PINN from Scratch

New
License: MIT · Updated 2026-11-05
My rating

The PINN-from-Scratch project implements physics-informed neural networks from minimal dependencies, using the 1D heat equation as the running benchmark. It opens with Python and machine-learning basics, contrasts classical PDE solvers (separation of variables, finite differences, finite elements) with the mesh-free PINN route, then builds the PINN model layer by layer and compares SGD, Adam and L-BFGS. Later chapters add Gaussian-mixture and curriculum-learning strategies for convergence, and extend the family to locELM, random-feature methods (PU-RFM) and random neural networks (RaNN).

Suggested path

  1. Work through the ML-basics and classical-solver chapters to fix the numerical reference
  2. Run the PINN chapters on the 1D heat equation and inspect the residual loss
  3. Reproduce the same problem in DeepXDE; compare hand-written vs framework
  4. Advance to NeuralOperator / industrial surrogates
Tags: