Please read these instructions before posting any event on Fermilab Indico

Indico search will be reestablished in the next version upgrade of the software:


This search is only for public events. Restricted events are not available.


IMPORTANT! Indico has been upgraded. Please let us know as soon as possible if you find any issues and email

Quantifying Reconstruction Uncertainty using Probabilistic Graphical Models


The success of convolutional neural networks (CNNs) in computer vision has resulted in techniques that have been widely adopted within the physical sciences. For instance, CNNs are often used in reconstruction of particle interactions, where the physical processes occurring within a detector are inferred from sensor observables.  However, such techniques provide a single locally optimal solution (i.e., energy or particle type) rather than a posterior probability distribution. This is a major deficiency of CNNs, as experimental science relies on uncertainties when comparing different measurements.  Additionally, a CNN rarely provides a physically interpretable model that can provide insight into the physical processes of the system.

Probabilistic graphical models (PGMs) represent a system as a graph whose nodes are probabilistic variables, while edges represent dependencies among the variables. Such a representation readily allows incorporating domain knowledge into the model through setting constraints over types and values of the variables as well as over the graph structure. These graph-based representations model complex systems compactly, capturing uncertainty that is inherent in both the modeled system and in the predictions made by using the model. PGMs can be used to compute the posterior probability over a variable of interest and are readily interpretable by a scientist who wishes to understand the model's properties.

Within the DIDACTS collaboration (, we have been exploring how PGMs can be applied within particle physics. In doing so, we aim to provide machine learning methods that are more appropriate to these applications. We aim to explicitly model uncertainty when using machine learning methods in the context of direct-detection of dark matter in XENON1T, which is a rare-event search whereby uncertainty on position is a critical parameter to sensitivity.

In this talk I will give an introduction to PGMs at the level of a tutorial. I will then demonstrate how these techniques translate to our application of reconstruction of dark matter events in XENON1T.

Recorded Meeting Video:

    • 1:00 PM 1:45 PM
      Quantifying Reconstruction Uncertainty using Probabilistic Graphical Models 45m
      Speaker: Christina Peters (University of Delaware)