New Autoencoder Spots Unusual Patterns Within Particle Collider Jets
Summarize this article with:
Scientists are exploring novel machine learning techniques to enhance the detection of rare and unusual events within the vast datasets generated by particle colliders. Emre Gurkanli and Michael Spannowsky present a multiscale tensor-network autoencoder, drawing inspiration from the MERA (Multiscale Entanglement Renormalisation Ansatz) structure, to improve anomaly detection in “jets”. Jets are collimated sprays of particles created during high-energy collisions, and identifying deviations from expected jet characteristics is crucial for discovering new physics. The application of a MERA-inspired autoencoder to collider anomaly detection represents a first, and the hierarchical compression it facilitates, aligned with the natural structure of jet data, demonstrably enhances performance, particularly when data compression is substantial. The findings support locality-aware hierarchical compression as a key set of tools for uncovering new physics at colliders. Multiscale autoencoders enhance rare signal detection in high-energy physics A MERA-inspired autoencoder significantly improves jet anomaly detection, achieving a sharp reduction in reconstruction error, a measure of how well the autoencoder can recreate the input data, compared to traditional dense autoencoders when compression is strong. Standard autoencoder architectures often struggle with highly compressed data, leading to a loss of information and reduced sensitivity to subtle anomalies that could indicate new physics. The architecture’s multiscale, locality-preserving structure mirrors the natural organisation of jets, which are formed through a branching cascade of particle production, offering a more efficient and physically motivated representation of complex data. This branching process occurs across a range of angular and momentum scales, meaning that the internal structure of a jet is inherently hierarchical. This enhanced performance unlocks the potential to identify subtle anomalies previously obscured by background fluctuations. Locality-aware hierarchical compression, enabled by concepts from quantum physics, specifically the MERA tensor network, serves as a valuable inductive bias for uncovering rare signals within collider experiments. An inductive bias is a set of assumptions built into a machine learning model to guide its learning process, and in this case, the bias reflects the known physical properties of jet formation. The autoencoder demonstrably improves performance on a benchmark dataset of simulated high-energy particle collisions, allowing for a more precise separation of signal from background noise. Crucially, a training-free diagnostic, involving the analysis of the autoencoder’s learned representations, revealed alignment between the model’s internal structure and the physical formation of jets through branching cascades of energy. This suggests the model isn’t simply memorising the training data, but is instead learning to represent the underlying physics. Further analysis, employing an ‘identity-disentangler ablation’ technique to systematically remove components of the autoencoder, showed that the MERA disentanglers, the specific layers inspired by MERA, enhance performance when compression is at its most intense. Disentangling refers to the process of separating different underlying factors of variation in the data, and the ablation study confirmed that these disentanglers play a critical role in preserving information during strong compression. This suggests a refined understanding of the model’s internal mechanisms and validates the effectiveness of the MERA-inspired design. These results currently focus on simulated data, generated using established Monte Carlo event generators, and do not yet show strong performance on real data from the Large Hadron Collider. The transition to real data is complicated by detector effects, the imperfections and limitations of the detector hardware, and more complex backgrounds, arising from multiple simultaneous collisions and other sources of noise. Translating this success to real-world collider data is acknowledged as difficult, given the extraordinarily complex events and substantial noise generated by the Large Hadron Collider, which requires robust and sophisticated data processing techniques. Nevertheless, demonstrating that a quantum-inspired approach can effectively compress and reorganise information within these events represents a valuable step forward. The autoencoder offers a new approach to identifying unusual events within particle collider data, reorganising complex information in a manner mirroring the natural branching structure of jets. The MERA structure allows the autoencoder to efficiently represent the data at multiple scales, capturing both short-range correlations (the relationships between nearby particles) and long-range correlations (the relationships between particles further apart). By focusing on locality and hierarchical compression, the architecture improves anomaly detection, particularly when data is heavily compressed, highlighting the value of incorporating prior knowledge into machine learning models used for particle physics. This prior knowledge, derived from the understanding of jet physics, helps the model to generalise better and avoid overfitting to the training data. This provides a new set of tools for anomaly detection, potentially enhancing the search for new physics beyond established models, such as supersymmetry or extra dimensions. The ability to efficiently and accurately identify anomalies is crucial for pushing the boundaries of our understanding of the fundamental laws of nature and could lead to groundbreaking discoveries in the field of particle physics. The research demonstrated that a multiscale autoencoder, inspired by MERA tensor networks, effectively compresses and reorganises information within simulated particle collider data. This architecture mirrors the natural hierarchical structure of jets, improving anomaly detection, especially when data compression is significant. The model’s ability to capture both short and long-range correlations within jets represents a valuable step towards identifying unusual events. Researchers combined benchmark comparisons with a local-compressibility diagnostic to validate the benefits of this locality-preserving, multiscale structure. 👉 More information🗞 Quantum-Inspired Tensor Network Autoencoders for Anomaly Detection: A MERA-Based Approach🧠 ArXiv: https://arxiv.org/abs/2604.06541 Tags: