Feyer is building algorithms that can discover new optical, quantum, and other physical technologies. For this they are building AI models that learn from differentiable physics simulators.
Their design methods, built on differentiable physics simulators, have already discovered new experiments that were built and tested in laboratories. With funding from SPRIND’s Next Frontier AI Challenge, they are now building a new generation of optimizers for physics design.
Most optimizers use only loss values and local gradients. The Neural Explorer also receives the simulator’s computational graph, including its topology, operations, intermediate values and gradients. It learns to propose larger moves that may escape local optima.
- Graph neural architectures and representations for large JAX computational graphs
- Learned optimizers, non-local search and combinations of neural and classical methods
- Scalable data generation, GPU training and rigorous equal-compute benchmarks
- An enrolled PhD researcher in ML, optimization, scientific computing, differentiable programming or a related field
- A careful researcher who can design clear experiments for open-ended questions
- Excellent Python skills; experience with JAX or a comparable autodiff stack
- Comfort working independently in a small research team
Physics expertise is welcome, but not required.
Duration: 3 to 6 months
Start: Flexible / ideally ASAP
Location: Tübingen AI Center or Munich Urban Colab
Format: Paid, 80% to full-time internship
Send your CV, a short note on fit, preferred dates and desired duration to [email protected].
For questions, contact [email protected].
