Giles Strong

A bstract The unsupervised search for overdense regions in high-dimensional feature spaces, where locally high population densities may be associated with anomalous contaminations to an otherwise more uniform population, is of relevance to applications ranging from fundamental research to industrial use cases. Motivated by the specific needs of searches for new phenomena in particle collisions, we propose a novel approach that targets signals of interest populating compact regions of the feature space. The method consists in a systematic scan of subspaces of a standardized copula of the feature space, where the minimum p -value of a hypothesis test of local uniformity is sought by greedy descent. We characterize the performance of the proposed algorithm and show its effectiveness in several experimental situations.

The full optimization of the design and operation of instruments whose functioning relies on the interaction of radiation with matter is a super-human task, due to the large dimensionality of the space of possible choices for geometry, detection technology, materials, data-acquisition, and information-extraction techniques, and the interdependence of the related parameters. On the other hand, massive potential gains in performance over standard, "experience-driven" layouts are in principle within our reach if an objective function fully aligned with the final goals of the instrument is maximized through a systematic search of the configuration space. The stochastic nature of the involved quantum processes make the modeling of these systems an intractable problem from a classical statistics point of view, yet the construction of a fully differentiable pipeline and the use of deep learning techniques may allow the simultaneous optimization of all design parameters. In this white paper, we lay down our plans for the design of a modular and versatile modeling tool for the end-to-end optimization of complex instruments for particle physics experiments as well as industrial and medical applications that share the detection of radiation as their basic ingredient. We consider a selected set of use cases to highlight the specific needs of different applications.

The design of instruments that rely on the interaction of radiation with matter for their operation is a quite complex task if our goal is to achieve near optimality on some well-defined utility fu...

Abstract The performance demands of future particle-physics experiments investigating the high-energy frontier pose a number of new challenges, forcing us to find improved solutions for the detection, identification, and measurement of final-state particles in subnuclear collisions. One such challenge is the precise measurement of muon momentum at very high energy, where an estimate of the curvature provided by conceivable magnetic fields in realistic detectors proves insufficient for achieving good momentum resolution when detecting, e.g., a narrow, high mass resonance decaying to a muon pair. In this work we study the feasibility of an entirely new avenue for the measurement of the energy of muons based on their radiative losses in a dense, finely segmented calorimeter. This is made possible by exploiting spatial information of the clusters of energy from radiated photons in a regression task. The use of a task-specific deep learning architecture based on convolutional layers allows us to treat the problem as one akin to image reconstruction, where images are constituted by the pattern of energy released in successive layers of the calorimeter. A measurement of muon energy with better than 20% relative resolution is shown to be achievable for ultra-TeV muons.

The full optimization of the design and operation of instruments whose functioning relies on the interaction of radiation with matter is a super-human task, given the large dimensionality of the space of possible choices for geometry, detection technology, materials, data-acquisition, and information-extraction techniques, and the interdependence of the related parameters. On the other hand, massive potential gains in performance over standard, "experience-driven" layouts are in principle within our reach if an objective function fully aligned with the final goals of the instrument is maximized by means of a systematic search of the configuration space. The stochastic nature of the involved quantum processes make the modeling of these systems an intractable problem from a classical statistics point of view, yet the construction of a fully differentiable pipeline and the use of deep learning techniques may allow the simultaneous optimization of all design parameters. In this document we lay down our plans for the design of a modular and versatile modeling tool for the end-to-end optimization of complex instruments for particle physics experiments as well as industrial and medical applications that share the detection of radiation as their basic ingredient. We consider a selected set of use cases to highlight the specific needs of different applications.

Beginning from a basic neural-network architecture, we test the potential benefits offered by a range of advanced techniques for machine learning, in particular deep learning, in the context of a typical classification problem encountered in the domain of high-energy physics, using a well-studied dataset: the 2014 Higgs ML Kaggle dataset. The advantages are evaluated in terms of both performance metrics and the time required to train and apply the resulting models. Techniques examined include domain-specific data-augmentation, learning rate and momentum scheduling, (advanced) ensembling in both model-space and weight-space, and alternative architectures and connection methods. Following the investigation, we arrive at a model which achieves equal performance to the winning solution of the original Kaggle challenge, whilst being significantly quicker to train and apply, and being suitable for use with both GPU and CPU hardware setups. These reductions in timing and hardware requirements potentially allow the use of more powerful algorithms in HEP analyses, where models must be retrained frequently, sometimes at short notice, by small groups of researchers with limited hardware resources. Additionally, a new wrapper library for PyTorch called LUMIN is presented, which incorporates all of the techniques studied.

Between the years 2015 and 2019, members of the Horizon 2020-funded Innovative Training Network named "AMVA4NewPhysics" studied the customization and application of advanced multivariate analysis methods and statistical learning tools to high-energy physics problems, as well as developed entirely new ones. Many of those methods were successfully used to improve the sensitivity of data analyses performed by the ATLAS and CMS experiments at the CERN Large Hadron Collider; several others, still in the testing phase, promise to further improve the precision of measurements of fundamental physics parameters and the reach of searches for new phenomena. In this paper, the most relevant new tools, among those studied and developed, are presented along with the evaluation of their performances.

In recent years novel inference techniques have been developed based on the construction of non-linear summary statistics with neural networks by minimising inferencemotivated losses. One such technique is inferno (P. de Castro and T. Dorigo, Comp. Phys. Comm. 244 (2019) 170) which was shown on toy problems to outperform classical summary statistics for the problem of confidence interval estimation in the presence of nuisance parameters. In order to test and benchmark the algorithm in a real world application, a full, systematics-dominated analysis produced by the CMS experiment, "Measurement of the top-antitop production cross section in the tau+jets channel in pp collisions at sqrt(s) = 7 TeV" (CMS Collaboration, The European Physical Journal C, 2013) is reproduced with CMS Open Data. The application of the inferno-powered neural network architecture to this analysis demonstrates the potential to reduce the impact of systematic uncertainties in real LHC analyses. This work also exemplifies the extent to which LHC analyses can be reproduced with open data.

We describe a software package, TomOpt, developed to optimise the geometrical layout and specifications of detectors designed for tomography by scattering of cosmic-ray muons. The software exploits differentiable programming for the modeling of muon interactions with detectors and scanned volumes, the inference of volume properties, and the optimisation cycle performing the loss minimisation. In doing so, we provide the first demonstration of end-to-end-differentiable and inference-aware optimisation of particle physics instruments. We study the performance of the software on a relevant benchmark scenarios and discuss its potential applications.

In this article we examine recent developments in the research area concerning the creation of end-to-end models for the complete optimization of measuring instruments. The models we consider rely on differentiable programming methods and on the specification of a software pipeline including all factors impacting performance -- from the data-generating processes to their reconstruction and the extraction of inference on the parameters of interest of a measuring instrument -- along with the careful specification of a utility function well aligned with the end goals of the experiment. Building on previous studies originated within the MODE Collaboration, we focus specifically on applications involving instruments for particle physics experimentation, as well as industrial and medical applications that share the detection of radiation as their data-generating mechanism.

Multi-slice CT system refers to the special CT system equipped with a multi-row detector array to simultaneously acquire data at different z locations. A 4-slice CT scanner has been developed. Here, the authors outline the principles of multi-slice helical CT in general and 4-slice helical CT in particular. The multi-slice helical CT has three key components: the preferred helical pitch for the efficient z sampling in data collection; the helical interpolation algorithms to correct for the fast simultaneous patient translation; and the z-filtration reconstruction for further control of the slice profile, image noise and artifacts in reconstruction. Preferred helical pitches for 4-slice CT are chosen for efficient z sampling and better cone-beam artifacts control. The helical reconstruction algorithms are proposed for multi-slice CT in general and for 4-slice helical CT in particular. The z-filtration reconstruction is developed and studied. Slice profiles, noises and artifacts of 4-slice helical CT are studied analytically and experimentally and compared with single slice helical CT. It is concluded that the 4-slice helical CT can provide equivalent image quality with improved z-axis resolution at 2 to 3 times of the volume coverage rate and at much lower mA of the single slice helical CT.

This record contains the fully-preprocessed training/validation and testing datasets used to train and evaluate the final models for "Calorimetric Measurement of Multi-TeV Muons via Deep Regression" by Jan Kieseler, Giles C. Strong, Filippo Chiandotto, Tommaso Dorigo, & Lukas Layer, (2021), arXiv:2107.02119 [physics.ins-det] (https://arxiv.org/abs/2107.02119). The files are LZF-compressed HDF5 format and designed to be used directly with the code-base available at https://github.com/GilesStrong/calo_muon_regression. Please use the 'issues' tab on the GitHub repo for any questions or problems with these datasets. The training dataset consists of 886,716 muons with energies in the continuous range [50,8000] GeV split into 36 subsamples (folds). The zeroth fold of this dataset is used as our validation data. The testing dataset contains 429,750 muons, generated at fixed values of muon energy (E=100, 500, 900, 1300, 1700, 2100, 2500, 2900, 3300, 3700, 4100 GeV), and split into 18 folds. The input features are the raw hits in the calorimeter (stored in a sparse COO representation), and the high-level features discussed in the paper.

Gluon splitting to b-quark pairs is shown to be an important process which is currently not well modelled by Monte Carlo event generators. In order to understand the splitting process in Monte Carlo generation, Pythia 8 and Sherpa are compared in terms of their production methods of b-quark pairs, and the parton shower approximation of gluon splitting is compared to its matrix element calculation using Sherpa. The parton shower in Sherpa shows reasonable agreement with the matrix element, but large differences are seen between the predictions of Pythia 8 and Sherpa, particularly in their b-quark pair production through the parton shower. In order to compare the generators to data, a non-prompt di-J/psi sample is prepared via a four-dimensional simultaneous binned-fit to data collected by the ATLAS detector in the Large Hadron Collider at CERN. The Monte Carlo result distributions are then smeared based on B->J/psi correlation functions derived from the analysis of a Pythia 8B J/psi sample. The generator and collider samples are then compared in terms of the distribution of the angular separations between J/psi pairs. It is found that of the two generators: Pythia 8 shows the best agreement with the data, and Sherpa primarily disagrees in the region of low angular-separation, where parton shower production is seen to dominate.

<p> </p><p>We propose to employ differentiable programming techniques in order to construct a modular pipeline that models all the aspects of a muon tomography task, from the generation and interaction of cosmic ray muons with a parameterized detector and passive material, to the inference on the atomic number of the passive volume.</p><p>This enables the optimization of the detector parameters via gradient descent, to suggest optimal detector configurations, geometries, and specifications, subject to external constraints such as cost, detector size, and exposure time.</p><p>The eventual aim is to release the package open-source, to be used to guide the design of futur detectors for muon scattering and absorption imaging.</p>

Within the context of studies for novel measurement solutions for future particle physics experiments, we developed a performant kNN-based regressor to infer the energy of highly-relativistic muons from the pattern of their radiation losses in a dense and granular calorimeter. The regressor is based on a pool of weak kNN learners, which learn by adapting weights and biases to each training event through stochastic gradient descent. The effective number of parameters optimized by the procedure is in the 60 millions range, thus comparable to that of large deep learning architectures. We test the performance of the regressor on the considered application by comparing it to that of several machine learning algorithms, showing comparable accuracy to that achieved by boosted decision trees and neural networks.

The full optimization of the design and operation of instruments whose functioning relies on the interaction of radiation with matter is a super-human task, given the large dimensionality of the space of possible choices for geometry, detection technology, materials, data-acquisition, and information-extraction techniques, and the interdependence of the related parameters. On the other hand, massive potential gains in performance over standard, "experience-driven" layouts are in principle within our reach if an objective function fully aligned with the final goals of the instrument is maximized by means of a systematic search of the configuration space. The stochastic nature of the involved quantum processes make the modeling of these systems an intractable problem from a classical statistics point of view, yet the construction of a fully differentiable pipeline and the use of deep learning techniques may allow the simultaneous optimization of all design parameters. In this document we lay down our plans for the design of a modular and versatile modeling tool for the end-to-end optimization of complex instruments for particle physics experiments as well as industrial and medical applications that share the detection of radiation as their basic ingredient. We consider a selected set of use cases to highlight the specific needs of different applications.

The second workshop on the HEP Analysis Ecosystem took place 23-25 May 2022 at IJCLab in Orsay, to look at progress and continuing challenges in scaling up HEP analysis to meet the needs of HL-LHC and DUNE, as well as the very pressing needs of LHC Run 3 analysis. The workshop was themed around six particular topics, which were felt to capture key questions, opportunities and challenges. Each topic arranged a plenary session introduction, often with speakers summarising the state-of-the art and the next steps for analysis. This was then followed by parallel sessions, which were much more discussion focused, and where attendees could grapple with the challenges and propose solutions that could be tried. Where there was significant overlap between topics, a joint discussion between them was arranged. In the weeks following the workshop the session conveners wrote this document, which is a summary of the main discussions, the key points raised and the conclusions and outcomes. The document was circulated amongst the participants for comments before being finalised here.

Data is in HDF5 format (with LZF compression). For specifics and details, please see https://github.com/GilesStrong/mode_diffprog_22_challenge, N.B. Link active after 01/08/22 The training file contains two datasets: `'x0'`: a set of voxelwise X0 predictions (float32) `'targs'`: a set of voxelwise classes (int): 0 = soil 1 = wall The format of the datasets is a rank-4 array, with dimensions corresponding to (samples, z position, x position, y position). All passive volumes are of the same size: 10x10x10 m, with cubic voxels of size 1x1x1 m, i.e. every passive volume contains 1000 voxels. The arrays are ordered such that zeroth z layer is the bottom layer of the passive volume, and the ninth layer is the top layer. It can be read using e.g. the code below: <em>with open('train.h5') as h5:</em> <em> inputs = h5['x0'][()]</em> <em> targets = h5['targs'][()]</em> The test file only contains the X0 inputs: <em>with open('test.h5') as h5:</em> <em> inputs = h5['x0'][()]</em>

Data is in HDF5 format (with LZF compression). For specifics and details, please see https://github.com/GilesStrong/mode_diffprog_22_challenge The training file contains two datasets: `'x0'`: a set of voxelwise X0 predictions (float32) `'targs'`: a set of voxelwise classes (int): 0 = soil 1 = wall The format of the datasets is a rank-4 array, with dimensions corresponding to (samples, z position, x position, y position). All passive volumes are of the same size: 10x10x10 m, with cubic voxels of size 1x1x1 m, i.e. every passive volume contains 1000 voxels. The arrays are ordered such that zeroth z layer is the bottom layer of the passive volume, and the ninth layer is the top layer. It can be read using e.g. the code below: <em>with h5py.File('train.h5', 'r') as f:</em> <em> inputs = h5['x0'][()]</em> <em> targets = h5['targs'][()]</em> The test file only contains the X0 inputs: <em>with h5py.File('test.h5', 'r') as h5:</em> <em> inputs = h5['x0'][()]</em> The private testing sample also contains targets. The private and public splits can be recovered using: <em>from sklearn.model_selection import train_test_split</em> <em>pub, pri = train_test_split(targets, test_size=25000, random_state=3452, shuffle=True)</em>

Data is in HDF5 format (with LZF compression). For specifics and details, please see https://github.com/GilesStrong/mode_diffprog_22_challenge The training file contains two datasets: `'x0'`: a set of voxelwise X0 predictions (float32) `'targs'`: a set of voxelwise classes (int): 0 = soil 1 = wall The format of the datasets is a rank-4 array, with dimensions corresponding to (samples, z position, x position, y position). All passive volumes are of the same size: 10x10x10 m, with cubic voxels of size 1x1x1 m, i.e. every passive volume contains 1000 voxels. The arrays are ordered such that zeroth z layer is the bottom layer of the passive volume, and the ninth layer is the top layer. It can be read using e.g. the code below: <em>with h5py.File('train.h5', 'r') as f:</em> <em> inputs = h5['x0'][()]</em> <em> targets = h5['targs'][()]</em> The test file only contains the X0 inputs: <em>with h5py.File('test.h5', 'r') as h5:</em> <em> inputs = h5['x0'][()]</em> The private testing sample also contains targets. The private and public splits can be recovered using: <em>from sklearn.model_selection import train_test_split</em> <em>pub, pri = train_test_split(targets, test_size=25000, random_state=3452, shuffle=True)</em>

Beginning from a basic neural-network architecture, we test the potential benefits offered by a range of advanced techniques for machine learning and deep learning in the context of a typical classification problem encountered in the domain of high-energy physics, using a well-studied dataset: the 2014 Higgs ML Kaggle dataset. The advantages are evaluated in terms of both performance metrics and the time required to train and apply the resulting models. Techniques examined include domain-specific data-augmentation, learning rate and momentum scheduling, (advanced) ensembling in both model-space and weight-space, and alternative architectures and connection methods. Following the investigation, we arrive at a model which achieves equal performance to the winning solution of the original Kaggle challenge, whilst requiring about 1% of the training time and less than 5% of the inference time using much less specialised hardware. Additionally, a new wrapper library for PyTorch called LUMIN is presented, which incorporates all of the techniques studied.

The performance demands of future particle-physics experiments investigating the high-energy frontier pose a number of new challenges, forcing us to find new solutions for the detection, identification, and measurement of final-state particles in subnuclear collisions. One such challenge is the precise measurement of muon momenta at very high energy, where the curvature provided by conceivable magnetic fields in realistic detectors proves insufficient to achieve the desired resolution. In this work we show the feasibility of an entirely new avenue for the measurement of the energy of muons based on their radiative losses in a dense, finely segmented calorimeter. This is made possible by the use of the spatial information of the clusters of deposited photon energy in the regression task. Using a homogeneous lead-tungstate calorimeter as a benchmark, we show how energy losses may provide significant complementary information for the estimate of muon energies above 1 TeV.

Beginning from a basic neural-network architecture, we test the potential benefits offered by a range of advanced techniques for machine learning, in particular deep learning, in the context of a typical classification problem encountered in the domain of high-energy physics, using a well-studied dataset: the 2014 Higgs ML Kaggle dataset. The advantages are evaluated in terms of both performance metrics and the time required to train and apply the resulting models. Techniques examined include domain-specific data-augmentation, learning rate and momentum scheduling, (advanced) ensembling in both model-space and weight-space, and alternative architectures and connection methods. Following the investigation, we arrive at a model which achieves equal performance to the winning solution of the original Kaggle challenge, whilst being significantly quicker to train and apply, and being suitable for use with both GPU and CPU hardware setups. These reductions in timing and hardware requirements potentially allow the use of more powerful algorithms in HEP analyses, where models must be retrained frequently, sometimes at short notice, by small groups of researchers with limited hardware resources. Additionally, a new wrapper library for PyTorch called LUMIN is presented, which incorporates all of the techniques studied.

HEP search analysis Monte Carlo simulation dataset for ML tutorials. Available in both CSV and ROOT format. Signal 'gen_target = 1' is standard model non-resonant di-Higgs via gluon fusion -&gt; bb tautau Background 'gen_target = 0' is fully-leptonic ttbar Sample is post skim selecting events in the mu tau_h channel b_0 = b-jet with highest pT b_0 = b-jet with second highest pT t_0 = hadronically decaying tau lepton t_1 = muon resulting from tau lepton decay mPT = missing transverse momentum h_bb = ℎ→𝑏𝑏, h→bb candidate (vector sum of b_0 and b_1) h_tt = ℎ→𝜏𝜏, h→ττ candidate (vector sum of t_0 and t_1 and mPT) diH = di-Higgs vector (the vector sum of h_bb and h_tt) Signal generation: MadGraph -&gt; Pythia -&gt; Delphes Background generation: Powheg -&gt; Pythia -&gt; Delphes Centre of mass energy is sqrt(13 TeV), Delphes detector simulation is for a general purpose in between the ATLAS and CMS detectors.

HEP search analysis Monte Carlo simulation dataset for ML tutorials. Available in both CSV and ROOT format. Signal 'gen_target = 1' is standard model non-resonant di-Higgs via gluon fusion -&gt; bb tautau Background 'gen_target = 0' is fully-leptonic ttbar Sample is post skim selecting events in the mu tau_h channel b_0 = b-jet with highest pT b_0 = b-jet with second highest pT t_0 = hadronically decaying tau lepton t_1 = muon resulting from tau lepton decay mPT = missing transverse momentum h_bb = ℎ→𝑏𝑏, h→bb candidate (vector sum of b_0 and b_1) h_tt = ℎ→𝜏𝜏, h→ττ candidate (vector sum of t_0 and t_1 and mPT) diH = di-Higgs vector (the vector sum of h_bb and h_tt) Signal generation: MadGraph -&gt; Pythia -&gt; Delphes Background generation: Powheg -&gt; Pythia -&gt; Delphes Centre of mass energy is sqrt(13 TeV), Delphes detector simulation is for a general purpose in between the ATLAS and CMS detectors.

The performance demands of future particle-physics experiments investigating the high-energy frontier pose a number of new challenges, forcing us to find improved solutions for the detection, identification, and measurement of final-state particles in subnuclear collisions. One such challenge is the precise measurement of muon momentum at very high energy, where an estimate of the curvature provided by conceivable magnetic fields in realistic detectors proves insufficient for achieving good momentum resolution when detecting, e.g., a narrow, high mass resonance decaying to a muon pair. In this work we show the feasibility of an entirely new avenue for the measurement of the energy of muons based on their radiative losses in a dense, finely segmented calorimeter. This is made possible by exploiting spatial information of the clusters of energy from radiated photons in a regression task. The use of a task-specific deep learning architecture based on convolutional layers allows us to treat the problem as one akin to image reconstruction, where images are constituted by the pattern of energy released in successive layers of the calorimeter. A measurement of muon energy with better than 20% relative resolution is shown to be achievable for ultra-TeV muons.

The unsupervised search for overdense regions in high-dimensional feature spaces, where locally high population densities may be associated with anomalous contaminations to an otherwise more uniform population, is of relevance to applications ranging from fundamental research to industrial use cases. Motivated by the specific needs of searches for new phenomena in particle collisions, we propose a novel approach that targets signals of interest populating compact regions of the feature space. The method consists in a systematic scan of subspaces of a standardized copula of the feature space, where the minimum p-value of a hypothesis test of local uniformity is sought by gradient descent. We characterize the performance of the proposed algorithm and show its effectiveness in several experimental situations.

This record contains the fully-preprocessed training/validation and testing datasets used to train and evaluate the final models for "Calorimetric Measurement of Multi-TeV Muons via Deep Regression" by Jan Kieseler, Giles C. Strong, Filippo Chiandotto, Tommaso Dorigo, &amp; Lukas Layer, (2021), arXiv:2107.02119 [physics.ins-det] (https://arxiv.org/abs/2107.02119). The files are LZF-compressed HDF5 format and designed to be used directly with the code-base available at https://github.com/GilesStrong/calo_muon_regression. Please use the 'issues' tab on the GitHub repo for any questions or problems with these datasets. The training dataset consists of 886,716 muons with energies in the continuous range [50,8000] GeV split into 36 subsamples (folds). The zeroth fold of this dataset is used as our validation data. The testing dataset contains 429,750 muons, generated at fixed values of muon energy (E=100, 500, 900, 1300, 1700, 2100, 2500, 2900, 3300, 3700, 4100 GeV), and split into 18 folds. The input features are the raw hits in the calorimeter (stored in a sparse COO representation), and the high-level features discussed in the paper.