Machine-learning assisted modeling of single-atom catalysts for CO₂ hydrogenation to methanol

Context

The catalytic conversion of CO₂ into methanol is a central route for turning an abundant carbon source into a liquid energy carrier and chemical intermediate. At the atomic scale, this reaction depends on a delicate balance between CO₂ activation, hydrogen transfer, stabilization of C1 oxygenated intermediates and desorption of methanol or competing products. Small changes in the local geometry or electronic structure of the active site can therefore redirect the reaction pathway, which makes surface-level understanding essential for rational catalyst design.

Single-atom catalysts supported by graphene or N-doped graphene offer a controlled platform for this problem. An isolated transition-metal atom coordinated in a carbon/nitrogen environment can provide a well-defined active site while maximizing metal efficiency. However, the relevant chemical space remains large: the metal identity, coordination motif, nearby defects, hydrogen coverage and reaction intermediates all modify activity and selectivity. Density functional theory can describe these effects accurately, but systematic exploration of many sites and elementary steps rapidly becomes computationally expensive. Conversely, general-purpose machine-learning potentials are not yet sufficiently reliable for rare reactive configurations at surfaces. The central idea of this PhD project is to build a focused, data-efficient modeling strategy for this specific class of reactive interfaces.

PhD project

The thesis will develop a computational workflow to study CO₂ hydrogenation on graphene-supported single-atom catalysts, with emphasis on transition-metal sites coordinated by carbon and nitrogen motifs. The first step will be to establish a robust DFT reference for representative adsorption states and elementary reaction steps, including CO₂ bending/activation, hydrogen addition, formate or carboxyl-like intermediates, methoxy formation and methanol release. This reference will be used to identify the structural motifs and reaction coordinates that control activity and selectivity.

The methodological core of the work will be the construction of an informative training set rather than the accumulation of a large, redundant database. Candidate configurations will be generated from DFT relaxations, reaction-path searches, molecular dynamics snapshots and targeted distortions around key intermediates. They will then be filtered using structural diversity, physical consistency and uncertainty indicators, so that expensive reference calculations are concentrated on configurations that genuinely improve the model. Based on this curated data, the PhD student will train and test machine-learning force fields or Δ-learning corrections able to reproduce energies, forces and selected reaction barriers in the relevant domain. The aim is to obtain models that are useful for catalyst screening because their limits are explicitly diagnosed, not simply because they perform well on equilibrium structures.

Scientifically, the project will address three linked questions: which metal–coordination environments activate CO₂ without trapping intermediates too strongly; which elementary steps are most sensitive to the local structure of the single-atom site; and how far active data selection can reduce the cost of atomistic modelling while preserving quantitative accuracy on reaction energetics. The project naturally builds on previous IJL work on alloy surfaces, adsorption and CO₂ hydrogenation, and on recent developments in machine-learning potentials for surfaces and reactive interfaces.

The Université de Lorraine is a leading multidisciplinary university in France, renowned for the quality of its research and its strong ties to national and international scientific networks. It brings together numerous laboratories of excellence, including the Jean Lamour Institute (IJL) and LORIA (Lorraine Laboratory for Research in Computer Science and its Applications). The IJL is one of the largest European laboratories in materials science and engineering, covering a broad spectrum of topics ranging from materials physics to their development and characterization. LORIA, for its part, is a leading laboratory in computer science, particularly in the fields of artificial intelligence, algorithms, robotics, and data science. The collaboration between the IJL and LORIA offers a particularly stimulating interdisciplinary research environment at the interface between materials science and advanced computer science, and provides an ideal framework for developing innovative research projects within the context of a doctoral thesis.

As part of this PhD project, the research is embedded in a dynamic and well-structured scientific environment at both the national and international levels. In France, the project is connected to major research networks such as the GDR IA-MAT and the PEPR DIADEM community, which promote interdisciplinary collaborations at the interface between materials science, artificial intelligence, and data science. Internationally, the work also benefits from active collaborations, notably through a joint associated laboratory with the Jožef Stefan Institute in Slovenia and participation in the European ECMetAC network dedicated to metallic alloys. This ecosystem provides the PhD candidate with a rich research environment that encourages collaboration, knowledge exchange, and strong connections with leading experts in materials science.

AApplicant skills: Strong background in chemistry, physical chemistry, materials science, or condensed matter  physics. Experience in data science, Python programming, high-performance computing and/or quantum chemistry will be considered an asset. Excellent communication skills are essential, with the ability to work and exchange ideas effectively both orally and in writing. English speaking is required. The application should include a statement of research interest, a CV and Master’s degree transcript.