On Friday June 20, 2025, Loïc Dumortier successfully defended his PhD thesis ‘Model-Informed Training Data Curation for Reactive All-Atom Potentials’. During his PhD research he was supervised by prof. Toon Verstraelen (CMM, Ghent University), dr. Jelle Vekeman (CMM, Ghent University), dr. Theodorus De Bruin (IFP Energies Nouvelles) and dr. Benoît Creton (IFP Energies Nouvelles).
Congratulations Loïc !
Summary of the PhD in laymen’s terms
This PhD thesis focuses on the development of predictive models, known as reactive potentials, used extensively in Molecular Dynamics. These potentials aim to simulate chemical reactions accurately and efficiently at the atomic scale, making it possible to study complex systems and processes over length and time scales that are computationally intractable for higher-accuracy ab initio methods. This task presents challenges, particularly concerning the high cost and scarcity of the high-quality reference data needed to build these models. The work concentrates on pioneering “model-informed training data curation” strategies to enhance the development of two major types of reactive potentials: the empirical Reactive Force Fields (ReaxFF) and the data-driven Machine Learning Interatomic Potentials (MLIPs). Two primary methodological advancements within this curation framework are presented in this thesis, demonstrating techniques tailored for ReaxFF and MLIPs, respectively, but holding potential for broader application:
- the “Balanced Loss” cost function, a novel approach designed to streamline the complex task of ReaxFF parameter optimization. (https://pubs.acs.org/doi/full/10.1021/acs.jctc.3c01009)
- an efficient active learning framework aimed at automating the creation of training datasets for dedicated MLIPs.
In essence, this research provides innovative, systematic, and data-efficient methodologies applicable within and beyond the immediate context of reactive potential development. By offering the Balanced Loss function and a node energy-based active learning scheme, it equips the field with valuable tools to advance the field of reactive Molecular Dynamics simulations.