Hacker News new | past | comments | ask | show | jobs | submit login
Physics-Based Deep Learning Book (physicsbaseddeeplearning.org)
333 points by sebg 9 months ago | hide | past | favorite | 24 comments



In this dense overview presentation (Oct 2022), Chris Rackauckas introduced Sci ML with diverse examples from many fields: epidemics, gravitational waves, pharmacometrics, ocean simulation... and some open source and proprietary Julia libraries for SciML. Highly informative!

https://www.youtube.com/watch?v=yHiyJQdWBY8


Does anyone know what the job market looks like for a "Physics-simulation ML engineer" (or whatever it's called)?


I'd strongly rephrase the title, this is NOT a book on physics-based deep learning.

This is a book on the deep learning approaches for physics problems DEVELOPED BY THIS RESEARCH GROUP. I think that is a very very important disclaimer to this book.

In addition, it is essentially used to strongly push their simulation framework Phi-Flow.

I would NOT call this an accurate depiction of the field.


Chris has done good work on this genre. His differential equations Julia package with support for physics or sci ML is pretty cool.

https://www.stochasticlifestyle.com/the-essential-tools-of-s...


Hopefully this is a great book, what a great topic to write a book about. Kudos to the author.


Maybe I'm blind, but how do I download the entire book as PDF? I only find the download button up top for individual pages?

Afaik, it's produced by Jupyter book[1], but find nothing in their docs either.

[1] https://jupyterbook.org/en/stable/intro.html



Direct link to the arXiv abstract page, where one can download the PDF: https://arxiv.org/abs/2109.05237


Some other good resources-

1. CRUNCH group YouTube (talks on Math + ML) - https://m.youtube.com/channel/UC2ZZB80udkRvWQ4N3a8DOKQ

2. Steve Brunton's Physics Informed Machine Learning playlist - https://m.youtube.com/playlist?list=PLMrJAkhIeNNQ0BaKuBKY43k...

3. The book "Data Driven Science and Engineering" from Steve Brunton

4. Deep Learning in Scientific Computation from ETH Zurich - https://m.youtube.com/playlist?list=PLJkYEExhe7rYY5HjpIJbgo-...



Sounds like a valuable resource for both beginners and experienced


I was wondering : does deep learning have the potential to make large-scale quantum physics simulations more tractable? How about plasma physics for fusion reactors?


It obviously has the potential.... work in weather and biology points to it.


TBC, this is about deep learning for physics problems, not a general approach to deep learning from a physicist's perspective.

> This document contains a practical and comprehensive introduction of everything related to deep learning in the context of physical simulations. As much as possible, all topics come with hands-on code examples in the form of Jupyter notebooks to quickly get started. Beyond standard supervised learning from data, we’ll look at physical loss constraints, more tightly coupled learning algorithms with differentiable simulations, training algorithms tailored to physics problems, as well as reinforcement learning and uncertainty modeling. We live in exciting times: these methods have a huge potential to fundamentally change what computer simulations can achieve.


That would have been the more interesting book. There is a lot of intuition that statistical mechanics could bring to deep learning.

I would have called this one Deep Learning for Physics.


I guess considering their group is called the Physics-based Simulation Group [1], I'm thinking maybe that's just the terminology they've always used? Or maybe it's a German->English translation thing?

[1] https://ge.in.tum.de/


I think the other one is more commonly known as “physics informed deep learning”.


I don't think so? Wikipedia suggests that "Physics-informed neural networks" is

> a type of universal function approximators that can embed the knowledge of any physical laws that govern a given data-set in the learning process

https://en.wikipedia.org/wiki/Physics-informed_neural_networ...

In other words, that seems to refer to giving the model prior info (a bias) about physical laws that generated the data. What I'm talking about is more abstract: using physics-y type math ideas to understand the internal behavior of the networks. Here are a couple examples:

https://proceedings.neurips.cc/paper_files/paper/2023/hash/6...

https://cgad.ski/blog/where-is-noethers-principle-in-machine...


Yeah, I came here expecting methods of mathematical physics applied to deep learning. I'd suggest a title change.

Still a great topic though, no doubt!


The title is misleading, no? It seems to be about how to apply deep learning to physics simulations. It is not about borrowing physics concepts and applying them to the NN landscape.

That said, it is a lovely set of topics.


It is misleading. The is not DL based on physics. It is physics based on DL.


> The title is misleading, no?

No. I got the correct meaning at first glance.


How many IBM Technical Support workers does it take to change a lightbulb? None, we have an identical model here and ours is working fine.


The most important question - how to apply these methods to contact dynamics?




Join us for AI Startup School this June 16-17 in San Francisco!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: