Differentiable Physics & Neural Networks | Guest Lecture for "Computer Graphics in AI Era"

This is a recording of my second guest lecture for CS8803/4803 CGA -- "Computer Graphics in AI Era", a Georgia Tech course taught by Prof. Bo Zhu (https://faculty.cc.gatech.edu/~bozhu/ ). In this lecture, we explore different ways to weave physical principles directly into machine learning algorithms. Along the way, we touch on fascinating ideas like Sparse Identification of Nonlinear Dynamics (SINDy), physics-informed neural networks (PINNs), solver-in-the-loop training, and more. 🌐 Course website: https://cgai-gatech.vercel.app 🖼️ Code demo: https://github.com/bobarna/diff-physi... 📅 Recorded on: April 3., 2025 --- 🎞️ Timestamps: 00:00 Introduction and Overview 03:30 Big Picture: Differentiable Physics Simulation 09:06 Examples 10:49 NNs are Universal Function Approximators 18:17 Example: Learning Material Parameters 22:05 Physics-inspired Loss Functions 35:34 People & Learning Resources 39:32 How to leverage our knowledge (of physics)? 41:50 Spring System & PINNs 48:10 Physics-based Loss Functions in NNs: Pressure Predicting CNNs 52:45 Solver-in-the-Loop training 1:01:03 Learning to Throw (Detailed Example) 1:07:10 Supervised vs. Differentiable Physics 1:11:22 Final Project Ideas & Q&A 📺 Missed the first lecture? You can watch my previous talk on 3D Gaussian Splatting:    • 3D Gaussian Splatting | Guest Lecture for ...   Errata: From 1:05:20, the derivatives shown on the slides are actually correct. When computing dL/dv, the term sin(2θ) is treated as a constant with respect to v. In the last line, we’re differentiating with respect to θ, so the derivative of sin(2θ) becomes 2cos(2θ), with v² acting as a constant factor. I mistakenly mentioned during the lecture that I had written the partial derivatives incorrectly — sorry for any confusion!