Learning PathPINNNeural OperatorsFinite Element

Physics AI Learning Path

A four-stage learning path from numerical foundations to neural operators — 15 tutorials and surveys sequenced for a smooth ride from zero background to industrial practice.

Updated Oct 12, 2026 15 resources
Introduction

Why this collection

Physics AI resources are scattered across papers, course notes and GitHub repos, and newcomers rarely know where to start. This collection arranges 15 curated resources into four stages, so you always know what to read next.

  • Stage 1 · Foundations — build intuition for numerical simulation before touching neural networks.
  • Stage 2 · Core methods — get hands-on with PINNs and understand why they converge (or don't).
  • Stage 3 · Neural operators — move from solving one PDE to learning a family of solution operators.
  • Stage 4 · Domain deep-dives — see how physics priors land in power systems, robotics and continuous-time models.

Each item carries an editorial note explaining *why* it sits at that position and *what to focus on*.

Stage 1 · Foundations — Numerical Methods & Simulation 3 resources

Physics Simulation & ML Integration Roadmap

Learning Roadmap / Systematic Overview
Systematic path from classical PDE numerics to data-driven physics AI - FEM, FVM, surrogates, PINN and their evolution.
CC BY 4.0 RoadmapPDENumerical Method
💬 Editor note: Start here for the big picture: the evolution logic from finite element and finite volume to surrogates and PINNs — after this you know what every later stage is solving.

Finite Element from Scratch in Python

Finite Element / From Scratch
Build a 2D linear-elastic FEM solver from NumPy/SciPy - no FEniCS/Abaqus - to deeply understand every matrix derivation and assembly step.
GPL-3.0 FEMFinite ElementPython
💬 Editor note: The second must-read: derives and assembles a 2D elastic FEM matrix from scratch in NumPy, with no framework at all. This is where discretisation actually clicks.

Physics-Based Deep Learning (Online Book)

Textbook / Systematic Course
A free online textbook from TU Munich covering physics + deep learning: differentiable physics, PINNs, neural operators and turbulence modeling, with runnable code.
CC BY-NC-SA 4.0 TextbookDifferentiable PhysicsPINN
💬 Editor note: A free online textbook from TUM with runnable code — keep it as a standing reference for the whole path rather than reading it cover to cover.

Stage 2 · Core Methods — PINNs 3 resources

PINN from Scratch

Beginner Tutorial
A Chinese structured tutorial implementing PINN from minimal dependencies, built chapter by chapter around the 1D heat equation — ML basics, separation of variables, FDM/FEM comparisons, PINN model construction and optimizer comparison, GMM and curriculum learning, plus locELM, random-feature and RaNN extensions. MIT licensed - the best start for Chinese-speaking learners.
New MIT PINNTutorialBeginner
💬 Editor note: The first hands-on project of the path: a well-structured tutorial that pairs with the original Raissi repo — write your first PINN from scratch, reading and coding side by side.

Machine Learning with Physics Knowledge for Prediction: A Survey

Survey / Reading Material
A systematic survey of physics-knowledge + ML methods for PDE prediction, splitting the field into architectural-level (loss, structured prediction, augmentation) and data-level (multi-task, meta, contextual) approaches, with industrial perspective and open-source ecosystem review.
arXiv preprint (copyright held by authors) Physics-informed MLPDEMulti-task Learning
💬 Editor note: Come back to the theory once the code runs: a systematic taxonomy of physics-knowledge injection by architecture layer and data layer, plus an industrial deployment perspective and an open-source ecosystem list.

Adaptive Physics-informed Neural Networks: A Survey

Survey / Reading Material
A focused survey of PINN improvements via transfer learning and meta-learning, addressing PINN's notorious convergence difficulty and the need to re-optimize on every PDE parameter change - the engineering reference for productionizing PINNs.
arXiv preprint (copyright held by authors) PINNTransfer LearningMeta Learning
💬 Editor note: Read this when the PINN will not converge: it surveys transfer- and meta-learning improvements — focus on the loss-weighting section to tackle slow convergence and the retrain-every-parameter-change bottleneck.

Stage 3 · Advanced — Neural Operators 5 resources

DeepONet Getting Started

Neural Operator / Beginner Tutorial
Systematic DeepONet theory and DeepXDE implementation - for PINN developers handling multiphysics-parameter scenarios.
MIT DeepONetNeural OperatorDeepXDE
💬 Editor note: The first stop in operator learning: from theory to DeepXDE code, showing how learning a function-to-function map differs fundamentally from a PINN — read it against Stage 2.

FNO (Fourier Neural Operator) Tutorial

Fourier Neural Operator / Beginner Tutorial
Systematic FNO from theory to code - for engineers and researchers wanting neural operators as PDE surrogates.
MIT FNONeural OperatorSurrogate
💬 Editor note: The second stop: the spectral view of Fourier neural operators, whose mesh-invariance is the key advantage over CNN-style methods for PDE surrogates.

Operator Learning: A Survey

Survey / Reading Material
A systematic survey of operator learning covering the theory, architectures and error analysis of DeepONet, FNO and their variants - the reference entry point to learning infinite-dimensional maps.
arXiv preprint (copyright held by authors) Operator LearningSurveyDeepONet
💬 Editor note: Fill in the theory: a systematic treatment of DeepONet, FNO and their variants — theory, architecture and error analysis, the best reference for infinite-dimensional mappings.
💬 Editor note: Aimed at engineering deployment: compares architectures side by side on Poisson and linear-elasticity problems, and demonstrates their use as surrogates for Bayesian inversion.
💬 Editor note: Closes Stage 3: maps the full architecture spectrum along scalability, generalisation and data sensitivity axes, giving you the judgement to choose between them.

Stage 4 · Deep Dives — Domains & Mathematical Foundations 4 resources

💬 Editor note: A power-systems deployment reference: surveys hybrid physics-informed networks for power system design, monitoring and control, with the emphasis on how the physical constraints are embedded.

Embedding Physics Priors in Robot Learning: A Survey

Multibody
A survey on embedding physics priors in robot learning, covering how physical constraints like contact mechanics, rigid-flexible coupling, and multibody dynamics improve sample efficiency and reliability of learned policies.
See LICENSE file robot learningphysics priorsmultibody dynamics
💬 Editor note: Robotics angle: how physical priors raise sample efficiency and reliability of learned policies, covering contact mechanics, rigid-soft coupling and multibody dynamics.
💬 Editor note: Mathematical reinforcement: a unified framework for continuous-time machine learning that clarifies the relation between neural differential equations and stochastic processes — for those who want to dig into theory.
💬 Editor note: A systematic set of course notes with a from-scratch PyTorch implementation — useful as a review and gap-filler for the whole path.