Machine Learning in Particle Physics

High-energy physics (HEP) is characterized by large amounts of high-dimensional data, collected for example with the experiments at the Large Hadron Collider (LHC) at CERN in Switzerland or simulated according to theoretical predictions based on quantum field theory.

These simulations follow a chain of several steps, starting with the high-energy process of interest, subsequently decaying heavy, unstable particles (sometimes via a parton shower), and hadronizing colored particles into colorless observable particles. Then, the interaction of all produced particles with the detector material is simulated, resulting in data format that is identical to the signatures recorded in experiment.

For analysis, the detector signatures, so-called hits, are first grouped in either clusters (inside the calorimeter system) or tracks (in the tracker system). These are then merged into particle candidates, which are used to select events of interest for analyses.

Simulation Chain. Image from SciPost Phys. 14 (2023) 4, 079 and Ramon Winterhalder.

From detector signatures to analyses. Image based on Comput.Softw.Big Sci. 7 (2023) 1, 1 and Claudius Krause.

Modern Machine Learning (ML) brought a lot of new ideas and improvements to HEP in recent years. On the experimental side, this includes data collection, particle reconstruction and selection as well as subsequent analysis. On the theoretical side, ML contributed a lot to better and faster simulation and parameter inference. Overall, ML has not only led to significant improvements of existing algorithms, but also to new ideas for previously considered impossible-to-solve problems.

ML has a broad range of applicability in HEP, almost all different types of objectives (like regression, classification, generation) on various different types of data (like tabular, point clouds, graphs) are being explored, as can also be seen in the HEP-ML Living Review.

Machine Learning Research at MBI

At MBI, we are investigating how ML can improve individual steps of the simulation chain above, both in 'forward' and 'inverse' direction. So far, we have focused on two main directions: faster simulations with deep generative models and anomaly detection in search for new physics.

Deep Generative Models in particle simulations

Deep generative models have gotten a lot of attention in recent years through large language models and image generators. In HEP, they can be used as surrogate models, speeding up otherwise slow simulation codes. The prime application of this idea is the simulation of particle showers in the calorimeter segments of the detectors. A simulation based on first principles using the GEANT4 library is slow, as it tracks all particles and their secondaries through the entire detector volumen. A deep generative model, however, learns to sample energies deposited in the detector and samples multiple showers in parallel, speeding up the shower generation by up to four orders of magnitude.

In contrast to the generation of natural images (photos or art), where we prefer highly-realistic single samples, we are more interested in correctly reproducing the entire distribution of showers accross phase space (i.e. in all energy ranges). These different needs require therefore different architectures compared to image generators and also specialized evaluation metrics, which are taylored to HEP. The HEP ML group at MBI is currently coordinating and evaluating the Fast Calorimeter Simulation Challenge, which aims to find the best-suited generative architecture for fast, faithful, and light-weight calorimeter shower generation. These results provide the architecture that the next generation of detector fast simulation could be based on. Recently, we also published a Review on Deep Generative Models for Detector Simulation.

Another application of deep generative models, especially architectures based on normalizing flows, in HEP is for improved efficiency in importance sampling. Importance sampling is a numerical tool that improves Monte Carlo estimates for complicated integrals and it is at the core of all event generators like Sherpa or MadGraph.

Anomaly Detection in experimental data

Physics beyond the Standard Model could hide in current experimental data, and our dedicated searches could be blind to it if we focus on the wrong signatures. With anomaly detection we are ameliorating this by trying to identify the needle in the haystack, without actually knowing how a needle looks like. At the core of many ML methods in anomaly detection lies the observation that a classifier, trained to distinguish two mixed sets of different signal/background composition is automatically also optimal to distinguish signal from background. This is known as "Classification WithOut LAbels (CWoLa)". Conditional deep generative models can now be used to construct, in a data-driven manner, a signal-depleted data sample that can be used in training a CWoLa anomaly detector. We called this method "Classifying Anomalies THrough Outer Density Estimation (CATHODE)" and recently, the CMS collaboration used in a search for new physics.