Dr. Timo Janßen

Machine learning is increasingly explored to accelerate Monte Carlo event generation, but precision applications require strictly unbiased results. I present ML-based approaches that improve efficiency while leaving the underlying MC estimators—and thus the physics—unchanged. I first review multi-channel phase-space sampling, where analytic mappings are combined into a mixture and optimized with VEGAS, and then replace this step with Normalizing Flows. Focusing on Continuous Normalizing Flows trained via Flow Matching, I demonstrate a lightweight integration with the portable generator Pepper that requires only a minimal interface and is readily transferable to other codes. The resulting sampler substantially improves unweighting efficiency and yields measurable resource savings for HL-LHC–scale runs. Finally, I outline extensions to higher orders: for NNLO cross sections in the STRIPPER subtraction scheme, flows learn an efficient separation of positive and negative contributions, reducing Monte Carlo uncertainties at fixed compute cost.