High-Level Synthesis for Machine Learning (HLS4ML) has created an open-source library to optimize the resources and time needed to run machine learning inference on hardware devices. This library has converted common machine learning algorithms into C files and subsequently into Verilog code that can be uploaded onto these devices. This process works on creating efficient techniques to produce neural networks and convolutional neural networks that can be mapped to Field-Programmable Gate Arrays (FPGAs). Our group seeks to analyze the strengths and weaknesses of this process through exploring the performance, resources, time, and other criteria of models derived from a high-level neural network library, Keras. We have chosen to explore the KERAS1_Layer and KERAS_conv2d models. To evaluate the efficiency of the low level code produced by HLS4ML, we implemented our own SystemVerilog versions that mimic the machine learning algorithms. In doing so, we are able to understand what HLS4ML is doing that is replicable in just Verilog, and what is not. We are currently working on matching HLS functionality, but preliminary results show that our KERAS_1Layer implementation has faster clock speeds while lagging behind in terms of resource usage and latency. This is due to a higher number of pipeline stages and we are working on rectifying this. We expect to see similar results for the KERAS_conv2d model.

Semi-visible Jets may occur from strongly coupled hidden sectors produced at the Large Hadron Collider, as suggested in the Hidden Valley models. While dark hadrons interact strongly with each other, they interact only weakly with visible states through the portal, which will undergo a QCD-like shower and ultimately hadronize, producing collimated sprays of dark hadrons. These states are invisible to colliders’ detectors unless they are able to decay to the Standard Model. A portion of these states are likely to be stable, providing good dark-matter candidates. Yet, many of the hadrons should decay back to the visible sector through the portal coupling, which result in a spray of stable invisible dark matter along with unstable states that decay back to the Standard Model. The signature of such Semi-visible Jets is characterized by the missing energy aligned along the direction of one of the jets. In this research, we generated the Semi-visible Jets s-channel samples using standalone MadGraph5, Pythia8, and Delphes and conducted data analysis using uproot and pyjet package in Python. We analyzed the kinematics of Semi-visible Jets with different parameter settings such as event selection cuts and jet clustering algorithms by creating kinematic plots of physical quantities including jet momentum, invariant mass, transverse mass, and missing energy using Python. Also, we have created the ATLAS JobOption, which has already been used for sample generation in the CERN ATLAS framework. We have noticed differences in kinematic distribution with different event selection and jet clustering algorithms and we expect to find the parameter settings for them that will optimize the Semi-visible Jets signal. By applying optimized parameter settings, we can locate the possible region where Semi-visible Jets can be observed in the Large Hadron Collider, which is a significant step forward in the discovery of dark matter.

Detecting physics beyond the standard model is an important task that utilizes cutting edge new models. In the “Anomaly Detection Data Challenge 2021,” we develop a novel anomaly detection algorithm for the task of finding a priori unknown and rare New Physics data. The challenge uses simulated anomaly detection data that emulates the strict bandwidth, latency and resource constraints of the L1 trigger of the Large Hadron Collider whose dataset is composed of 10 particle jets, 4 muons, 4 electrons, and missing transverse energy. Autoencoders and variational autoencoders are powerful neural network models that are widely used to approximate input distributions and reduce latent dimensions. We reproduce simple autoencoders/variational autoencoders to detect anomalies within the dataset using Mean Squared Error (MSE) loss and Kullback- Leibler divergence (K-L Divergence) as the anomaly metrics respectively. Using these well-performing autoencoder models as a baseline, we develop and test novel, powerful, generative and autoencoder based models for the anomaly detection task.