Real-Time Anomaly Detection for Exotic Higgs Decays Using Decision Trees

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.May 6 2024

In a paper published in the journal Nature Communications, researchers presented an interpretable implementation of an autoencoding anomaly detection algorithm using a forest of deep decision trees (DT) on a field programmable gate array (FPGA) tailored for scenarios at the large hadron collider (LHC) at the European organization for nuclear research (CERN). 

The autoencoder was trained on known standard model processes and deployed in real-time trigger systems for detecting unknown phenomena, such as rare exotic decays of the Higgs boson. This method offered edge artificial intelligence (AI) users efficient anomaly detection capabilities, achieving inference with a 30 ns latency and minimal resource usage.

Background

Previous works utilized LHC to explore physics beyond the standard model (BSM) using AI algorithms. Researchers focused on detecting rare events amidst the numerous standard model processes in the collider's data. Various methods, including neural network-based approaches, were employed to trigger potential real-world BSM events to uncover new physics phenomena.

The drawbacks of utilizing AI algorithms for BSM physics exploration at the LHC include the complexity of analyzing vast amounts of data, leading to computational challenges and increased processing times. Additionally, reliance on AI may introduce biases or limitations in detecting rare events, potentially overlooking significant discoveries. 

Simulation-Based Methodology 

The methods employed in this study involved using simulated samples to assess the performance of the autoencoder in real-time triggers. These samples encompassed proton-proton collision events, including background processes and anomaly processes simulating the production and decay of scalar bosons. 

Various tools such as MadGraph5_aMC, Pythia8, and Delphes were used to simulate proton collisions, decay processes, and detector effects. The input variables for the autoencoder were reconstructed values calculated by Delphes, utilizing algorithms for reconstructing jets and photons. Pileup effects were neglected due to the focus on hadronic jets, with further details provided in the samples.

The firmware design was based on existing structures, utilizing an autoencoder processor for anomaly detection. The design incorporated deep DT engines for encoding and decoding, with modifications to output a vector of values for distance computation. While further modifications for efficient data transmission were possible, they were not within the scope of this study.

Verification and validation procedures were conducted to ensure the accuracy and functionality of the algorithm. It included validation through C simulation in Vivado high-level synthesis (HLS), co-simulation to compare register transfer level (RTL) models against C designs, and physical verification through programming configurations onto the xcvu9p FPGA. These steps confirmed the consistency between simulated outputs and expected results, validating the algorithm's efficacy in practical applications.

Autoencoder Anomaly Detection

The study outlines a DT-based autoencoder's design and training approach, emphasizing its application as an anomaly detector in real-time trigger systems. By benchmarking against conventional methods, such as diphoton triggers, the autoencoder's efficacy in detecting exotic Higgs decays is demonstrated. Furthermore, the study explores its adaptability to signal contamination in training data, showcasing its potential robustness in practical settings.

The autoencoder architecture extends beyond traditional random forest-based approaches, utilizing deep decision tree structures for encoding and decoding input vectors. Through a series of decision paths, input variables are mapped to a latent space, enabling efficient anomaly detection based on distance metrics between input and output.

Extensive simulations using Monte Carlo methods generate training and testing samples to evaluate performance, simulating proton-proton collision events at 13 TeV. These samples encompass a range of SM processes and benchmark signals for exotic Higgs decays, facilitating comprehensive validation of the autoencoder's effectiveness.

Benchmark results illustrate significant improvements in signal acceptance rates compared to conventional trigger methods, particularly evident in scenarios with exotic Higgs decay signals. The autoencoder's performance is quantified regarding acceptance rates and receiver operating characteristic (ROC) curves, highlighting its superior performance in detecting rare BSM signals.

Moreover, the study investigates the autoencoder's resilience to signal contamination in training data, a crucial consideration in real-world applications. By training models with varying levels of signal contamination, the study demonstrates the autoencoder's ability to maintain performance even in scenarios with up to 33% signal contamination.

The study underscores the potential of decision tree-based autoencoders as powerful tools for anomaly detection in high-energy physics experiments. Their adaptability, as demonstrated through benchmarking and signal-contaminated training scenarios, suggests promising prospects for their integration into future trigger systems.

Conclusion

In summary, implementing a DT-based autoencoder for anomaly detection in high-energy physics experiments showed promise, offering efficient real-time detection of exotic signals comparable to neural network-based methods. The study explored training machine learning (ML) models on collected data, presenting opportunities for detecting signals beyond the standard model.

However, challenges remained in accommodating diverse input channels and accurately mapping input space to anomaly scores for identifying BSM events within anomalous samples. Rigorous statistical treatment was crucial for precise BSM composition extraction. Continued research on decision tree-based anomaly detectors was essential for advancing real-time anomaly detection capabilities in high-energy physics experiments.