UCL - Scalable Collider Data Analysis with Optimised Predictions and Machine-Learning Anomaly Detection
Existing background work
The Large Hadron Collider (LHC) has generated an immense amount of data, which, when combined with advanced Standard Model (SM) predictions, presents new opportunities for discovering physics beyond the SM. Current analysis frameworks, such as Rivet, allow for data-theory comparisons, but incorporating higher-order calculations (NLO, NNLO) into these comparisons remains challenging. Additionally, there is potential to enhance traditional statistical analyses with machine learning-based anomaly detection to identify mismodelling or uncover hints of new physics. This project aims to improve both the affordability and accuracy of SM calculations and apply them to large-scale metadata analysis of collider data.
Main objectives of the project
- Optimising SM Calculations: The project seeks to make state-of-the-art SM predictions computationally affordable and accessible for large-scale collider data analysis. This will be achieved through performance profiling of existing tools and integrating higher-order calculations into Monte Carlo (MC) particle-level predictions using reweighting techniques.
- Comprehensive Metadata Analysis: The student will perform large-scale metadata analysis using Rivet, comparing experimental data from LHC experiments like ATLAS and CMS with optimised SM predictions. P-value distributions will be generated to quantify the level of agreement between theory and data across multiple observables.
- Anomaly Detection Framework: Machine learning algorithms will be applied to p-value distributions to detect anomalies in data-theory comparisons. This will enhance sensitivity in identifying potential Monte Carlo mismodelling or uncovering signals of new physics.
Details of Software/Data Deliverables
- Performance-Optimised Prediction Tools: The student will profile existing SM tools (e.g. Sherpa, Herwig) to identify and address computational bottlenecks, optimising them for large-scale collider data comparisons.
- Integration of Higher-Order Calculations: Incorporation of NLO and NNLO corrections into MC particle-level predictions through reweighting techniques, making these sophisticated calculations accessible for widespread use.
- Data-Theory Comparison Framework: A Rivet-based framework for generating detailed p-value distributions, which will provide a robust statistical foundation for comparing experimental data with SM predictions.
- Machine Learning-Based Anomaly Detection: Development of a machine learning-enhanced system to detect inconsistencies in data-theory comparisons, using unsupervised learning techniques to identify potential new physics signals or MC mismodelling.
