Session T19: HEP Data Analysis In the Post-Moore Era

Live

This contribution describes the Jas4pp framework for exploring physics cases and for detector-performance studies of future particle collision experiments. Jas4pp is a multi-platform Java program for numeric calculations, scientific visualization in 2D and 3D, storing data in various file formats and displaying collision events and detector geometries. It also includes complex data-analysis algorithms for function minimisation, regression analysis, event reconstruction (such as jet reconstruction), limit settings and other libraries widely used in particle physics. The framework can be used with several scripting languages, such as Python/Jython, Groovy and JShell. Several benchmark tests discussed in the paper illustrate significant improvements in the performance of the Groovy and JShell scripting languages compared to the standard Python implementation in C. The improvements for numeric computations in Java are attributed to recent enhancements in the Java Virtual Machine.   

Live

In high energy experiments, reconstructing charged particle trajectories is one of the most CPU intensive tasks. A Common Tracking Software (ACTS) is a collaborative project to generate track reconstruction tools that are agnostic to specific detector geometries. Track seeding is the first stage of track reconstruction which produces short track candidates for further refinement. Currently, the seeding algorithm in ACTS relies on many hand tuned, detector specific cuts inherited from the ATLAS experiment and configurable parameters that are not optimal for other geometries. In this paper, we investigate the application of machine learning methods to optimize the seed finding algorithm parameters for multiple detector geometries. We find that an evolutionary algorithm performed best on both the generic detector in ACTS and the Light Dark Matter Experiment (LDMX) tagging tracker, with efficiencies of 99.4\% and 97.89\% respectively. We find that evolutionary algorithms are a powerful tool for parameter optimization. respectively.   

Live

NOvA is a long-baseline neutrino oscillation experiment. It measures both the disappearance of muon neutrinos and the appearance of electron neutrinos at a large distance from a source of muon neutrinos in order to constrain neutrino oscillation parameters. In particular, NOvA aims to precisely measure $\theta_{23}$ and $\Delta{}m^{2}_{32}$, as well as determine the neutrino mass hierarchy and probe the existence of CP-violation in the neutrino sector. NOvA is a low-statistics experiment and the neutrino oscillation model contains several physical boundaries. Constructing statistically correct confidence intervals is therefore challenging. NOvA ensures a correct statistical coverage by following the computationally expensive Feldman-Cousins prescription. This approach requires the generation and fitting of millions of pseudo-experiments to build empirical test-statistics distributions. This talk will describe this technique and the framework implemented on supercomputers that reduced the time necessary to produce statistically robust results from several months down to a few days.   

Live

In many phenomenological studies in which the full accuracy offered by the detector simulator GEANT4 is not required, faster alternatives are used in which the detector response is approximated as a parametric function. One drawback to this method is that the parametric function must be hand-coded, and should the experiment change for any reason the detector response must be re-coded. Instead of hand coding, Falcon seeks to use deep generative models to learn the detector response function. As part of the efforts of the Falcon group, conditional generative adversarial networks were used to learn the mapping from parton level jets to reconstruction level jets. Results from this model using simulated events in the Compact Muon Solenoid detector will be presented. The performance of the machine learning models will be compared with existing detector simulators.   

Not Participating

We use machine learning to achieve faster, more complete profile-likelihood ratio (PLR) calculations in the LUX dark matter experiment -- a method which can be straightforwardly applied to other rare event searches. One of the main drawbacks of the PLR method with histogrammed probability density functions (PDFs) is that it becomes computationally intractable to include many input dimensions. This can be partially addressed by breaking up the variable space into products of lower-dimensional PDFs, but this potentially sacrifices information in the form of correlations between variables. To combat this issue, we first compress the inputs into a single dimension using a neural network which is then used as the input to the PLR method. We ensure all relevant information is preserved by imposing that the mutual information between the output of the neural network and the signal/background designation matches that of the inputs; this also provides an absolute calibration for when to stop the training process. This approach reduces the PLR computation time by an order of magnitude or more while allowing straightforward inclusion of additional highly-correlated variables such as an S1 prompt fraction variable.   

Live

The unified approach of Feldman and Cousins allows for estimating confidence intervals for datasets with small statistics that commonly arise in high energy physics. It has gained widespread use, for instance, in measurements of neutrino oscillation parameters in long-baseline experiments. However, the approach is computationally intensive as it is typically done in a grid-based fashion over the entire parameter space. In this talk, I will discuss a more efficient algorithm for the Feldman-Cousins approach using Gaussian processes to construct confidence intervals iteratively. I'll show that in the neutrino oscillation context, one can obtain confidence intervals fives times faster in one dimension and ten times faster in two dimensions, while maintaining an accuracy above 99.5{\%}. I'll also discuss next steps related to the implementation in the NOvA FC framework at NERSC.   

Live

Clustering of charged particle tracks along the beam axis is the first step in reconstructing the positions of proton-proton (p-p) collisions at Large Hadron Collider (LHC) experiments. In this talk, we formulate this problem for a 2048 qubit D-Wave quantum computer that works by quantum annealing. We showcase the performance of the quantum annealer on artificial events generated from p-p collision and track distributions measured by the Compact Muon Solenoid experiment at the LHC. This performance is enhanced via multiple hardware optimizations which are outlined in the talk. The quantum clustering algorithm is found to be limited by the connectivity of the qubits and the overall efficiency of the algorithm in addressing event topologies with more than 5 collisions. Current research directions are highlighted in extending this algorithm to be compatible with operating at the full LHC-scale problem complexities relevant for particle physics research.   

Live

Using quantum computers to perform ab-initio calculations of the real-time evolution of quarks and gluons with lattice quantum chromodynamics is a long-term goal with potentially high impact on the interpretation of particle collider experiments. Currently available NISQ machines only allow us to study the real-time evolution for very simple quantum field theory models such as the quantum spin or gauge Ising models in small spatial volumes. We discuss how the successful roadmap starting with these simple models and leading to lattice QCD calculations of masses and form factors using Euclidean time and importance sampling can be adapted to the problem of real-time evolution. We report on recent quantum computations of real-time evolution for a quantum Ising model on three different IBMQ computers. We show that a metric that we proposed demonstrates significant progress for the most recent machines. We discuss a new method to estimate phase shifts from the early stage of a scattering process. We report on our most recent results with NISQ quantum computers at the time of the conference.   

Live

A framework is presented to extract and understand decision-making information from a deep neural network classifier of jet substructure tagging techniques. There are two methods studied. The first is using expert variables that augment the inputs ("expert-augmented" variables, or XAUGs). These XAUGs are concatenated to the classifier steps immediately before the final decision. The second is layerwise relevance propagation (LRP).The results show that XAUG variables can be used to interpret classifier behavior, increase discrimination ability when combined with low-level features, and in some cases capture the behavior of the classifier completely. The LRP technique can be used to find relevant information the network is using, and when combined with the XAUG variables, can be used to rank features, allowing one to find a reduced set of features that capture part of the network performance. These XAUGs can also be added to low-level networks as a guide to improve performance.