marcel.science / home
Hello! 👋 I’m Marcel, a postdoc in the COSMO Lab at EPFL in Lausanne, Switzerland. Previously, I was a doctoral student at the Fritz Haber Institute (in the NOMAD Laboratory) and TU Berlin (in the Machine Learning group).
Here’s the brief, jargon-y pitch for my research:
Computational quantum-mechanical modelling methods such as density functional theory can predict the properties of materials from first principles, but are limited by their high computational cost. In settings involving many simulations, machine learning can reduce these costs, often by orders of magnitude, by interpolating between reference calculations. In my research, I study how to design such fast surrogate models.
In practical terms, this means that I’m trying to understand how various machine-learning models can be used to approximate the relationship between the structure (i.e. the position of atoms and the unit cell) of a material and certain properties, in particular the energy and forces. This requires getting into the nitty-gritty of how these methods work, how physical requirements can be incorporated into them, and doing lots of parameter searches.
Here are some other places on the internet where you can find me:
- Twitter for banter and updates,
- notes.marcel.science for some technical notes,
- Gitlab and Github for code,
- Google Scholar or ORCID for official scientific activity,
- marcel.computer for my personal (online) life!
If you have any questions, or just want to say hi, you should be able to reach me on twitter, or at mail@marcel.science
.
Thanks for stopping by! 🚀
Projects
These are the projects I’m working on (or have worked on):
repbench
(2018-2022): Review and benchmark of different representations of molecules and materials, evaluating them as features for regression,cmlkit
(2018-2020): Minimalist, modular framework to specify, optimise, and evaluate machine learning models for condensed matter physics, developed largely for therepbench
project,gknet
(2020-2023): Adaptation of the Green-Kubo method for computing thermal conductivities to message-passing graph neural networks, and application to zirconium dioxide, a strongly anharmonic material,glp
(2023-now):jax
-based implementation of stress and heat flux for graph-based machine-learning potentials,gkx
(2023-now): Green-Kubo molecular dynamics injax
,doctor
(2017-2023): My doctoral dissertation,eqt
(2024): Testing to what extent non-invariant models can reasonably predict physical observables, i.e., checking the practical impact of breaking invariances.
I've also written some tutorials for the NOMAD Analytics Toolkit:
cmlkit
: Toolkit for Machine Learning in Computational Condensed Matter Physics and Quantum Chemistry, which covers the basics of thecmlkit
framework, in a hopefully beginner-friendly way.krr4mat:
Kernel Ridge Regression for Materials Property Prediction: A Tutorial Introduction, which gives a short, and very pragmatic, introduction to kernel ridge regression.
I also started the Fritz Sessions, an intermittent series of lectures at the Fritz Haber Institute, covering topics related to the future.
Publications
- Title: Representations of molecules and materials for interpolation of quantum-mechanical simulations via machine learning
- Published: npj Comput. Mater. 8, 41 (2022)
- DOI: 10.1038/s41524-022-00721-x
- Preprint: arxiv.org/abs/2003.12081 (2020)
- Title: Heat flux for semilocal machine-learning potentials
- Published: Phys. Rev. B 108, L100302 (2023)
- DOI: 10.1103/physrevb.108.l100302
- Preprint: arxiv:2303.14434 (2023)
- Title: Stress and heat flux via automatic differentiation
- Published: J. Chem. Phys. 159, 174105 (2023)
- DOI: 10.1063/5.0155760
- Preprint: arxiv:2305.01401 (2023)
- Title: Machine Learning for Atomistic Modeling: Representations and Thermal Transport (Dissertation)
- DOI: 10.14279/depositonce-18647 (2023)
- Title: Probing the effects of broken symmetries in machine learning
- Published: Mach. Learn.: Sci. Technol. 5 04LT01 (2024)
- DOI: 10.1088/2632-2153/ad86a0
- Preprint: arxiv:2406.17747 (2024)