Materials Informatics
The Materials Informatics group’s aim is to leverage the recent developments in computer science and informatics to accelerate the discovery of sustainable materials. This includes but is not limited to machine learning techniques ranging from machine-learned interatomic potentials to large language models.
For the design of sustainable materials, it is essential to drastically rethink the materials design process. It is no longer sufficient to just slightly optimize material properties of already existing alloys, but rather the design of sustainable materials requires rethinking the whole materials design process. To accelerate the materials discovery, simulation and machine learning models are applied to sample the periodic table and identify suitable alloy candidates. In this perspective, the field of materials informatics combines experiment, simulation and machine learning in one discipline and is consequently an interdisciplinary research field. This combination is enabled by scientific workflows which cover the input parameter, processing conditions and final measurement technique to determine the desired material property, in one document which is both human-readable and machine-readable.
The Materials Informatics group, established in September 2023, is leading the development of the pyiron workflow framework as a toolbox to construct workflows to discover sustainable materials. While the initial version of the pyiron workflow framework was designed to calculate thermodynamic properties of materials with atomistic ab-initio simulation, it has since been extended to support the calculation of defect properties including defect phase diagrams and kinetics as well as to interface with experimental measurements.
Method Development
Based on the workflow-centric approach of the Materials Informatics group, we develop data-driven machine-learned models to accelerate the discovery of sustainable materials:
- Uncertainty Propagation for Density Functional Theory (DFT): While DFT is in principle exact, the exchange functional remains unknown, which limits the accuracy of DFT calculation. Still, in addition to the accuracy of the exchange functional, the quality of material properties calculated with DFT is also restricted by the choice of finite bases sets. For the case of plane-wave DFT these are the reciprocal k-point grid and the plane-wave energy cut-off. The resulting uncertainty directly effects the prediction of materials properties with DFT.
- Machine-Learned Interatomic Potentials (MLIP): Plane-wave DFT typically scales cubically with the number of electrons in the simulation cell. This limits the number of atoms to a few thousands. To address this limitation interatomic potentials, approximate the interaction between atoms based on the position of their nuclei, which enables linear scaling with the number of atoms.
- Statistical Sampling of Chemical Space: Brute Force sampling of the chemical space is prohibitive even for simulation approaches. To enable the efficient sampling and accelerate the discovery of sustainable materials, Bayesian approaches are developed to guide the scientific discovery.
Research Software Development
In addition to the pyiron workflow framework, the Materials Informatics group is involved in a number of open-source scientific software projects, to accelerate the discovery of sustainable materials:
- Scientific Foundation Models: While Large Language Models (LLM) still struggle to consistently generate trust worthy scientific workflows, this challenge is addressed with the development of agentic interfaces to scientific workflow frameworks. This enables the LLM to access and combine predefined interfaces, which have previously been validated by expert users.
- High-Performance Computing (HPC): To up-scale simulation workflows from a laptop to the worlds leading Exascale Computers, the Materials Informatics group closely collaborates with leading research laboratories in the US on developing workflow solutions for the Exascale.
- Reproducible Workflows: To enable the transferability and reproducibility of scientific workflows between different computer centers, the Materials Informatics group collaborates in a network of open-source projects to identify Workflow Standards and contributes to the conda-forge community project to distribute open-source Materials Science software.
If the research directions of the Materials Informatics group sound interesting to you, feel free to reach out and get in contact with Jan. We are always open for collaboration and looking for exceptional students to join us on the quest to discover the materials of the future.
People | Projects | Publications |