TAGRA 2025

Physics-Informed Neural Networks (PINNs) are a deep learning approach that directly integrates the differential equations governing physical systems into the learning process of artificial neural networks. This method is increasingly preferred in engineering applications due to its low data requirements and its ability to produce more consistent results by embedding physical laws within the model. An RL circuit, on the other hand, consists of a resistor (R) and an inductor (L) connected in series, where the current responds to sudden changes due to the influence of the magnetic field. Accurately modeling how the current changes over time in such circuits is crucial for understanding their dynamic behavior. In this study, current–time data obtained from simulation results were used. The PINN method was applied to predict the time-dependent current response of a DC-powered RL circuit. The differential equation describing the circuit was directly incorporated into the PINN model’s loss function; both the ODE loss and the initial condition loss (i(0) = 0) were employed to ensure the model’s accuracy. The regular decrease of the loss function throughout the training process indicated that the model increasingly approached the analytical solution with each iteration. Notably, by the 5000th epoch, the loss value had dropped to 0.11, demonstrating that the PINN achieved a stable solution. The initially high error gradually decreased and approached minimum values, confirming that the model successfully learned the physical behavior of the system. During training, the model was first trained using the Adam optimization algorithm and then further optimized with the LBFGS method to obtain a more precise solution. This two-stage process enabled the model to reach the correct solution more quickly and reliably. The findings show that PINNs reliably model physical systems and can be effective for analyzing complex circuits or dynamic systems.