Speaker
Description
While the elusive nature of neutrinos makes their detection non-trivial, they are an excellent candidate for long-range monitoring of nuclear reactors. Of the multitude of byproducts produced when an atom splits during nuclear fission, only antineutrinos can readily pass through meters of water and concrete shielding surrounding nuclear fuel. Antineutrino reactions in a water-Cherenkov detector produce a variety of complex signatures in the detector. To model these processes, Monte-Carlo simulations are employed to understand the probabilities of observing certain detector signals given neutrino parameters such as event vertex, direction and energy. Once the detector is operational, we aim to understand the opposite: the relative probabilities of competing neutrino event hypotheses given an observed detector signal. To accomplish this task, a likelihood function can be used to infer the characteristics of the event given a detector signal.
The process of generating likelihood functions often involves intractable integrals or model simplifications. Rather than explicitly calculating these integrals and sacrificing complex model behaviors, it is possible to create a predictive model using an Approximate Bayesian Computation (ABC) with machine learning tools. In particular, we generate a predictive algorithm with a technique called Likelihood-Free Inference. In this talk, we explore this deep learning technique and its applicability for reconstruction of low-energy neutrino inverse beta decay events in these detectors. Of particular interest with this technique is the flexibility of incorporating a broad range of detector configurations (geometry, scintillator, etc.) and physics hypotheses (kinematics, particle identification, etc.). Beyond neutrino physics, this technique can find utility where Bayesian analysis of a model is required for a system where a Monte Carlo simulation exists.
| Collaboration | EOS |
|---|
