Modeling larger length and time scales in machine learning force fields

Molecular dynamics simulations describe the collective motion of atoms over time, enabling the study of fundamental processes across physics, chemistry, and biology. This has established molecular dynamics as a cornerstone of modern science and improving its accuracy has a long standing history in the computational sciences. The accuracy of a molecular dynamics simulation is determined by the accuracy of the interatomic forces. These forces can be obtained from fast but approximate empirical force fields, or slow but accurate quantum mechanical methods. Machine learning force fields have emerged as a promising alternative, bridging this gap by approximating ("learning") the interatomic forces from quantum mechanical reference data. Despite great advances over the past years, many challenges remain on the path towards generally applicable machine learning force fields. A key limitation is the accurate treatment of long time scales and large length scales. However, both are crucial for the correct prediction of experimental outcomes, which is the ultimate measure of predictive usefulness. In this thesis, we address some challenges that are related to the modeling of long time scales and large length scales within machine learning force fields. We propose a deep neural network architecture, that reduces the computational cost while maintaining accuracy and reliability. This is accomplished by proposing a new design space for geometric neural network architectures, extending the time scales that are accessible in machine learning force field simulations. As a next step, we combine this model with physically inspired long-range corrections, enabling a meaningful description of large length scales. This approach allows simulating time and length scales relevant for biomolecular simulations, further demonstrating agreement with experimental data. Recognizing that current approaches often resort to physically inspired post-hoc correction for long-range interactions, we also explore a complementary direction: Learning all interactions without physically derived correction terms but in a purely data-driven way. We propose a computationally efficient algorithm capable of learning short- and long-range interactions with equal accuracy. This paves the way for machine learning models, that can simultaneously learn across all relevant length scales in atomistic systems.

Author dc:creator

• Frank, Thorben

Advisor dc:contributor.advisor

• Müller, Klaus-Robert

Licence dc:rights.uri

• https://creativecommons.org/licenses/by-sa/4.0/

Language dc:language.iso

en

Identifier URI

https://doi.org/10.14279/depositonce-24446

OAI identifier oai:identifier

oai:depositonce.tu-berlin.de:11303/25622