Abstract
A modified physics-informed neural network (PINN) tailored for solving inverse problems in data-driven engineering applications was demonstrated. The inherited PINN framework enabled the network to integrate ill-posed, noisy experimental data while enforcing a wide range of governing equations, initial and boundary conditions. The network was designed to predict system parameters in governing equations based on a limited number of data points down to three. The general network architecture was further refined to predict thermal conductivity and its performance was validated under various cases. Furthermore the prediction of 18650 Li-ion battery cell thermal conductivity values based on experimental temperature measurements was also conducted.