Matthias Rupp

Matthias Rupp is a researcher who builds machine-learning models of molecules and materials, and he heads the PROMAR group at the Luxembourg Institute of Science and Technology. After a PhD in bioinformatics at the University of Frankfurt in 2009, he coauthored a 2011 model that predicted the atomization energies of organic molecules from nuclear charges and atomic positions, with each molecule encoded as a Coulomb matrix. From the University of Basel, he helped publish the QM9 data set of 134k small organic molecules in 2014, and in 2015 he coauthored the Δ-machine learning approach. His later work includes the many-body tensor representation in 2017, a Machine Learning for Materials group at the Fritz Haber Institute in Berlin, a post as applied scientist at Citrine Informatics, and ultra-fast interpretable machine-learning potentials in 2023. He is also an ELLIS Fellow.

Now at: Luxembourg Institute of Science and Technology
Born: 1975
Badges: Bridge

Years link to their sources.

Timeline

2009

A PhD in bioinformatics

After studying computer science, he earns a PhD in bioinformatics at the University of Frankfurt in 2009. He then works as a postdoctoral researcher on cheminformatics at the Helmholtz centre in Munich, on machine learning at the Technical University of Berlin, on computer-assisted drug design at ETH Zurich and on physical chemistry at the University of Basel.
2011

Machine-learned molecular energies

In September 2011, he, Alexandre Tkatchenko, Klaus-Robert Müller and O. Anatole von Lilienfeld introduce a machine-learning model that predicts the atomization energies of organic molecules from nuclear charges and atomic positions only, with molecules encoded as Coulomb matrices. In December 2011, with John C. Snyder, Katja Hansen, Müller and Kieron Burke, he uses machine learning to approximate density functionals.
2014–2015

QM9 and Δ-machine learning

From the University of Basel's Institute of Physical Chemistry, he coauthors the QM9 data set, published in Scientific Data in August 2014 with computed properties for 134k stable small organic molecules. In March 2015, he, Raghunathan Ramakrishnan, Pavlo O. Dral and von Lilienfeld introduce Δ-machine learning, which adds machine-learning corrections to inexpensive approximate quantum-chemistry methods.
2017

The many-body tensor representation

In April 2017, he and Haoyan Huo introduce the many-body tensor representation, one description for both molecules and crystals that stays invariant to translations, rotations and permutations of same-element nuclei. The work appears in Machine Learning: Science and Technology in 2022.
By 2020

Machine Learning for Materials in Berlin

At the Fritz Haber Institute of the Max Planck Society in Berlin, he leads the newly established Machine Learning for Materials group in the Theory Department, which Matthias Scheffler directs. With Christopher Sutton, Mario Boley, Luca M. Ghiringhelli, Jilles Vreeken and Scheffler, he coauthors a 2020 Nature Communications paper from the institute's NOMAD Laboratory on a subgroup-discovery method that detects domains of applicability of materials models.
2020

Citrine Informatics and Konstanz

In 2020, he is an applied scientist at Citrine Informatics in Redwood City, California, a start-up working on active learning for industrial materials discovery and optimisation. He then joins the Department of Computer and Information Science at the University of Konstanz as a faculty member.
2023

Ultra-fast machine-learning potentials

In September 2023, he, Stephen R. Xie and Richard G. Hennig publish ultra-fast interpretable machine-learning potentials in npj Computational Materials, built from two- and three-body potentials in a cubic B-spline basis fitted by regularized linear regression. These potentials run two to four orders of magnitude faster than state-of-the-art machine-learning potentials. In 2023, he holds affiliations at the University of Konstanz and at the Materials Research and Technology Department of the Luxembourg Institute of Science and Technology.
As of October 2026

Heading PROMAR at LIST

He heads the Process Modelling, Automation and Robotisation (PROMAR) group in the Scientific Instrumentation and Process Technology unit of the Luxembourg Institute of Science and Technology, which develops machine-learning interatomic potentials and Bayesian optimisation for the design of chemicals and materials. The group's SIMINDENT project, funded through FNR BRIDGES 2025, uses ultra-fast machine-learning potentials to predict nanoindentation in cemented carbides with industry partner CERATIZIT. He is an ELLIS Fellow in the Quantum and Physics Based Machine Learning program.