(10/n) Supported by @bifold.berlin, @zuseschooleliza.bsky.social.

(9/n) Check out our paper and code:
🌐 Website & Colab: ml4molsim.github.io/hamiltonian-...
📄 Paper: arxiv.org/abs/2601.22123
💻 Code: github.com/ML4MolSim/ha...

(8/n) Done with a brilliant team: @plainer.bsky.social, @gregorlied.bsky.social, @thorbenfrank.bsky.social, Oliver Unke, Stefan Chmiela, @franknoe.bsky.social, and Klaus-Robert Müller.

(7/n) With this, you get a single model that provides:
- Instantaneous forces (like a standard MLFF)
- Stable large-timestep updates far beyond classical integrators
- Training and inference cost comparable to MLFFs

(6/n) How?

Inspired by recent advances in few-step generative modeling, our tailored loss function combines force matching with a consistency constraint that enforces agreement of the predicted flow across different time horizons.

(5/n) Our approach:

In contrast, our method learns continuous-time, large-timestep dynamics directly from decorrelated phase-space samples without requiring expensive reference trajectories, and supports arbitrary timesteps during inference.

(4/n) Existing approaches rely on trajectory data, typically generated using another model simulated with small timesteps. This potentially introduces artifacts from the teacher and is computationally expensive.

Can we learn large timesteps, without ever seeing trajectories?

(3/n) The solution: Hamiltonian Flow Maps

The core idea is to model the Hamiltonian evolution directly in phase space over a finite time interval. Concretely, we aim to learn a 𝗛𝗮𝗺𝗶𝗹𝘁𝗼𝗻𝗶𝗮𝗻 𝗙𝗹𝗼𝘄 𝗠𝗮𝗽 that advances positions and momenta over an interval Δt.

(2/n) The problem: Simulations are fundamentally limited by the small timesteps required for stable numerical integration. Even with ML, most of the cost still comes from taking millions of tiny integration steps.

Ever get tired of tiny timesteps bottlenecking your MD simulations?

We show how to train a model for large-timestep Hamiltonian dynamics directly on standard MLFF datasets. 𝗡𝗼 𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝘁𝗿𝗮𝗷𝗲𝗰𝘁𝗼𝗿𝗶𝗲𝘀, 𝗻𝗼 𝘂𝗻𝗿𝗼𝗹𝗹𝗶𝗻𝗴, 𝗻𝗼 𝘁𝗲𝗮𝗰𝗵𝗲𝗿 needed!

🧵👇

Link to github: github.com/ML4MolSim/di...
Paper on arxiv: arxiv.org/abs/2506.15378

✨ Via a modular architecture, we enable a fair comparison of symmetries changing both attention mechanism and embedding strategies of our model
✨ We transfer the powerful DiT architecture from Computer Vision to the molecular domain, proposing two complementary graph-based conditioning strategies

We introduce DiTMC, a new way to predict molecular conformers - the different 3D shapes molecules can flex into.
✨ We learn to predict 3D geometry from molecular structure
✨ We achieve state-of-the-art results on the GEOM benchmarks

I’m excited to be at NeurIPS 2025 next week and present our latest paper on molecular conformer generation! Huge thanks to my co-authors @thorbenfrank.bsky.social , Gregor Lied, Klaus-Robert Müller, Oliver Unke and Stefan Chmiela for an incredible collaboration. Supported by: @bifold.berlin