Ruhr-Universität Bochum

Research at MPI SusMat · Teaching at Ruhr University Bochum

Janssen Lab Materials Informatics

Research

Computational methods for sustainable materials

Designing sustainable materials requires rethinking the whole materials design process. The Materials Informatics group at MPI SusMat combines physical modelling, statistical learning, workflow technology and AI to make computational materials science predictive, reproducible and scalable.

Computational thermodynamics

From thermodynamics and statistical mechanics to predictive materials modelling.

Whether a material is stable, how it transforms with temperature and how defects affect its properties is governed by free energies. Computing them from first principles requires density functional theory, statistical mechanics and careful sampling of vibrational, electronic and configurational contributions.

We develop automated approaches for finite-temperature properties – from melting temperatures of elemental crystals to complete composition–temperature phase diagrams – and quantify the numerical precision of the underlying electronic-structure calculations.

  • Phase stability
  • Free energies
  • Defects
  • Finite temperature
  • DFT
Workflow diagram in four panels connected by arrows: data generation with DFT, fitting of EAM, neural-network and ACE potentials, validation of bulk modulus, convex hull and phonons with LAMMPS and phonopy, and calculation of free energies and a phase diagram with calphy – all orchestrated by pyiron, with active-learning loops back to data generation.
From electrons to phase diagrams: an automated pyiron workflow generates DFT data, fits machine-learned interatomic potentials, validates them and computes the Al–Li phase diagram. Figure 1 from S. Menon et al., npj Comput. Mater. 10, 261 (2024), doi:10.1038/s41524-024-01441-0, licensed under CC BY 4.0.

Related publications

  • S. Menon et al., From electrons to phase diagrams with machine learning potentials using pyiron based automated workflows, npj Computational Materials 10, 261 (2024). doi:10.1038/s41524-024-01441-0
  • J. Janssen, E. Makarov, T. Hickel, A. V. Shapeev and J. Neugebauer, Automated optimization and uncertainty quantification of convergence parameters in plane wave density functional theory calculations, npj Computational Materials 10, 263 (2024). doi:10.1038/s41524-024-01388-2
  • L.-F. Zhu, J. Janssen, S. Ishibashi, F. Körmann, B. Grabowski and J. Neugebauer, A fully automated approach to calculate the melting temperature of elemental crystals, Computational Materials Science 187 (2021). doi:10.1016/j.commatsci.2020.110065

Machine learning for materials

From machine-learning fundamentals to predictive atomistic models.

Machine-learned interatomic potentials bring near-DFT accuracy to length and time scales that are inaccessible to electronic-structure methods. Their predictive power depends on the training data, the model's expressivity and the precision of the reference calculations – trade-offs that we analyse systematically.

Beyond potentials, we use Bayesian and statistical methods, such as (deep) Gaussian processes and batch Bayesian optimisation, to explore complex chemical spaces like high-entropy alloys with quantified uncertainty and to decide which calculation or experiment is most informative next.

  • Interatomic potentials
  • Active learning
  • Bayesian optimisation
  • Uncertainty
  • Materials discovery
Nine contour maps, one per fcc metal from calcium to gold, showing the error in the computed bulk modulus as a function of plane-wave energy cutoff and k-point mesh; red contour lines mark target precisions of 5, 1, 0.5 and 0.1 GPa.
How precise is a DFT result? Error in the bulk modulus of nine fcc metals across energy cutoff and k-point sampling. Contour lines mark target precisions from 5 GPa down to 0.1 GPa. Figure 1 from J. Janssen et al., npj Comput. Mater. 10, 263 (2024), doi:10.1038/s41524-024-01388-2, licensed under CC BY 4.0.

Related publications

  • I. Baghishov et al., Application-specific machine-learned interatomic potentials: exploring the trade-off between DFT convergence, MLIP expressivity, and computational cost, Digital Discovery 5 (2026). doi:10.1039/d5dd00294j
  • Sk M. A. A. Alvi et al., Deep Gaussian process-based cost-aware batch Bayesian optimization for complex materials design campaigns, npj Computational Materials 12, 105 (2026). doi:10.1038/s41524-026-01981-7
  • Sk M. A. A. Alvi et al., Accurate and uncertainty-aware multi-task prediction of HEA properties using prior-guided deep Gaussian processes, npj Computational Materials 11, 306 (2025). doi:10.1038/s41524-025-01811-2
  • Sk M. A. A. Alvi et al., Hierarchical Gaussian process-based Bayesian optimization for materials discovery in high entropy alloy spaces, Acta Materialia 289 (2025). doi:10.1016/j.actamat.2025.120908
  • V. F. Grizzi et al., ALF: Open-Source Active Learning Framework for Atomistic Modeling, Journal of Chemical Theory and Computation (2026). doi:10.1021/acs.jctc.6c01122
  • A. Rohskopf et al., FitSNAP: Atomistic machine learning with LAMMPS, Journal of Open Source Software 8, 84 (2023). doi:10.21105/joss.05118

Scientific workflows & HPC

From Python functions to reproducible simulations on supercomputers.

Modern simulation studies combine many codes, thousands of calculations and heterogeneous computing resources. We develop the open-source infrastructure that makes this manageable: the pyiron framework, scalable execution of Python functions on HPC systems with executorlib and pysqa, and the Python Workflow Definition for exchanging workflows between workflow managers.

Provenance, interoperability and semantic annotation make workflows reusable scientific objects – connecting simulation with experiments, data platforms and knowledge graphs.

  • pyiron
  • Provenance
  • Interoperability
  • HPC
  • Knowledge graphs

Research software

Flow chart: an Executor submits a Python function, which requests resources from the job scheduler; the function is executed, its status can be checked with done(), and the result is communicated back via result() before shutdown().
executorlib extends Python's standard Executor interface: a submitted function requests resources from the job scheduler, runs asynchronously and returns its result. From J. Janssen et al., J. Open Source Softw. 10, 7782 (2025), doi:10.21105/joss.07782, licensed under CC BY 4.0.

Related publications

  • J. Janssen et al., A Python workflow definition for computational materials design, Digital Discovery 4, 3149–3161 (2025). doi:10.1039/d5dd00231a
  • J. Janssen et al., Executorlib – Up-scaling Python workflows for hierarchical heterogeneous high-performance computing, Journal of Open Source Software 10, 7782 (2025). doi:10.21105/joss.07782
  • J. Janssen and J. Neugebauer, pysqa – Python Simple Queuing System Adapter, Journal of Open Source Software (2026). doi:10.21105/joss.10961
  • J. Janssen et al., Toward Full Interoperability in Materials Science: Integrating Workflows With Knowledge Graphs, Advanced Engineering Materials (2026). doi:10.1002/adem.71258
  • S. Brinckmann et al., A Lightweight Procedural Layer for Hybrid Experimental–Computational Workflows in Materials Science, Advanced Engineering Materials (2026). doi:10.1002/adem.202503176
  • M. Stricker et al., Computationally accelerated experimental materials characterization – drawing inspiration from high-throughput simulation workflows, npj Computational Materials 12, 2 (2026). doi:10.1038/s41524-025-01919-5
  • J. Janssen et al., pyiron: An integrated development environment for computational materials science, Computational Materials Science 163, 24–36 (2019). doi:10.1016/j.commatsci.2018.07.043

AI for science

From language models to autonomous scientific workflows.

Large language models can translate a scientific question into a sequence of simulation steps – but only if they are coupled to validated tools rather than generating results themselves. We develop agents that use established simulation codes and workflow frameworks, so that every number an agent reports is backed by a reproducible calculation.

Digital, well-annotated workflows are the prerequisite: they make simulation protocols machine-actionable and prepare materials science data for AI.

  • LLM agents
  • LangSim
  • Autonomous workflows
  • AI readiness

Related publications

  • Y. Zimmermann et al., 32 examples of LLM applications in materials science and chemistry: towards automation, assistants, agents, and accelerated scientific discovery, Machine Learning: Science and Technology 6, 030701 (2025). doi:10.1088/2632-2153/ae011a
  • M. Bruns et al., Supporting AI Readiness Through Digital Workflows in Materials Science, Advanced Engineering Materials (2026). doi:10.1002/adem.71262

Interested in one of these themes?

Student projects are connected to ongoing research in all four themes and adapted to your background.

Student projects