Recent Advances in Machine Learning Accelerated Molecular Dynamics
Computer simulation with molecular dynamics (MD) acts as a bridge between microscopic models and macroscopic phenomena. Machine learning (ML) - an emerging data-driven approach in this context - can provide new impetus and accelerate MD simulations to tackle new challenges in both method developments and applications , .
Traditionally, force fields behind MD simulations in biophysics and materials chemistry applications were constructed using simple functional forms, which do not always properly describe complex chemical environments such as, for example, metal sites in enzymes or (electro)chemical interfaces. In this context, ML based reactive force fields (or potentials) are now emerging as a promising alternative approach, with their ability to give quantum mechanical accuracy without explicitly including the electronic degrees of freedom. This will allow simulation of a range of relevant systems using time and length scales that are unfeasible today , , particularly when combined with enhanced sampling methods such as metadynamics and Markov state models.
The success of ML relies on the availability of data and knowledge of how to design a dataset which is representative yet small enough. Thus, the synergy between sampling methods, such as the Markov state model, metadynamics and ML techniques, such as active learning, may stimulate new ideas , . ML methods that can also provide property predictions in both compositional and conformational spaces are urgently needed , . In turn, such methods should also allow a posteriori data analytics for exploring correlations and discovering new descriptors in real-world applications.
All in all, it is now particularly timely to bring scientists from MD simulation and ML backgrounds within the broad and simulation-intense areas of biophysics and material science under the same roof to spark discussions and new ideas. In particular, leading questions that we would like to address in this workshop are as follows:
- How to include prior knowledge such as symmetries, specific boundary conditions, quantitative observations and covariant requirements, and long-range interactions into atomistic ML models?
- How to develop the synergy between advanced sampling methods and ML methods?
- How to combine knowledge from massively large, heterogeneous and/or experimental/MD simulation datasets with ML in order to identify novel targets in material (e.g. for a new material design in all-solid-state batteries) and drug design (e.g. in a responsive mode as was shown to be necessary for the Covid-19 pandemic)?
This workshop will be supported by both the Juelich node and Pisa node. We will also seek support from EU-supported consortia, specifically the Bioexcel CoE (www.bioexcel.eu) and the Dynion Forschergruppe.
Paolo Carloni ( Juelich Research Center ) - Organiser
GiovanniMaria Piccini ( Humboldt Universität zu Berlin - Max Planck Gesellschaft ) - Organiser
Andrew Plested ( Leibniz-Institut für Molekulare Pharmakologie ) - Organiser
Giuseppe Brancato ( Scuola Normale Superiore, Pisa ) - Organiser
Kersti Hermansson ( Uppsala University ) - Organiser
Chao Zhang ( Uppsala University ) - Organiser