Deep Learning
Learn how our hybrid deep-learning models identified rare subatomic signatures and advanced national-lab research initiatives.
CHALLENGE
Physicists at a national laboratory needed to identify rare b‑jet signatures to advance nuclear research. There was no existing dataset, and the team lacked machine‑learning expertise. The project required building a dataset from simulations, training a model without GPUs and deploying it on an FPGA.
SOLUTION
We partnered with the simulation team to generate labelled data and visualisations, using AWS SageMaker and XGBoost as a baseline before moving to a CNN‑LSTM hybrid model. We tuned hyperparameters and used L1 regularisation to achieve initial results.
Impact
The prototype achieved roughly 70% accuracy, winning management support and additional resources. It set the stage for more sophisticated deep‑learning models on custom hardware and showcased how data science accelerates fundamental research.
