Machine learning for event reconstruction and particle identification at CLAS12
Explore the source record for details and available documents.
Engineering topics
Publications and source records attributed to Gavalian, Gagik.
Explore the source record for details and available documents.
In this work, we present studies of track parameter reconstruction from raw information in CLAS12 detector's Drift Chambers, using Machine Learning (ML). We study the resolution of tracks reconstructed with different types of ML models/algorithms, including Multi-Layer Perceptron (MLP), Extremely Randomized Trees (ERT) and Gradient Boosting Trees (GBT) using simulated data. We find that the resulting ML model is capable of reconstructing track parameters (particle momentum, and polar and azimuthal angles) with accuracy similar to Hit Based (HB) tracking code, but $150$ times faster. Moreover, physics reactions can be identified using the particles reconstructed by the neural network in real-time (with a rate of about $34~kHz$) during experimental data collection. The developed model can be used in numerous applications, such as triggering specific physics reactions in real-time, detector performance monitoring, and real-time detector calibration.
Imagine being able to predict — with unprecedented accuracy and precision — the structure of the proton and neutron, and the forces between them, directly from the dynamics of quarks and gluons, and then using this information in calculations of the structure and reactions of atomic nuclei and of the properties of dense neutron stars (NSs). Also imagine discovering new and exotic states of matter, and new laws of nature, by being able to collect more experimental data than we dream possible today, analyzing it in real time to feed back into an experiment, and curating the data with full tracking capabilities and with fully distributed data mining capabilities. Making this vision a reality would improve basic scientific understanding, enabling us to precisely calculate, for example, the spectrum of gravity waves emitted during NS coalescence, and would have important societal applications in nuclear energy research, stockpile stewardship, and other areas. This review presents the components and characteristics of the exascale computing ecosystems necessary to realize this vision.
Particle track reconstruction is the most computationally intensive process in nuclear physics experiments. Traditional algorithms use a combinatorial approach that exhaustively tests track measurements ("hits") to identify those that form an actual particle trajectory. In this article, we describe the development of four machine learning (ML) models that assist the tracking algorithm by identifying valid track candidates from the measurements in drift chambers. Several types of machine learning models were tested, including: Convolutional Neural Networks (CNN), Multi-Layer Perceptrons (MLP), Extremely Randomized Trees (ERT) and Recurrent Neural Networks (RNN). As a result of this work, an MLP network classifier was implemented as part of the CLAS12 reconstruction software to provide the tracking code with recommended track candidates. The resulting software achieved accuracy of greater than 99% and resulted in an end-to-end speedup of 35% compared to existing algorithms.
Modern Nuclear Physics experimental setups run experiments with higher beam intensity resulting in increased noise in detector components used for particle track reconstruction. Increased uncorrelated signals (noise) result in decreased particle reconstruction efficiency. In this paper, we investigate the usage of Machine Learning, specifically Convolutional Neural Network Auto-Encoders (CAE), for de-noising raw hits from drift chambers in the CLAS12 detector. To the best of our knowledge, this is the first time CAE is employed to perform such an operation in this field. During the de-noising phase, it is important to remove as much noise as possible while retaining the valid hits to avoid losing crucial information about the experiment. Here, we show that using CAE, it is possible to remove noise hits while retaining up to 94% of valid tracks for a beam current of 110nA while for lower beam currents (45-55nA), we get up to 98% efficiency. Studies on experimental conditions with increasing noise show that CAE performs better than conventional tracking algorithms in isolating hits belonging to tracks. Specifically, the de-noising algorithm results in tracking efficiency improvements greater than 15%, in real data production procedures with nominal conditions, and up to two times better efficiency in synthetically generated data with high luminosity conditions (90-110nA), indicating that machine learning can lead to significantly shorter times for conducting physics experiments.
Charged particle tracking represents the largest consumer of CPU resources in high data volume Nuclear Physics (NP) experiments. An effort is underway to develop machine learning (ML) networks that will reduce the resources required for charged particle tracking. Tracking in NP experiments represent some unique challenges compared to high energy physics (HEP). In particular, track finding typically represents only a small fraction of the overall tracking problem in NP. This presentation will outline the differences and similarities between NP and HEP charged particle tracking and areas where ML learning may provide a benefit. The status of the specific effort taking place at Jefferson Lab will also be shown.
Explore the source record for details and available documents.
We report the first measurement of the photoproduction cross section of the $\phi$ meson in its neutral decay mode in the reaction $\gamma p \to p\phi(K_SK_L)$. The experiment was performed with a tagged photon beam of energy $1.6 \le E_\gamma \le 3.6$ GeV incident on a liquid hydrogen target of the CLAS spectrometer at the Thomas Jefferson National Accelerator Facility. The $p \phi$ final state is identified via reconstruction of $K_S$ in the invariant mass of two oppositely charged pions and by requiring the missing particle in the reaction $\gamma p \to p K_S X$ to be $K_L$. The presented results significantly enlarge the existing data on $\phi$-photoproduction. These data, combined with the data from the charged decay mode, will help to constrain different mechanisms of $\phi$ photoproduction.
The spin structure functions $g_1$ for the proton and the deuteron have been measured over a wide kinematic range in $x$ and \Q2 using 1.6 and 5.7 GeV longitudinally polarized electrons incident upon polarized NH$_3$ and ND$_3$ targets at Jefferson Lab. Scattered electrons were detected in the CEBAF Large Acceptance Spectrometer, for $0.05 < Q^2 < 5 $\ GeV$^2$ and $W < 3$ GeV. The first moments of $g_1$ for the proton and deuteron are presented -- both have a negative slope at low \Q2, as predicted by the extended Gerasimov-Drell-Hearn sum rule. The first result for the generalized forward spin polarizability of the proton $\gamma_0^p$ is also reported, and shows evidence of scaling above $Q^2$ = 1.5 GeV$^2$. Although the first moments of $g_1$ are consistent with Chiral Perturbation Theory (\ChPT) calculations up to approximately $Q^2 = 0.06$ GeV$^2$, a significant discrepancy is observed between the $\gamma_0^p$ data and \ChPT\ for $\gamma_0^p$,even at the lowest \Q2.
The $e p\to e^\prime p \rho^0$ reaction has been measured, using the 5.754 GeV electron beam of Jefferson Lab and the CLAS detector. This represents the largest ever set of data for this reaction in the valence region. Integrated and differential cross sections are presented. The $W$, $Q^2$ and $t$ dependences of the cross section are compared to theoretical calculations based on $t$-channel meson-exchange Regge theory on the one hand and on quark handbag diagrams related to Generalized Parton Distributions (GPDs) on the other hand. The Regge approach can describe at the $\approx$ 30% level most of the features of the present data while the two GPD calculations that are presented in this article which succesfully reproduce the high energy data strongly underestimate the present data. The question is then raised whether this discrepancy originates from an incomplete or inexact way of modelling the GPDs or the associated hard scattering amplitude or whether the GPD formalism is simply in
The exclusive channel p-vectore-vector,e{sup '}p){pi}{sup 0} was studied in the first and second nucleon resonance regions in the Q{sup 2} range from 0.187 to 0.770 GeV{sup 2} at Jefferson Lab using the CEBAF Large Acceptance Spectrometer. Longitudinal target and beam-target asymmetries were extracted over a large range of center-of-mass angles of the {pi}{sup 0} and compared to the unitary isobar model MAID, the dynamic model by Sato and Lee, and the dynamic model DMT. A strong sensitivity to individual models was observed, in particular for the target asymmetry and in the higher invariant mass region. This data set, once included in the global fits of the above models, is expected to place strong constraints on the electrocoupling amplitudes A{sub 1/2} and S{sub 1/2} for the Roper resonance N(1400)P{sub 11} and the N(1535)S{sub 11} and N(1520)D{sub 13} states.
The helicity amplitudes of the electroexcitation of the Roper resonance are extracted for 1.7<Q{sup 2}<4.5 GeV{sup 2} from recent high precision JLab-CLAS cross section and longitudinally polarized beam asymmetry data for {pi}{sup +} electroproduction on protons at W=1.15-1.69 GeV. The analysis is made using two approaches, dispersion relations and a unitary isobar model, which give consistent Q{sup 2} behavior of the helicity amplitudes for the {gamma}*p{yields}N(1440)P{sub 11} transition. It is found that the transverse helicity amplitude A{sub 1/2}, which is large and negative at Q{sup 2}=0, becomes large and positive at Q{sup 2}{approx_equal}2 GeV{sup 2}, and then drops slowly with Q{sup 2}. The longitudinal helicity amplitude S{sub 1/2}, which was previously found from CLAS e-vectorp{yields}ep{pi}{sup 0},en{pi}{sup +} data to be large and positive at Q{sup 2}=0.4,0.65 GeV{sup 2}, drops with Q{sup 2}. Available model predictions for {gamma}*p{yields}N(1440)P{sub 11} allow us to conclude that these results provide strong evidence in favor of N(1440)P{sub 11} as a first radial excitation of the 3q ground state. The results of the present paper also confirm the conclusion of our previous analysis for Q{sup 2}<1 GeV{sup 2} that the presentation of N(1440)P{sub 11} as a q{sup 3}G hybrid state is ruled out.
Electroproduction of exclusive {phi} vector mesons has been studied with the CLAS detector in the kinematic range 1.4{<=}Q{sup 2}{<=}3.8 GeV{sup 2},0.0{<=}t{sup '}{<=}3.6 GeV{sup 2}, and 2.0{<=}W{<=}3.0 GeV. The scaling exponent for the total cross section as 1/(Q{sup 2}+M{sub {phi}}{sup 2}){sup n} was determined to be n=2.49{+-}0.33. The slope of the four-momentum transfer t{sup '} distribution is b{sub {phi}}=0.98{+-}0.17 GeV{sup -2}. Under the assumption of s-channel helicity conservation, we determine the ratio of longitudinal to transverse cross sections to be R=0.86{+-}0.24. A two-gluon exchange model is able to reproduce the main features of the data.