Doctoral Researcher in Trustworthy World Models for Critical Systems — Power Grids
This PhD position focuses on designing novel neuro-symbolic learning encoders for world models applied to power grid systems. The candidate will investigate architectures that embed the electrical laws governing power flows (e.g. Kirchhoff’s laws, power balance constraints) and battery chemical dynamics directly into the learning process, so that the resulting world models remain physically consistent and generalize better than purely data-driven approaches.
Beyond architectural design, the trustworthiness dimension of the thesis will focus on logic-informed certification of these models — inspired by algorithms such as alpha-beta-CROWN — to formally verify properties of the learned world models against the symbolic constraints of the domain. The choice of which properties to certify, and their priority, will be driven directly by the operational needs of CREOS, ensuring the work targets what matters most for the real-world reliability of power grid monitoring and simulation tools.
Profile sought:
- Master’s degree in Computer Science, with a specialization in machine learning preferred
- Strong programming and analytical skills
- Knowledge of foundational and contemporary machine learning approaches (including foundation and generative models)
- Prior experience in ICT for energy is a plus
How to apply: Applications (CV + cover letter) must be submitted online through the University of Luxembourg’s HR system (applications by email are not considered). Early applications are strongly encouraged, as applications are processed upon receipt: Application here
The PhD position is part of the our research effort on Trustworthy World Models for multimodal vision-graph-tabular settings, and the candidate will work closely with 2 Postdocs and another PhD student on shared neuro-symbolic encoder architectures and certification methods, while each specializing in the mathematical/physical structure and stakeholder priorities of their own application domain.