Instructor: Peter Yichen Chen (peter.chen@ubc.ca)
TA: Weijia Zeng (owenzeng@student.ubc.ca)
Class Time: MW 15:30-17:00
Meeting times: MW 15:30-17:00
Classroom location:Â CEME 1212
Communication: Please join this slack group for asking questions and finding teammates. This is required as you need to turn in your slides on slack to receive credit for your presentation.
Pre-class survey: https://forms.gle/5EpXeQ12fMvDHexVA
Class Overview
This course explores the integration of modern machine learning techniques with physics-based modeling and simulation. Structured around foundational principles and emerging research, we will examine differentiable simulations, physics-informed learning, reduced-order modeling, and generative AI methods for physical systems. Topics span from graph neural networks and neural operators to diffusion models and AI-driven scientific discovery. Students will engage with key readings and hands-on projects, culminating in the creation of a differentiable simulator for a chosen application. Case studies include 3D content creation, engineering design/control, and scientific material discovery.
Seminar Structure
Unlike traditional lecture-style courses, this research seminar uses a role-playing format to immerse students in the research lifecycle. Each week, we’ll read a technical paper on learning and simulation. During role-play, students will rotate through various positions. Throughout the semester, each student should sign up for roughly 10 presentations.
Role-playing (Heavily borrowed from Columbia’s E6998)
Every week, we will focus on one/two papers and organize the discussion around different "roles" played by students: paper reviewer, archaeologist, industry expert, and hacker. All role players should make some slides and turn them in on slack.
Paper reviewer. Complete a full, critical, but constructive (not negative) review of the paper. Answer all questions of the NeurIPS Review Form.
# of roles: 3.
In class: report your reviews to the rest of us in 20 min (total).
Archaeologist. A research paper should not be reviewed as a piece of isolated scientific work. Investigate where the paper sites in the context of previous and subsequent work. Find and report one older paper cited by the discussed paper, and tell us why this discussed paper advances the technique. Meanwhile, find a subsequent paper (newer than the discussed one) and report why it merits a publication. Chaining the series of papers will help to reveal a chronological flow of technical advancement.
# of roles: 1.
In class: report your finding with your slides in 15 min (total).
Industrial R&D expert. Convince us (your industry bosses) that the paper can bring technical strength to the company. So it worth the company's resource to integrate the technique into our commerical products or pipeline. You can choose an appropriate company and product or application (e.g., software company like Audodesk or animation studio like Pixar).
# of roles: 1.
In class: Present your argument with your slides in 15 min (total).
Hacker. Implement a small part of the paper or simplified (e.g. 2D instead of 3D) version of the paper. If the paper introduces a new system (e.g. Taichi), then you can also design some cool demo with it. Prepare a demo of your work for the class.Â
# of roles: 2.
In class: Present your demo/slides in 25 min (total).
This format follows the role-playing seminar model proposed by Jacobson and Raffel: https://colinraffel.com/blog/role-playing-seminar.html, and is inspired by similar courses such as Columbia’s E6998 (https://www.cs.columbia.edu/~cxz/teaching/E6998_f16/), USC's CSCI-699 (https://odedstein.com/teaching/hs-2023-csci-699/index.html), and UC Berkeley’s CS294-173 (https://sites.google.com/berkeley.edu/learningfor3d-seminar/schedule?authuser=0)
Prerequisites/Corequisites
Advanced undergraduate and beginning graduate students in computer science and applied sciences.
Course Goal
Develop differentiable simulations by integrating physical principles with data-driven methods using modern differentiable programming frameworks.
Critically compare modeling approaches along the spectrum from physics-based to purely data-driven, understanding their trade-offs in accuracy, generalization, and interpretability.
Apply deep learning techniques—including graph neural networks, neural operators, and transformers—to problems in physics, engineering, and scientific computing.
Evaluate and select computational tools based on criteria such as simulation speed, scalability, and fidelity to physical phenomena.
Design and manage open-ended research projects, from proposal to presentation, while demonstrating collaboration, planning, and iteration.
Communicate complex technical ideas through oral presentations, written reports, and reproducible experiments.