Julian was working on "Enhanced Sampling Applied to Data Generation for Machine-Learned Interatomic Potentials" supervised by Fabian Zills. We congratulate him on this wonderful achievement and invite you to attend the defense.
Abstract: The accuracy and robustness of machine-learned interatomic potentials (MLIPs) depends on the quality and configurational diversity of the provided training data. Traditional molecular dynamics (MD) sampling often fails to explore the configurational space efficiently, resulting in unseen configurations and more need for data extrapolation.In this thesis I investigate the use of enhanced sampling techniques to improve data generation for MLIPs.Using alanine dipeptide as a benchmark system, we compare umbrella sampling and metadynamics in constructing free energy surfaces based on backbone torsion angles. Additionally, we introduce a force decomposition approach that allows direct manipulation of translational, rotational and vibrational force components, effectively controlling intermolecular interactions and inducing pseudo-diffusion in bulk water simulations. My results demonstrate that models trained on biased datasets, particularly those using enhanced sampling, outperform those trained on unbiased MD trajectories.
Best of luck Julian on future endeavors!
