most. The product ratio ΔCO/ΔH2 has a stronger dependence on the CO2/H2O ratio in the fuel compositions than on other operating condition parameters. The large 300 cm2 planar cells were modeled under both adiabatic and furnace environment. The simulation and sensitivity analysis indicate that the large cell operating in adiabatic environment usually provides relatively lower performances than the one in furnace environment. Additionally, the adiabatic environment also causes the cell’s performances are more sensitive to the operation conditions, such as external voltage, fuel flow rate, and fuel compositions. The large cell in adiabatic environment also leads to higher internal temperature variation, which may impact the structure reliability.