Dice Question Streamline Icon: https://streamlinehq.com

Out-of-Distribution simulation of PXD detector signatures beyond known conditions

Develop generative methods that simulate Pixel Vertex Detector (PXD) hits conditioned on background rate and luminosity in out-of-distribution regimes beyond available experimental data, including both length extrapolation of per-sensor hit multiplicities and context extrapolation from only event-level attributes, without access to high-luminosity training data.

Information Square Streamline Icon: https://streamlinehq.com

Background

The thesis highlights that real PXD background data from Belle II is substantially sparser and exhibits high variance compared to Geant4 simulations, making image-based approaches inadequate. Moreover, reliance on random trigger data ties availability to specific experimental runs, leaving higher luminosity regimes without real background data.

Within this context, out-of-distribution (OOD) simulation is essential to enable extrapolation to new beam parameters, energies, luminosities, and detector geometries. The chapter frames two key scenarios: length extrapolation (the model accesses per-sensor conditions and must generate larger cardinalities than seen in training) and context extrapolation (the model receives only event-level attributes and must infer per-sensor multiplicities and correlations).

References

This is one of the most intriguing open problems in the realm of Deep Generative Models for Particle Physics is the possibility of simulating particle detector signatures that extend beyond the known experimental conditions. From the real PXD data perspective, the idea of generating PXD hits conditioned over the amount of background~(rate of background) and luminosity even beyond the current experimental data is still an open problem.

— Deep Generative Models for Ultra-High Granularity Particle Physics Detector Simulation: A Voyage From Emulation to Extrapolation (2403.13825 - Hashemi, 5 Mar 2024) in Chapter 6, Introduction