Physics-informed neural network assessment and multi-objective design selection for thermoelectric coolers

Thermoelectric cooler (TEC) design couples Peltier pumping, Fourier back-conduction and Joule heating, making Pareto solutions sensitive to calibration, bounds and omitted transport. We combine a manufacturer-calibrated lumped model, an algebraic physics-informed neural network (PINN) surrogate audit that does not solve a partial differential equation, direct multi-objective optimization and requirement-based selection. Thermal scaling fitted at 5.2 A yielded withheld cooling-capacity RMSE values of 0.601 W at 1.7 A and 1.410 W for nonzero-lift 8.7 A points. With 40,000 samples, surrogate holdout RMSE values were 0.655 W for cooling capacity and 1.174 W for input power. A matched ablation confirmed that the algebraic residual adds no independent information at labelled points. Collocation reduced mean sparse-label errors in all five paired runs, but the exact tests gave p = 0.0625 two-sided and p = 0.03125 for the directional alternative, limiting inference. At N = 127 , MOPSO achieved the best mean HV and IGD + in a matched four-algorithm benchmark. The source-bounded balanced design gave COP = 0.854 and PD = 55.21 kW m − 2 . Relaxing bounds raised PD to 76.34 kW m − 2 without increasing COP. Contact resistance changed COP ranking and selected geometry; a one-dimensional Thomson audit preserved Pareto membership but changed exact selections. The workflow supports conditional TEC design selection within the stated bounds.