Advanced Deep Learning for Physics (ADL4P)

Welcome to Advanced Deep Learning for Physics (ADL4P)! This course explores cutting-edge techniques at the intersection of physics simulations and AI / deep learning.

This course explains how to combine AI / deep learning techniques and numerical simulation algorithms to simulate, reconstruct and estimate materials such as fluids and deformable objects. In particular, this course will focus on advanced deep learning concepts such as generative / foundation models and time series prediction, with possible applications in many fields, from engineering over medical to computer graphics and vision.

Course Structure

Lectures: Weekly
Exercises/Homework: Weekly coding assignments based on Jupyter notebooks and Python

Topics Covered:

Introduction to Physics-based Deep Learning
Neural Surrogates, Operators and Architecturs
Physical Loss Terms
Differentiable Physics
Graph Neural Networks, Foundation Models
Diffusion Models and Score-based Methods

Lecture

Topic	Slides	Recording
Introduction	Lecture 01	Recording
Supervised Learning	Lecture 02	Recording
Architectures, Differentiable Physics I	Lecture 03a	Recording
Differentiable Physics II	Lecture 03b	Recording
Differentiable Physics II (cont'd)		Recording
Graph-based NNs I	Lecture 04a	Recording
Graph-based NNs II	Lecture 04b	Recording
Graph-based NNs III	Lecture 04c	Recording
SBI and Generative Models I	Lecture 05a	Recording
SBI and Generative Models II	Lecture 05b	Recording
Reinforcement Learning	Lecture 06	Recording
Foundation Models, Conclusions	Lecture 07	Recording

Tutorials

Week	Exercise
Week 1	ADL4P Ex1 - Introduction to Phiflow
Week 2	ADL4P Ex2 - Convergence rate and Momentum
Week 3	ADL4P Ex3 - Sphere Packing
Week 4	ADL4P Ex4 - Supervised Network Training
Week 5	ADL4P Ex5 - Manual Differentiation
Week 6	ADL4P Ex6 - Auto Differentiation
Week 7	ADL4P Ex7 - Optimal Path
Week 8	ADL4P Ex8 - GNNs
Week 9	ADL4P Ex9 - Diffusion
Week 10	ADL4P Ex10 - Kuramoto Sivashinsky Simulator
Week 11	ADL4P Ex11 - Kuramoto Sivashinsky Learning

Prerequisites

You should be familiear with machine learning basics, e.g., from Deep Learning Book (Goodfellow et al.)
Mathematical background knowledge is recommended: linear algebra, calculus, partial differential equations Numerical Methods for Scientific Computing (van Kan et al.)
Basic understanding of physics simulations
Knowledge of Python is required for the exercises