GENERAL-PURPOSE MACHINE LEARNING FORCE FIELDS FOR (BIO)MOLECULAR SIMULATIONS - Dissertation Defense - Abstract

Sep 17, 2026

Uploaded by Finn

Accurate and efficient simulation of large (bio)molecular systems with quantum-mechanical fidelity remains a central computational-science challenge. Ab initio quantum chemistry is accurate but too expensive, while classical force fields are efficient but less accurate. This thesis introduces general-purpose machine learning force fields that preserve quantum fidelity and scale to large systems by combining quantum data, efficient atomic representations, and explicit long-range interaction models, including new datasets and descriptor-based global learning methods.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages120

Reading Time5 h 2 min

AccessFree to download

Abstract

Datasets: MD22 and QCell

Long-range global learning via rGDML

Pretrained global model: SO3LR

Validation and implications

PhD-FSTM-2026-002 The Faculty of Science, Technology and Medicine DISSERTATION Defense held on 16/01/2026 in Luxembourg to obtain the degree of DOCTEUR DE L’UNIVERSIT´E DU LUXEMBOURG EN PHYSIQUE by Adil KABYLDA Born on 15th February 1997 in Pavlodar (Kazakhstan) GENERAL-PURPOSE MACHINE LEARNING FORCE FIELDS FOR (BIO)MOLECULAR SIMULATIONS Dissertation defense committee Dr. Alexandre Tkatchenko, Supervisor Professor, University of Luxembourg Dr. Etienne Fodor, Chairman Professor, University of Luxembourg Dr. Massimiliano Esposito Professor, University of Luxembourg Dr. Johannes T. Margraf Professor, University of Bayreuth Dr. Siewert J. Marrink Professor, University of Groningen Abstract The accurate and eﬃcient simulation of large (bio)molecular systems with quantum mechanical ﬁdelity represents a grand challenge in computational science. Ab initio quantum chemistry methods provide accuracy but remain prohibitively expensive at realistic scales, whereas classical force ﬁelds achieve eﬃciency but sacriﬁce accuracy. Machine learning force ﬁelds (MLFFs) promise to close this gap, yet their predictive power is often limited by locality assumptions that miss the long-range eﬀects governing the structure, dynamics, and function of complex (bio)molecular systems. This thesis develops a framework for general-purpose MLFFs that preserves quantum mechanical ﬁdelity while scaling to large systems by combining quantum-mechanical data, eﬃcient atomic representations, and models explicitly designed to capture long-range interactions. To advance model development beyond small molecules, we introduce two quantum mechanical datasets that span the chemical space of cellular components: MD22 and QCell. MD22 oﬀers a benchmark featuring molecular dynamics trajectories for six biomolecular units and two supramolecular complexes. It represents a signiﬁcant increase in system size (up to 370 atoms) and conformational ﬂexibility, and is speciﬁcally designed to probe nonlocal correlations. To support the training of broadly applicable, general-purpose models, QCell takes this a step further by signiﬁcantly expanding coverage across all major classes of biomolecules, with ∼500k diverse fragments of carbohydrates, nucleic acids, lipids, as well as noncovalent dimers and ion-water clusters. We then make collective eﬀects tractable in global MLFFs that couple all atomic degrees of freedom by developing an eﬃcient interatomic descriptor. The resulting algorithm, reduced descriptor gradient-domain machine learning (rGDML), automatically constructs the minimal set of interatomic features required to capture long-range ﬂuctuations, converting the quadratic growth of global descriptors into linear scaling. rGDML improves accuracy over both local and baseline global models, and its eﬃciency and stability are demonstrated through a 50 ns molecular dynamics simulation of a tetrapeptide. Its enhanced interpretability enables systematic analysis across MD22 molecules, revealing that nonlocal features (atoms separated by up to 15 ˚A in the studied systems) are essential to retain overall accuracy for peptides, DNA base pairs, fatty acids, and supramolecular complexes. Building on these insights, we introduce SO3LR, a pretrained general-purpose MLFF that couples a fast SO(3)-equivariant neural network for semi-local interactions with universal, physically grounded pairwise potentials for short-range repulsion, long-range electrostatics, and dispersion. SO3LR is trained on a diverse set of four million neutral and charged molecular complexes computed at the PBE0+MBD level of quantum mechanics, ensuring broad coverage of covalent and noncovalent interactions. The model scales to 200k atoms on a single GPU and achieves reasonable to high accuracy across the chemical space of organic (bio)molecules. We validate this performance with polyalanine simulations from 300 to 800 K, accurate structural and spectroscopic observables

FAQs

Why are machine learning force fields needed for (bio)molecular simulations?

Quantum-accurate ab initio methods are prohibitively expensive at realistic scales, and classical force fields sacrifice accuracy. ML force fields aim to close this gap by improving predictive power while maintaining quantum mechanical fidelity.

What are MD22 and QCell used for in this thesis?

MD22 provides benchmark molecular dynamics trajectories for biomolecular units and supramolecular complexes to probe nonlocal correlations. QCell expands chemical coverage across major biomolecule classes using a large set of diverse fragments and related complexes.

How does rGDML address long-range effects in global MLFFs?

rGDML develops an efficient interatomic descriptor that captures long-range fluctuations while converting the quadratic growth of global descriptors into linear scaling, enabling improved accuracy and efficiency in long simulations.

Sep 17, 2026

Uploaded by Finn

GENERAL-PURPOSE MACHINE LEARNING FORCE FIELDS FOR (BIO)MOLECULAR SIMULATIONS - Dissertation Defense - Abstract_Page 1

Open in WPS to view all 120 pages