Transferable Implicit Transfer Operators
Title: |
Transferable Implicit Transfer Operators |
DNr: |
Berzelius-2025-189 |
Project Type: |
LiU Berzelius |
Principal Investigator: |
Simon Olsson <simonols@chalmers.se> |
Affiliation: |
Chalmers tekniska högskola |
Duration: |
2025-05-31 – 2025-12-01 |
Classification: |
10210 |
Keywords: |
|
Abstract
Computing properties of molecular systems rely on estimating expectations of the
(unnormalized) Boltzmann distribution. Molecular dynamics (MD) is a broadly adopted
technique to approximate such quantities. However, stable simulations rely on tiny
integration time steps (10^(−15) s), whereas convergence of some moments, e.g., binding free energy or rates, might rely on sampling processes on time scales as long as 10^(−1) s, and these simulations must be repeated for every molecular system independently. We recently proposed Implicit Transfer Operator (ITO) Learning [1], a framework to learn surrogates of the simulation process with multiple time resolutions. Our initial work shows that ITO models can simulate challenging molecular systems, such as fast-folding proteins, with time steps at least six orders of magnitude larger than traditional molecular dynamics [1]. We implement ITO models using equivariant flow-matching models with bespoke SE(3) equivariant architectures. While the ITO architecture, in principle, allows for training general models that work for all molecules, current datasets have been missing. This deficiency limits current ITO models to be molecule specific and their application potential in drug and material design. We aim to train the first-ever transferable ITO models. Such a model would open tremendous potential for applications and lower computational needs to study the properties of molecules. Our group has experience with ITO models and transferable generative models for molecules. Simon Olsson (PI) initially proposed and implemented the ITO framework. Simon Olsson and Juan Viguera Diez (Ph.D. student) developed the second generation of ITO models leveraging sample efficiency gains from Boltzmann Generators [2]. Simon Olsson and Juan Viguera Diez have previous experience with transferable generative models for molecular structures [3,4]. In previous projects, we have developed prototypes, generated promising preliminary results, and optimized our code base for efficient and scalable training and inference. We are applying for this project to perform final production runs.
[1] Mathias Schreiner, Ole Winther and Simon Olsson. Implicit Transfer Operator Learning: Multiple Time-Resolution Surrogates for Molecular Dynamics. 37th Conference on Neural Information Processing Systems (NeurIPS 2023). https://arxiv.org/abs/2305.18046
[2] Juan Viguera Diez, Mathias Schreiner, Ola Engkvist, and Simon Olsson. Boltzmann
Priors for Implicit Transfer Operators, 2025. International Conference of Learning
Representations (ICLR) 2025. https://arxiv.org/abs/2410.10605
[3] Juan Viguera Diez, Sara Romeo Atance, Ola Engkvist and Simon Olsson. Generation of conformational ensembles of small molecules via surrogate model-assisted molecular
dynamics. 2024 Mach. Learn.: Sci. Technol. 5 025010.
[4] Juan Viguera Diez, Sara Romeo Atance, Ola Engkvist, Rocío Mercado, Simon Olsson. A transferable Boltzmann generator for small-molecules conformers.