AI for High-Energy Physics

Speeding deployment of AI models in high-energy physics

AI can bring many benefits for studies in high-energy physics, but training AI models in this space takes a great deal of time and data. This project, titled “Designing Efficient Edge AI with Physics Phenomena” aims to speed the deployment and efficiency of AI models by embedding physics knowledge into the machine learning process.

To study high-energy physics, scientists rely on experimental datasets that are highly dimensional and complex. The next generation of studies in this field will require ML to be deployed in real-time with stringent constraints in terms of computational speed, resource usage, and energy consumption while retaining accuracy.

One way to improve accuracy and learn more efficiently is to embed domain-specific knowledge into the learning process. To achieve this, our team is developing methods that provide strong control on how to understand, propagate, and track constraints and associated uncertainties for edge ML models. Specifically, our work explores the connection between efficient ML and “physics-inspired” neural networks (PINNs), where the highest efficiency can be achieved by reducing the training dataset, the number of model parameters, and (bitwise) operations for a given NN, where the latter can be controlled through sparsity and quantization techniques.

To develop and demonstrate these methods, we are focusing on experiments at the CERN Large Hadron Collider, neutrino experiments based on Liquid Argon Time Projection Chamber technology, and super-conducting magnet technology applications.

Project Team

About