Seminar LAMBDA: Simulation-based inference for scientific applications: a neutrino-physics case study

Аннотация: 
In many problems in experimental physics, the relation between the parameters of a physical model and detector-level observations cannot be written in closed form: the likelihood p(x | θ) is intractable. One therefore resorts to detailed Monte Carlo simulations, which make it straightforward to generate observations x for given parameter values θ. In practice, however, one needs to solve the inverse problem: to infer the parameters of the underlying model from observed data.

Simulation-based inference (SBI) addresses exactly this setting by training machine learning models on simulated data. Depending on the formulation, such a model either estimates the posterior distribution of the parameters directly or approximates the likelihood, which then enters a standard statistical analysis.

This talk introduces the main ideas of SBI through two simple examples, an idealised coin toss and a toy model of a scintillation detector, with particular emphasis on how the resulting inference can be validated. I will then turn to our application to the JUNO experiment (Commun. Phys. 9 (2026) 1, 63), where neural density estimation is used to tune the detector energy response model from calibration data. I will conclude with preliminary results on simulation budget scaling laws: how the performance of such models depends on the number of simulated events, and how large a Monte Carlo sample is required to reach a given level of accuracy.

Место проведения: online
Дата: 18.09.2026 
Время: 12:00-13:20