Below are examples of projects our group can offer as semester or Master thesis projects. This list is not exhaustive and other topics related to our research are equally possible. Several of the projects could also be adapted to the Bachelor level.
If you are interested in doing a project with us, please make yourself familiar with general guidelines we follow during the project. We will in general not accept or decline any project until six weeks before the semester begins.
Overview of projects:
A key step when simulating the properties of a material is structure optimisation, in which an approximate arrangement of atoms in a crystal is optimized to the most stable configuration. In this work we will investigate how reducing the size of the discretisation basis impacts the quality of the obtained structure in such a structure optimisation. Moreover we will employ recent perturbative error estimates[1] in order to quantify the expected error in the obtained structure versus a fully converged discretisation. For this work we will employ the density-functional toolkit (DFTK), a first-principle materials simulation code based on density-functional theory in which the aforementioned error estimates are implemented and readily available.
Requirements: Strong programming skills, ideally Julia or python; Basic knowledge of numerical methods for partial differential equations; Experience in numerical analysis of PDEs is a bonus; Experience in running DFT calculations is a bonus;
| [1] | E. Cancès, G. Dusson, G. Kemlin and A. Levitt. SIAM J. Sci. Comp. 44 (2022). ArXiv 2111.01470 |
Building large datasets with materials properties from density-functional theory (DFT) calculations is a challenge. Active learning techniques try to efficiently query simulators iteratively, based on a statistical model [2]. The computational cost of DFT is however not uniform across materials. Understanding the cost for a given target accuracy is a problem of error control with numerical parameters.
One of the core parameters determining the cost of a DFT calculation is the discretisation. The baseline active-learning approach is computing with a fixed discretisation (plane-wave cutoff) chosen a priori for the whole dataset (e.g. [3] and [4]).
The goal of this project is to formulate and implement an augmented active learning model which can choose not only the next material structure to query but also choose an appropriate discretisation adaptively, trading off per-example uncertainty reduction against computational cost.
Requirements: Strong programming skills, ideally Julia or python; Basic knowledge of numerical methods for partial differential equations; Experience with probabilistic machine learning methods is a bonus; Experience in running DFT calculations is a bonus;
| [2] | R. Garnett. Bayesian Optimization. Cambridge University Press (2023). |
| [3] | C. van der Oord, M. Sachs, D. P. Kovács, C. Ortner and G. Csányi . Hyperactive learning for data-driven interatomic potentials. npj Comput Mater 9, 168 (2023). DOI 10.1038/s41524-023-01104-6 |
| [4] | A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon and E. D. Cubuk. Scaling deep learning for materials discovery. Nature 624, 80–85 (2023). DOI 10.1038/s41586-023-06735-9 |
Data-driven materials modeling has been shown to be essential in modern materials discovery. Statistical models trained on large datasets of first-principle simulations provide efficient and accurate predictions of materials properties, reducing the need for costly computations. However, the underlying assumption of uniformly high-quality training data doesn't always meet reality. When faced with data from diverse sources, incorporating the different level of uncertainty in the data is necessary to ensure accurate predictions of the quantity of interest.
In this project, we will use Gaussian Process (GP) regression that offers an approach to efficiently handle data with varying quality, providing probabilistic predictions which enable quantification of uncertainty [5]. We will explore the potential of error propagation within GP regression with non-uniform noise model, and evaluate the accuracy of the developed model to ensure its applicability for practical data-driven materials modelling.
Requirements: Strong programming skills, ideally Julia or Python; experience with probabilistic machine learning methods, Gaussian Processes, Bayesian optimization; experience with DFTK is a bonus; basic knowledge of numerical methods for partial differential equations is a bonus.
| [5] | C. E. Rasmussen & C. K. I. Williams, Gaussian Processes for Machine Learning, the MIT Press, 2006. DOI 3206.001.0001 |
The problem underlying density-functional theory (DFT) is a minimisation of an energy functional with respect to the density matrix. In turn the space of density matrices themselves has the structure of a smooth manifold. However, instead of solving this minimisation problem directly (termed direct minimisation, DM), most codes solve DFT by satisfying the first-order stationarity conditions, which leads to the self-consistent field (SCF) equations. For some settings the convergence of direct minimisation can be superior[6], which has been stimulating for the recent mathematical studies of DM approaches to DFT[7]. In this project we want to build upon the readily available software stacks for optimisations on manifolds[8] and explore the use of their readily available manifold optimisation routines in the context of DFT approaches.
Requirements: Strong programming skills, ideally Julia or python; Basic knowledge of Riemannian optimisation; Basic knowledge of numerical methods for partial differential equations; Experience in numerical analysis of PDEs is a bonus; Experience in running DFT calculations is a bonus;
| [6] | E. Cancès, G. Kemlin, A. Levitt SIAM J. Mat. Anal. Appl. 42, 243 (2021). DOI 10.1137/20m1332864 |
| [7] | X. Dai, S. de Gironcoli, B. Yang, A. Zhou. Multiscale Model. Simul. 21, 777 (2023). DOI 10.1137/22M1472103 |
| [8] | https://www.manopt.org/ and https://manoptjl.org/ |
Solutions from atomic calculations are a key building block to reduce the computational cost of larger calculations of molecules and materials [9]. For example, the construction of pseudopotentials depends on solving challenging inverse problems on top of single atom calculations. Automatic differentiation, the ability to compute arbitrary parameter derivatives, enables development of novel methods using gradient-based optimization and error propagation[10][11].
The goal of this project is to implement a differentiable DFT code for isolated atoms in Julia. An important focus will be mathematical correctness and careful validation against high accuracy reference results.
Requirements: Strong numerical programming skills. Working knowledge of numerical methods for differential equations. Understanding of automatic differentiation (e.g. Julia, JAX, PyTorch) is a bonus. Knowledge of quantum physics is a bonus but not required.
| [9] | Martin RM. Electronic Structure: Basic Theory and Practical Methods. 2nd ed. Cambridge University Press; 2020. DOI 10.1017/9781108555586 |
| [10] | Blondel, Mathieu, and Vincent Roulet. "The elements of differentiable programming." arXiv preprint arXiv:2403.14606 (2024). |
| [11] | Sapienza, F., Bolibar, J., Schäfer, F., Groenke, B., Pal, A., Boussange, V., Heimbach, P., Hooker, G., Pérez, F., Persson, P.O. and Rackauckas, C., 2024. Differentiable Programming for Differential Equations: A Review. arXiv preprint arXiv:2406.09699 |
