marcel.science / home

My current research focuses on the design, implementation, and application of machine-learning interatomic potentials, in particular for materials science. I've worked on a few different aspects of this overall topic, starting with representations of molecules and materials (repbench), through thermal transport with message-passing neural networks (gknet, glp), to the investigation of the role of physical priors like equivariance (eqt), and derivative-based forces (nc). Recently, I've been focusing on including long-range effects in these models (lorem). I'm also currently dabbling in closing the loop to predict experimental observables... ⚗️

Most of what I do is build computer programs that simulate the movement of atoms: The kind of simulations chemists, physicists, and materials scientists use to understand and design everything from batteries to thermal insulators to new drugs. The difficulty here is this: You can either solve the underlying quantum mechanics directly, which is very accurate but also very expensive (even tiny systems for short times need a supercomputer), or you can use a hand-crafted approximation (a "force field") that runs fast but encodes a lot of guesses about how atoms will behave. Neither is great if you actually want to predict, say, how heat flows through a real material at a real temperature, or if you want to look at materials that have never been studied before. I work on using machine learning to bridge these extremes: Train a model on a relatively small set of expensive quantum calculations, and then let it stand in for them. This is much cheaper, almost as accurate, and (crucially) flexible enough to keep getting better as you feed it more data. My concrete everyday work involves mostly prototyping new models, arguing with people about derivatives, and trying to write software that is somewhere in the sweet spot between "well-engineered/efficient enough to work" and "flexible/hackable enough for research".

Read this with jargon / without it.