READS will develop ML methods and their edge implementation within large scale accelerator systems. The Main Injector enclosure houses two accelerators; the Main Injector (MI), which is a 120 GeV conventional powered magnet synchrotron and the Recycler (RR), which is an 8 GeV permanent magnet ring. In periods of joint operation, when both machines contain high intensity beam, radiative beam losses from MI and RR overlap on the enclosure’s beam loss monitoring (BLM) system, making it difficult to attribute those losses to a single machine. Incorrect diagnoses result in unnecessary downtime that incurs both financial and experimental cost. Using a continuous adaptation of the popular UNet architecture in conjunction with a novel data augmentation scheme, we developed an ML model which accurately infers the machine of origin on a per-BLM basis in periods of joint and independent operation. Using streamed, distributed BLM readings and real-time ML inference hardware, this project aims to replicate and then improve upon the machine expert’s ability to de-blend, or disentangle, each machines’ contribution to the measured losses.

The project forms part of the Phase-2 Endcap Muon Track Finder (EMTF) Upgrade to be implemented within the Compact Muon Solenoid (CMS) detector of the Large Hadron Collider (LHC) at CERN.

One of the 3 objectives of EMTF Phase-2 is to do pT assignment. This involves using machine learning and deep learning (e.g. BDT, NN) to determine the track pT using all the discriminating variables: Δф, Δθ, bend, η, etc. of different muon tracks.

However, the baseline NN for pT assignment proved to be difficult to implement using a single FPGA (which was desirable). My role in the project was to compress the model to fit on-chip in a way that reduces the size and latency of the NN in inference without comprising much of its performance/ accuracy.

This project aims to demonstrate machine learning using graph neural networks as a game-changing solution for imaging calorimeter data in future high-energy physics (HEP) experiments. The CMS high luminosity upgrade calorimeter (HGCal) and Large Hadron Collider (LHC) will be used as a case study.

My contribution(s) to this project include:

• Building Pytorch equivalent of TensorFlow kernels for use in integration with PyTorch’s torch_geometric library to optimize the present operations (KNN for GNNs), for large datasets and bypass the need to use TF kernels which may add to latency given the need to adapt the TF kernels to work with torch_geometric.

• Integrate and implement custom object condensation loss function for GNN training to enable a single pass, end-to-end reconstruction of particles trajectory in the CMS Phase-2 High-Granularity Calorimeter (HGCAL) and LHC at CERN.

• Assisted in testing the success rate of the GNN to predict clusters of hits originating from the same incident particle in a multi-particle environment.