PhD: PhD candidate, Generative (X)AI for the Discovery of Domain-Specific Optimizers
Where: Leiden, 
Netherlands
What:
Within the NWO AI4optics project, Leiden University leads Work Package B.2, “Generative (X)AI for the discovery of domain-specific optimizers”. This project advances new methods for automatically discovering optimization algorithms tailored to specific optical design problems, in both imaging and illumination (non-imaging) optics.
The approach builds on Leiden's own evolutionary framework for automatic black-box optimization algorithm discovery (LLaMEA), which places a Large Language Model inside a generate-test-refine loop: the LLM proposes solver code, the candidate is benchmarked, and the LLM is asked to improve it based on structured feedback — repeated many times to evolve families of algorithms that rival or beat hand-crafted baselines. In this project, you will further integrate optical domain knowledge, so that discovered solvers can tackle real-world optical physics problems, and combine LLM-driven code generation with in-the-loop hyper-parameter optimization (e.g. Bayesian optimization), so the LLM can focus on algorithmic structure while numerical tuning is handled automatically.
A second, equally important line of research is explainability. Building on Leiden's “code evolution graphs” and mechanistic interpretability research, which analyze how LLM-generated algorithms evolve through iterative mutation and refinement, you will develop richer visualization and explanatory methods that reveal which code changes and evolutionary strategies actually improve solver performance, linking these directly to improvements in key optical performance metrics such as wavefront error and stray light. This will help domain experts validate and trust automatically discovered solvers. You will also design prompt templates and retrieval-augmentation techniques that embed physics- and optics-specific constraints into the code generation process, and develop curated benchmarks reflecting industrially relevant optics use cases.
Who: Niki van Stein - n.van.stein@liacs.leidenuniv.nl
When: Until 2026-08-30 22:00
Presented at next GECCO?: yes