Dmitrii Kochkov

Discovering synergies between ML, computing and physics.

[Stephan Hoyer @shoyer]

NeuralGCM was a huge team effort, and I would especially like to acknowledge the core team who has worked on the project over the past three years: @dkochkov1, @janniyuval, @langmore, @pnorgaard314, @singledispatch

[Stephan Hoyer @shoyer]

New open source release from my team at Google: Dinosaur, a differentiable dynamical core for global atmospheric modeling, written in JAX: https://t.co/5WASoZ1GLM

Dinosaur is a core component of NeuralGCM and we hope it is useful for the weather/climate research community.

[Kim Stachenfeld (neurokim.bsky.social) @neuro_kim]

We have a paper in ICLR! The title is “Learned Coarse Models for Efficient Turbulence Simulation.” We wanted to see if we could train general-purpose ML models to predict turbulent dynamics accurately at low spatial and temporal resolutions (1/n).

[Dmitrii Kochkov @dkochkov1]

2/2
Key learnings:

1) Graph-nets learn amplitudes and phases that generalize to larger lattice sizes
2) Learning phases and amplitudes with separate networks gives better results
3) GNNs have easier time restoring symmetries than learning in a fully symmetric subspace

[Dmitrii Kochkov @dkochkov1]

1/2
Our work "Learning ground states of quantum Hamiltonians with graph networks" is out!
https://t.co/XSIj574c8Z

TLDR: we do Variational Monte Carlo on Heisenberg Hamiltonians using graph-nets as variational ansatz.

w/
@spectralhippo
Alvaro Sanchez
@PeterWBattaglia
@bkclark4

[Stephan Hoyer @shoyer]

Can machine learning improve physics constrained optimization tasks, like those at the heart of numerical weather forecasting?
Happy to share some our work on "Variational Data Assimilation with a Learned Inverse Observation Operator", tomorrow/today at ICML.
🧵

[Shirley Ho @cosmo_shirley]

Come here our selected contributed talk on simulating turbulence with deep learning at #SimDL (https://t.co/0PnK0SbAat) today

Led by @neuro_kim, Alvaro Sanchez, @DrumBushField, Dmitrii Kochkov, @MilesCranmer, @spectralhippo, Jonathan Goodwin, Elaine Cui and @PeterWBattaglia

[Dmitrii Kochkov @dkochkov1]

@MilesCranmer I think once the particle symmetry is introduced we are forced to make "sets" out of elements of the corresponding Fock space. With that axiom of choice still feel natural.

[Google DeepMind @GoogleDeepMind]

The paper “Learning Mesh-Based Simulation with Graph Networks” shows how to effectively learn physics simulation on irregular meshes. The model can simulate vastly different systems with significant speedups over the ground truth solver: https://t.co/MxW36V4TtY #ICLR2021 (3/3)

[Dmitrii Kochkov @dkochkov1]

@ylecun @JonathanTompson That's great work, but focused on computer graphics rather than scientific accuracy.

[Dmitrii Kochkov @dkochkov1]

1/2
Excited to share "Machine learning accelerated computational fluid dynamics"
https://t.co/8rXhLGTVZC

We use ML inside a CFD simulator to advance the accuracy/speed Pareto frontier

with/
Jamie A. Smith
Ayya Alieva
Qing Wang
Michael P. Brenner
@shoyer

[Dmitrii Kochkov @dkochkov1]

@ngutten Great question! We don't have any results on this yet, but it's an important area to advance. One challenge with boundaries is that the flow is often very different near the walls and might require additional augmentations.

[Dmitrii Kochkov @dkochkov1]

@ColinJMcAuliffe That's a good perspective! In fact the goal of such effective "regularization" is to minimize errors caused by the loss of details on coarse grids.

One thing to note is that our approach is unlikely to model fully missing detail since we achieve high pointwise accuracy.

[Dmitrii Kochkov @dkochkov1]

@whiskeyandwry Great question! My take on this is that it depends heavily on the application. One could use such accelerated methods to perform design by optimization and then evaluate the performance using methods with stronger convergence guarantees.