Atomistic modeling plays a central role in materials science, enabling researchers to understand and predict the behavior of materials at the atomic scale. It provides critical insights into structure-property relationships, guides the design of novel materials, and accelerates the discovery of functional compounds for applications ranging from energy storage to catalysis.
Global optimization of atomic structures encompasses a suite of techniques aimed at identifying the lowest-energy configuration of a given set of atoms with minimal computational overhead. These methods typically rely on efficient sampling and exploration of the potential energy surface, which is often derived from electronic structure calculations such as density functional theory (DFT).
In recent years, machine learning models have emerged as powerful alternatives to traditional electronic structure methods, offering comparable accuracy at a fraction of the computational cost. For instance, Gaussian processes, as implemented in the BOSS code, have proven particularly effective for active Bayesian optimization of atomic structures, even when starting from sparse datasets. Another promising approach involves universal interatomic potentials like MACE, which are trained on extensive DFT databases. These models generalize well across the periodic table for predicting interatomic forces, though they may face challenges in accurately distinguishing low-energy configurations within specific systems.
We propose two research projects focused on atomistic modeling and global optimization within the AI-based Materials Science (AI4MS) group at TUM.
Project 1: Multi-task learning
Multi-task Gaussian processes offer a robust framework for capturing correlations between multiple atomic structure properties. A notable application is multi-fidelity learning, where two computational models predict the same property - such as total energy - but differ in accuracy and computational cost. This approach has gained significant attention for its ability to accelerate tasks like global optimization by effectively combining fast, lower-fidelity models (e.g., MACE) with more accurate but computationally intensive methods such as DFT. Current research focuses on developing optimal task-sampling strategies that enhance model performance while keeping computational demands to a minimum.
Project 2: Grand canonical optimization
Atomic structure optimization traditionally assumes a fixed number of atoms or a predefined stoichiometry. However, many real-world applications demand the discovery of functional materials where the optimal elemental composition is unknown. Grand canonical optimization is an emerging approach that explores structures across varying stoichiometries, potentially integrating property optimization with free energy minimization. As this area is still in its infancy, it presents exciting opportunities for both methodological innovation and impactful applications in materials discovery.
Tasks
(i) Review existing literature and identify promising solutions.
(ii) Implement suitable algorithms, e.g., into the BOSS code.
(iii) Test and validate the implemented methods.
(iv) Apply the methods to atomic systems of interest to the AI4MS group or experimental collaborators.