A5031270964- Particle physics theoretical and experimental studies
- High-Energy Particle Collisions Research
- Quantum Chromodynamics and Particle Interactions
- Particle Detector Development and Performance
- Computational Physics and Python Applications
- Dark Matter and Cosmic Phenomena
- Cosmology and Gravitation Theories
- Neutrino Physics Research
- Medical Imaging Techniques and Applications
- Astrophysics and Cosmic Phenomena
- Radiation Detection and Scintillator Technologies
- Anomaly Detection Techniques and Applications
- Distributed and Parallel Computing Systems
- Particle Accelerators and Free-Electron Lasers
- Gaussian Processes and Bayesian Inference
- Algorithms and Data Compression
- Soil Moisture and Remote Sensing
- Radio Astronomy Observations and Technology
- Diamond and Carbon-based Materials Research
- Generative Adversarial Networks and Image Synthesis
- Black Holes and Theoretical Physics
- Particle accelerators and beam dynamics
- Gamma-ray bursts and supernovae
- Big Data Technologies and Applications
- International Science and Diplomacy
Universität Hamburg
2018-2025
École Polytechnique
2024
University of Iowa
2024
The University of Tokyo
2024
Centre National de la Recherche Scientifique
2024
Laboratoire Leprince-Ringuet
2024
Northern Illinois University
2024
University of Göttingen
2024
Karlsruhe Institute of Technology
2024
Institut National de Physique Nucléaire et de Physique des Particules
2024
Autoencoder networks, trained only on QCD jets, can be used to search for anomalies in jet-substructure. We show how, based either images or 4-vectors, they identify jets from decays of arbitrary heavy resonances. To control the backgrounds and underlying systematics we de-correlate jet mass using an adversarial network. Such autoencoder allows a general at same time easily controllable new physics. Ideally, it applied data phase space region, allowing us efficiently physics un-supervised learning.
Accurate simulation of physical processes is crucial for the success modern particle physics. However, simulating development and interaction showers with calorimeter detectors a time consuming process drives computing needs large experiments at LHC future colliders. Recently, generative machine learning models based on deep neural networks have shown promise in speeding up this task by several orders magnitude. We investigate use new architecture -- Bounded Information Bottleneck...
A critical question concerning generative networks applied to event generation in particle physics is if the generated events add statistical precision beyond training sample. We show for a simple example with increasing dimensionality how indeed amplify statistics. quantify their impact through an amplification factor or equivalent numbers of sampled events.
We propose a new model-agnostic search strategy for physics beyond the standard model (BSM) at LHC, based on novel application of neural density estimation to anomaly detection. Our approach, which we call Classifying Anomalies THrough Outer Density Estimation (CATHODE), assumes BSM signal is localized in region (defined e.g. using invariant mass). By training conditional estimator collection additional features outside region, interpolating it into and sampling from it, produce events that...
First-principle simulations are at the heart of high-energy physics research program. They link vast data output multi-purpose detectors with fundamental theory predictions and interpretation. This review illustrates a wide range applications modern machine learning to event generation simulation-based inference, including conceptional developments driven by specific requirements particle physics. New ideas tools developed interface will improve speed precision forward simulations, handle...
With the vast data-collecting capabilities of current and future high-energy collider experiments, there is an increasing demand for computationally efficient simulations. Generative machine learning models enable fast event generation, yet so far these approaches are largely constrained to fixed data structures rigid detector geometries. In this paper, we introduce EPiC-GAN - equivariant point cloud generative adversarial network which can produce clouds variable multiplicity. This flexible...
Abstract Machine learning-based anomaly detection (AD) methods are promising tools for extending the coverage of searches physics beyond Standard Model (BSM). One class AD that has received significant attention is resonant detection, where BSM assumed to be localized in at least one known variable. While there have been many proposed identify such a signal make use simulated or detected data different ways, not yet study methods’ complementarity. To this end, we address two questions....
Abstract Fast simulation of the energy depositions in high-granular detectors is needed for future collider experiments at ever-increasing luminosities. Generative machine learning (ML) models have been shown to speed up and augment traditional chain physics analysis. However, majority previous efforts were limited relying on fixed, regular detector readout geometries. A major advancement recently introduced CaloClouds model, a geometry-independent diffusion which generates calorimeter...
Physics beyond the Standard Model that is resonant in one or more dimensions has been a longstanding focus of countless searches at colliders and beyond. Recently, many new strategies for have developed, where sideband information can be used conjunction with modern machine learning, order to generate synthetic datasets representing background. Until now, this approach was only able accommodate relatively small number dimensions, limiting breadth search sensitivity. Using recent innovations...
Machine learning based on convolutional neural networks can be used to study jet images from the LHC. Top tagging in fat jets offers a well-defined framework establish our DeepTop approach and compare its performance QCD-based top taggers. We first optimize network architecture identify quarks Monte Carlo simulations of Standard Model production channel. Using standard we then multivariate tagger. find that both approaches lead comparable performance, establishing as promising new for...
For simulations where the forward and inverse directions have a physics meaning, invertible neural networks are especially useful. A conditional INN can invert detector simulation in terms of high-level observables, specifically for ZW production at LHC. It allows per-event statistical interpretation. Next, we allow variable number QCD jets. We unfold effects radiation to pre-defined hard process, again with probabilistic interpretation over parton-level phase space.
LHC analyses directly comparing data and simulated events bear the danger of using first-principle predictions only as a black-box part event simulation. We show how simulations, for instance, detector effects can instead be inverted generative networks. This allows us to reconstruct parton level information from measured events. Our results illustrate how, in general, fully conditional networks statistically invert Monte Carlo simulations. As technical by-product we maximum mean discrepancy...
Based on the established task of identifying boosted, hadronically decaying top quarks, we compare a wide range modern machine learning approaches. Unlike most methods they rely low-level input, for instance calorimeter output. While their network architectures are vastly different, performance is comparatively similar. In general, find that these new approaches extremely powerful and great fun.
We introduce a new and highly efficient tagger for hadronically decaying top quarks, based on deep neural network working with Lorentz vectors the Minkowski metric. With its novel machine learning setup architecture it allows us to identify boosted quarks not only from calorimeter towers, but also including tracking information. show how performance of our compares QCD-inspired image-recognition approaches find that significantly increases strongly quarks.
Bayesian neural networks allow us to keep track of uncertainties, for example in top tagging, by learning a tagger output together with an error band. We illustrate the main features versions established deep-learning taggers. show how they capture statistical uncertainties from finite training samples, systematics related jet energy scale, and stability issues through pile-up. Altogether, offer many new handles understand control deep at LHC without introducing visible prior effect...
Given the increasing data collection capabilities and limited computing resources of future collider experiments, interest in using generative neural networks for fast simulation events is growing. In our previous study, Bounded Information Bottleneck Autoencoder (BIB-AE) architecture generating photon showers a high-granularity calorimeter showed high accuracy modeling various global differential shower distributions. this work, we investigate how BIB-AE encodes physics information its...
Abstract Motivated by the computational limitations of simulating interactions particles in highly-granular detectors, there exists a concerted effort to build fast and exact machine-learning-based shower simulators. This work reports progress on two important fronts. First, previously investigated Wasserstein generative adversarial network bounded information bottleneck autoencoder models are improved successful learning hadronic showers initiated charged pions segment calorimeter...
Abstract Motivated by the high computational costs of classical simulations, machine-learned generative models can be extremely useful in particle physics and elsewhere. They become especially attractive when surrogate efficiently learn underlying distribution, such that a generated sample outperforms training limited size. This kind GANplification has been observed for simple Gaussian models. We show same effect simulation, specifically photon showers an electromagnetic calorimeter.
Abstract We explore the use of normalizing flows to emulate Monte Carlo detector simulations photon showers in a high-granularity electromagnetic calorimeter prototype for International Large Detector (ILD). Our proposed method — which we refer as “Layer-to-Layer Flows” ( L2LFlows ) is an evolution CaloFlow architecture adapted higher-dimensional setting (30 layers 10× 10 voxels each). The main innovation consists introducing 30 separate flows, one each layer calorimeter, where flow...
We introduce a new technique named Latent CATHODE (LaCATHODE) for performing "enhanced bump hunts", type of resonant anomaly search that combines conventional one-dimensional hunts with model-agnostic score in an auxiliary feature space where potential signals could also be localized. The main advantage LaCATHODE over existing methods is it provides well behaved when evaluating beyond the signal region, which essential to prevent sculpting background distributions hunt. accomplishes this by...
Abstract The demands placed on computational resources by the simulation requirements of high energy physics experiments motivate development novel tools. Machine learning based generative models offer a solution that is both fast and accurate. In this work we extend Bounded Information Bottleneck Autoencoder (BIB-AE) architecture, designed for particle showers in highly granular calorimeters, two key directions. First, generalise model to multi-parameter conditioning scenario, while...
Choosing which properties of the data to use as input multivariate decision algorithms—also known feature selection—is an important step in solving any problem with machine learning. While there is a clear trend towards training sophisticated deep networks on large numbers relatively unprocessed inputs (so-called automated engineering), for many tasks physics, sets theoretically well-motivated and well-understood features already exist. Working such can bring benefits, including greater...
Weakly supervised methods have emerged as a powerful tool for model-agnostic anomaly detection at the Large Hadron Collider (LHC). While these shown remarkable performance on specific signatures such dijet resonances, their application in more manner requires dealing with larger number of potentially noisy input features. In this paper, we show that using boosted decision trees classifiers weakly gives superior compared to deep neural networks. Boosted are well known effectiveness tabular...