Jannes Nys

I design neural architectures — equivariant graph neural networks, attention-based models, and physics-informed networks — for challenging scientific problems. Recent threads include generative models trained with optimal-transport ideas (Wasserstein Quantum Monte Carlo, NeurIPS 2023), deep learning for time-dependent many-body dynamics (Nature Communications 2024), message-passing wave functions for the homogeneous electron gas (PRB 2024), higher-order optimization techniques for neural wave functions (arXiv 2025), and LLM-based multi-agent systems for automating parts of the scientific discovery loop.

I design quantum algorithms for digital and analog devices, with three current threads. For simulating fermionic matter and 2D materials, I work on Clifford-friendly stabilizer methods for fermion dynamics with reduced non-Clifford gate counts (with I. Tavernelli's group at IBM Research, arXiv 2025), quantum circuits for local fermion-to-qubit mappings (Quantum 2023), and generalized fermion-to-qudit encodings (Phys. Rev. A 2025). For quantum dynamics, I develop variational time-evolution algorithms that bypass the quantum geometric tensor (with J. Gacon, Phys. Rev. Research 2024). And for validating quantum devices, I build ML methods that recover hard-to-measure observables from analog simulators — e.g. predicting topological entanglement entropy on a Rydberg simulator (Nature Physics 2025).

Underneath the model design sits a steady focus on scalable numerics and linear algebra: foremost neural networks and deep learning, scaled with JAX and PyTorch on multi-GPU and HPC clusters, and supporting techniques like Monte Carlo, tensor networks, and custom optimization algorithms. I originally cut my teeth on this in industry with TensorFlow at Boltzmann, and now contribute to NetKet, a JAX-based open-source toolbox for many-body quantum systems (SciPost 2022).