2017–2021 ·
Machine learning for particle physics simulations
How can machine learning accelerate computationally expensive physics simulations for particle collision processes at CERN without losing reliability or uncertainty information?
Physics-informed neural networks integrated with Monte Carlo generators to efficiently generate outputs with attached uncertainties.
The original challenge
High-precision particle-physics simulations repeatedly evaluate mathematically complex scattering amplitudes (probabilities). At high multiplicity (large numbers of output particles) these calculations become a major computational bottleneck inside Monte Carlo event generators.
Research contribution
This work developed ensembles of neural networks to serve as fast surrogate functions for the underlying amplitudes. Physics-informed phase-space partitioning improved performance near challenging singular regions, while explicit validation examined the effect of approximation errors on integrated and differential observables.
This approach was integrated into the SHERPA Monte Carlo generation software.
Connection to current work
This background forms the theoretical and technical foundation of much of my work: scalable simulations, the use of surrogate models, uncertainty propagation and the design of computational methods that remain reliable when embedded in larger decision processes.
Selected outputs