I develop machine-learning methods for problems in the natural sciences. My work sits between approximation theory and scientific machine learning: I use invertible neural networks and normalizing flows to learn adaptive approximation spaces, and apply them to quantum molecular physics and, more recently, image compression.

I am currently a consultant at d-fine GmbH and continue my research independently. Before that I was a postdoctoral researcher at the Department of Mathematics, Universität Hamburg, and did my PhD there and at the Center for Free-Electron Laser Science (CFEL), DESY.

• Inducing Riesz Bases in L² via Composition Operators
  Y. Saleh, A. Iske · Complex Analysis and Operator Theory 20 (2026) · paper · arXiv
• Convergence theory for Hermite approximations under adaptive coordinate transformations
  Y. Saleh · arXiv preprint (2026) · arXiv
• Computing Excited States of Molecules Using Normalizing Flows
  Y. Saleh, Á. F. Corral, E. Vogt, A. Iske, J. Küpper, A. Yachmenev · Journal of Chemical Theory and Computation 21 (2025) · paper · arXiv
• Bounds on the Generalization Error in Active Learning
  V. Menden, Y. Saleh, A. Iske · Proceedings of the 6th Northern Lights Deep Learning Conference (NLDL) 265 (2025) · paper · arXiv
• Active learning of potential-energy surfaces of weakly bound complexes with regression-tree ensembles
  Y. Saleh, V. Sanjay, A. Iske, A. Yachmenev, J. Küpper · The Journal of Chemical Physics 155 (2021) · paper · code

• 2026 · Learning, sharing, and using optimized vibrational coordinates — ExoMol 2026, University College London, United Kingdom (invited)
• 2026 · Learning bases via normalizing flows, applications to solving molecular Schrödinger equations and image compression — International Conference on Scientific Computing and Machine Learning, Bath, United Kingdom
• 2025 · Normalizing flows: from learning probability distributions to basis-set discovery — Linköping University, Sweden (invited)
• 2025 · Learning basis sets using unitary and bounded bijective operators — International Conference in Numerical Mathematics and Scientific Computing, Uppsala University, Sweden
• 2025 · Bounds on the generalization error in active learning — 6th Northern Lights Deep Learning Conference (NLDL), Tromsø, Norway (poster)

20 talks and posters in total, 6 of them invited — full list.

_{This profile is generated from the same data as my website and CV.}