In July 2026 I spent five days at the Deep Learning for Science (DL4SCI) Summer School, hosted by Computing Sciences at Berkeley Lab and run with NERSC on the UC Berkeley campus. This year's edition was organized around three ideas that are reshaping how science is done with machine learning: foundation models, reasoning, and agentic AI. The program followed the full lifecycle of a scientific model, from data and training at scale to adaptation and evaluation, with a recurring emphasis on what actually works on HPC systems.
These pages are my write-up of the lectures, one post per day. I wrote them for two readers: my future self, and anyone who could not attend and wants a technical but readable account of what was discussed.
How to read these notes
Most of the material is lecture notes: what I captured live from the talks, edited into prose. Where I could not take notes during a session, I wrote a background primer afterwards that explains the topic from general knowledge. Those sections carry a visible banner and do not represent the speaker's content. Every day ends with key takeaways and open questions, and the references page collects the models, methods and papers mentioned across the week.
Themes of the week
- Generative models for science. Diffusion and flow models for weather, molecules and proteins, how to make them fast through distillation, and how to make them respect hard physical constraints.
- Reasoning and reinforcement learning. From the basics of MDPs to RLHF, DPO, RL with verifiable rewards and GRPO, and what "reasoning" means for a model.
- World models and structured data. Mathematical AI, 4D world models for physical systems, and foundation models for time series and tabular data.
- Agents for science. Trustworthy agentic systems in high-stakes clinical settings, and how natural-language feedback can be used as an optimization signal.
The five days
Agenda
Times are local (Berkeley, CA). Italic rows are breaks / events.
Monday, July 20, 2026 — Day 1
| Start | End | Event / Talk — Speaker |
|---|---|---|
| 8:45 AM | 9:00 AM | Breakfast & Coffee |
| 9:00 AM | 9:10 AM | Welcome from LBNL — Jonathan Carter |
| 9:10 AM | 9:15 AM | Safety Talk — Moe Abdelwahhab |
| 9:15 AM | 10:30 AM | Generative AI for Science — Arash Vahdat |
| 10:30 AM | 11:00 AM | Coffee Break |
| 11:00 AM | 12:30 PM | Accelerated Diffusion Models — Julius Berner |
| 12:30 PM | 2:00 PM | Lunch & Events |
| 2:00 PM | 3:30 PM | Transformers, Scaling and Hardware — Jacob Austin |
| 3:30 PM | 4:00 PM | Coffee Break |
| 4:00 PM | 4:30 PM | Science Application Talk — Dar Mehta |
| 4:30 PM | 5:00 PM | Science Application Talk — Alex Morehead |
Tuesday, July 21, 2026 — Day 2
| Start | End | Event / Talk — Speaker |
|---|---|---|
| 8:45 AM | 9:00 AM | Breakfast & Coffee |
| 9:00 AM | 10:30 AM | Discrete Diffusion — Ferdinando Fioretto |
| 10:30 AM | 10:40 AM | Welcome from NERSC — Sudip Dosanjh |
| 10:40 AM | 11:00 AM | Coffee Break |
| 11:00 AM | 12:30 PM | Foundations — Andrew Wilson |
| 12:30 PM | 2:00 PM | Lunch & Events |
| 2:00 PM | 3:30 PM | Science is still not AGI-pilled enough — Daniel Freeman |
| 3:30 PM | 4:00 PM | Coffee Break |
| 4:00 PM | 6:00 PM | Reception & Poster Presentations |
Wednesday, July 22, 2026 — Day 3
| Start | End | Event / Talk — Speaker |
|---|---|---|
| 8:45 AM | 9:00 AM | Breakfast & Coffee |
| 9:00 AM | 10:30 AM | Reasoning — Aditya Grover |
| 10:30 AM | 11:00 AM | Coffee Break |
| 11:00 AM | 12:30 PM | Introduction to RL — Elynn Chen |
| 12:30 PM | 2:00 PM | Lunch & Events |
| 2:00 PM | 3:30 PM | Understanding the Visual World Through Physical Intrinsics — Joseph Redmon |
| 3:30 PM | 4:00 PM | Coffee Break |
| 4:00 PM | 4:30 PM | Science Application Talk — Danielle Maddix Robinson |
| 4:30 PM | 5:00 PM | Science Application Talk |
Thursday, July 23, 2026 — Day 4
| Start | End | Event / Talk — Speaker |
|---|---|---|
| 8:45 AM | 9:00 AM | Breakfast & Coffee |
| 9:00 AM | 10:30 AM | The Rise and Ramifications of Artificial Mathematical Intelligence — Tony Feng |
| 10:30 AM | 11:00 AM | Coffee Break |
| 11:00 AM | 12:30 PM | Scaling 4D World Models for Scientific AI: Balancing Physics and Data — Peter Yichen Chen |
| 12:30 PM | 2:00 PM | Lunch & Events |
| 2:00 PM | 3:30 PM | Reasoning / Time series / Tabular Data — Bernie Wang |
| 3:30 PM | 4:00 PM | Coffee Break |
| 4:00 PM | 5:00 PM | World Models — Jiajun Wu |
Friday, July 24, 2026 — Day 5
| Start | End | Event / Talk — Speaker |
|---|---|---|
| 8:45 AM | 9:00 AM | Breakfast & Coffee |
| 9:00 AM | 10:30 AM | Trustworthy AI Agents for Science — Sheng Liu |
| 10:30 AM | 11:00 AM | Coffee Break |
| 11:00 AM | 12:30 PM | (Multi-)Modality, Accuracy and Creativity of Diffusion Models — Molei Tao |
| 12:30 PM | 2:00 PM | Lunch & Events |
Confirmed Lecturers
| Lecturer | Affiliation |
|---|---|
| Andrew Wilson | NYU |
| Joseph Redmon | Ai2 |
| Jiajun Wu | Stanford University |
| Bernie Wang | MIT |
| Arash Vahdat | Nvidia |
| Jacob Austin | Anthropic |
| Tony Feng | UC Berkeley / Google DeepMind |
| Peter Yichen Chen | University of British Columbia |
| Elynn Chen | NYU |
| Sheng Liu | Stanford University |
| Julius Berner | Nvidia |
| Daniel Freeman | Anthropic |
| Aditya Grover | UCLA |
| Dar Mehta | Casual Labs |
| Danielle Maddix Robinson | Siemens |
| Molei Tao | Georgia Tech |
| Ferdinando (Nando) Fioretto | University of Virginia |
| Hayden Helm | Helivan |
| Alex Morehead | Berkeley Lab |
More lecturers to be announced.
These are my own notes and may contain errors. Not official course material. Source on GitHub.