Fig 6. Diffusivity equation—inverse modeling: PINN performance as function of training set size.

This figure shows the estimated error dependency on the amount of training data, N_tr. The error bars show mean and standard derivation (± 1 SD) based on 24 realizations. Note that there is no noise added in this investigation, and all physical constants are set to one, i.e., θ₁ = θ₂ = 1.0. The reported error values of θ₁ and θ₂ are percentage errors while the relative ${\mathcal{L}}^2$ error is shown for p.