Session K17: Data Science I

Live

A new paradigm for data-driven, model-agnostic particle searches at colliders is emerging, which aims to leverage recent breakthroughs in anomaly detection and machine learning. In order to develop and benchmark new anomaly detection methods within this framework, it is essential to have standard datasets. To this end, we have created the LHC Olympics 2020, a community challenge accompanied by a set of simulated collider events. Participants in these Olympics have developed their methods using an R\&D dataset and then tested them on black boxes: datasets with an unknown anomaly (or not). This talk will review the LHC Olympics 2020 challenge, including an overview of the competition, a description of methods deployed in the competition (including those submitted by non-collider physicists), lessons learned from the experience, and implications for data analyses at colliders and beyond.   

Live

Histogram-based template fits are a common technique for parameter estimation in high energy physics and other areas of physics where Monte Carlo event generators are reliable. Parameterized neural network reweighting can be used to extend this fitting procedure to many dimensions and does not require binning. If the fit is to be performed using reconstructed data, then expensive detector simulations must be used for training the neural networks. We introduce a new two-level fitting approach that only requires one dataset with detector simulation and then a set of additional generation-level datasets without detector effects included. This Simulation-level fit based on Reweighting Generator-level events with Neural networks (SRGN) is demonstrated using simulated datasets for a variety of examples including a simple Gaussian random variable, strong force radiation modeling (parton shower tuning), and the top quark mass extraction.~ This presentation is based on~\underline {https://arxiv.org/abs/2010.03569}.   

Live

The use of machine learned classifiers for the measurement of parameters of interest has become ubiquitous in High Energy Physics (HEP) experiments. These models offer significant improvement in sensitivity compared to the traditional cut-based approach by exploiting subtle patterns in the high dimensional feature space, however, this also makes them highly sensitive to systematic uncertainties which lead to differences between the training and application datasets. Contrary to the traditional wisdom of keeping the decision criteria invariant to systematic effects, we study the use of a classifier that is fully aware of the systematic uncertainty in order to provide a better sensitivity to the parameter of interest. Studies are performed on a toy dataset as well as a more realistic HEP dataset, comparing our approach to typical baseline machine learning based approaches.   

Live

To allow a direct comparison with theoretical predictions, experimentally measured particle physics data must be corrected for detector effects, or ``unfolded'' (a process often also called deconvolution). The recently introduced OmniFold method uses machine learning to unfold data at particle-level, thereby simultaneously unfolding all variables and accounting for effects that would be neglected using traditional unfolding schemes. Because the OmniFold method uses the full phase space of the data to perform the unfolding, it should naturally allow for the potential presence of Beyond the Standard Model (BSM) physics without concern over potential distortions that the BSM physics might cause in the truth-to-measured response matrix. In this study we examine the power of the OmniFold method to preserve BSM information and the feasibility of performing searches in fully unfolded data. This approach applies to any application of deconvolution in a complex phase space that models detector response with simulation.   

Live

The discovery of symmetries in physical laws is of both pure and applied interest to various areas of physics. Besides being meaningful in and of itself, symmetry identification in a data set magnifies the statistical power of the data set by reducing its effective dimension. We propose a customised loss function for a modified generative adversarial network (GAN) to allow the neural net to discover non--trivial symmetries in data sets. The loss function and the associated neural net are analysed both analytically and experimentally to study their ability to discover non--trivial symmetries. The loss function we propose is specific yet flexible, and may be adjusted to reflect various symmetry groups by a suitable choice of the mean squared error term.   

Live

We treat an LHC event sample as a spatial point pattern in the relevant phase space of the final state signature. We then demonstrate how methods from spatial statistics and computational geometry can be applied to address classical problems like statistical inference, density estimation and manifold learning in an unsupervised fashion.   

Live

In this work, we will present the implementation templates for two types of recurrent neural network layers within the HLS4ML library – Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). These templates provide the lower-level hardware implementation for a neural network model based on these layers, allowing them to be mapped to an FPGA. Using the HLS4ML library, the latency per inference and resource utilization can be adjusted for the targeted FPGA, and the application requirements. Several particle physics problems are used to characterize the template and test its efficiency after the High-Level Synthesis. Design space exploration was performed across different features - resource utilization, latency, model performance, and fixed-point precision. As an example, LSTM and GRU based models for the task of jet identification on simulated proton-proton collision at the Large Hadron Collider were considered. Also, the implementation templates were evaluated against varying numbers of model parameters, and synthesized for larger neural network models based on LSTM and external recursion for jet flavor classification in high energy collision for an FPGA.   

Live

Estimation of calorimeter hit energy resolution will be a significant challenge in proton-proton collisions during the HL-LHC era. This study is dedicated to a novel neural network approach to identify physics signatures in ATLAS LAr calorimeter to overcome that challenge. The ATLAS Readout Electronics Upgrade Simulation (AREUS) software is used to produce simulation samples, which contain necessary information obtained by optimal filtering techniques for various pile-up values. This information is used for training the architecture to predict the optimal filtering coefficients and thus the shape of a calorimeter hit on the fly. Moreover, we aim to use FPGAs in order to create a low-latency real-time interface that would allow the architecture to run on the LAr calorimeter back-end systems. Implementation of the architecture into a format suitable for deployment on FPGAs is in production.   

Live

The development of data-driven models to describe physical phenomenon requires frameworks that are physically principled and generalizable. We present a data-driven framework for reduced order modeling of continuum physics by harnessing the Mallat Scattering Transform (MST). The MST acts as an analogue to the traditional Convolutional Neural Network framework with predefined, physics-informed weights to aggregate information in a scale-dependent manner while preserving conservation properties. The transform provides a state based representation of the physics and time-independent rate coefficients describing their evolution, which can be determined using a modified version of the Generalized Master Equation. This framework reduces the amount of parameters that must be independently optimized compared to similar data-driven models, while enforcing the necessary conditions like diffeomorphic continuity. The framework also produces descriptors of the dynamics that are translatable to the underlying physical principles governing the behavior of the PDE. We demonstrate the efficacy of the framework by surrogating the behavior of one-dimensional linear and non-linear Partial Differential Equations, such as the Burgers' Equation. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.