I’m using a PINN to solve the damped oscillator differential equation and simultaneously find the friction parameter of the latter, given noisy observations of a damped oscillator as an input. I wrote the code using Tensorflow using a custom training routine. The problem is that the trainable parameter that I’m defining isn’t approaching the correct value that I know from the noisy observations. Ultimately, the solution of the PINN is completely incorrect. However, my code works perfectly fine without the side quest of finding a trainable parameter, i.e., the friction parameter here.