Researchers propose a physics-informed neural network (PINN) that estimates key parameters of a DC buck converter from time-domain data. The network identifies inductance, capacitance, and parasitic resistances by embedding the converter's governing equations into the training process, rather than relying on separate measurement setups.
A distinctive feature is the staged release of gradients during training. The order begins with inductance (L), moves to capacitance (C), and then to parasitic resistance (rL), with further stages implied. This sequential approach is designed to improve convergence and parameter identifiability compared to updating all parameters at once.
The work is described in an arXiv preprint and focuses on a single power converter topology. The abstract does not provide quantitative results or comparisons, so the practical gains over conventional methods remain to be detailed in the full paper.