ICLR 2025 Workshop on Machine Learning Multiscale Processes

MLMP Logo depicting and art deco style image of an atom merging with a microchipGiven low-level theory and computationally-expensive simulation code, how can we model complex systems on a useful time scale?
Date: 27 April 2025
Location: ICLR 2025 conference, Singapore EXPO, Room Conference GHJ
Registration: iclr.cc; Conference Sessions and Workshops or Sunday Workshop 1 Day Pass
Accepted papers: OpenReview, iclr.cc, Constructor Model
Follow us:

Next: AI4X 2025 8 to 11 July 2025 in Singapore

The next big AI for science event in Singapore is the AI4X 2025 conference. Unlike the laser-focused workshop, it will cover all areas of AI for science (and finance). Discuss the future of the area with leading researchers from top universities and companies! Early bird registration until 30 April 2025.

Schedule

TimeEvent
08:15–08:45Poster Setup
08:45–09:00Opening Remarks
09:00–09:30Keynote: Kostya Novoselov
09:30–10:00Keynote: Sergei Gukov
10:00–10:30Coffee Break
10:30–10:45On Incorporating Scale into Graph Networks
10:45–11:15Keynote: Qianxiao Li
11:15–11:45Keynote: Charlotte Bunne
11:45–12:15Q&A and Panel Discussion
12:15–13:15Lunch
TimeEvent
13:15–13:30DoMiNO: Down‑scaling Molecular Dynamics with Neural Graph Ordinary Differential Equations
13:30–14:00Poster Session: Spotlight Talks
14:00–15:00Poster Session: Poster Session
15:00–15:30Coffee Break
15:30–15:45LOGLO‑FNO: Efficient Learning of Local and Global Features in Fourier Neural Operators
15:45–16:005D Neural Surrogates for Nonlinear Gyrokinetic Simulations of Plasma Turbulence
16:00–16:30Keynote: Daniel Polani
16:30–16:50Panel Discussion
16:50–17:40Workgroups discussions
17:40–17:50Collaboration pitches
17:50–18:00Closing Remarks

Keynote Speakers

Fundamental laws of Nature, Standard Model of Physics, and the most applied part of it, quantum mechanics, are well established. Theoretically, the dynamics of anything starting from a hydrogen atom and all the way to Earth's climate follow those equations. The problem is complexity [Dirac 1929]. An exact computation of a modest system containing 100 atoms is still beyond the capability of modern computers.

Some of the greatest scientific achievements resulted from breakthroughs in scale transitions: renormalization, density functional theory, Higgs boson, multiscale models for complex chemical systems, climate modeling, protein folding. Those achievements are highly regarded because they are impactful – but also unique and can't be readily applied to different systems.

Encouraged by the recent successes, this workshop aims to enable the development of universal AI methods that would be able to find efficient and accurate approximations, and use them for some of the most pressing and high-impact scientific problems that have computational complexity as the limiting factor to an in silico solution, such as:

  • High-temperature superconductivity
  • Fusion power
  • Weather prediction
  • Living organism digital twins
  • Catalysts

If we solve scale transition, we solve science.