8/7/2019 siwif-logo.png Summer Institute on Wigner Imaging and Femtography Simonetta Liuti University of Virginia CNF Symposium August 12-13, 2019 SURA Headquarters, Washington DC
2 8/13/19 Summer Institute for Wigner Imaging and Femtography Simonetta Liuti Matthias Burkardt Pete Alonzi Dustin Keller Olivier Pfister Principle Investigator Co Principle Investigator Co Principle Investigator Co Principle Investigator Co Principle Investigator University of Virginia New Mexico State University University of Virginia University of Virginia University of Virginia Data Management/ Wigner Theory Machine Learning Observables Communication Jake Grigsby Brandon Kriesten Librado Anglero Yao(Grace) Tong Machine Learning Group Leader University of Virginia University of Virginia University of Virginia University of Virginia Physics Observables Group Leader Mathematics and Economics Computer Science and Mathematics Fatma Aslan Evan Anders Magnusson Krisean D Allen New Mexico State University University of Virginia Virginia Union University Consultant PhD Computer Engineering and Physics Computer Science Kyle-Thomas Pressler Christopher Thompson Meg Graham Carlos Gonzalez Arciniegas University of Virginia Virginia Union University University of Virginia University of Virginia Physics Physics and Engineering Computer Science Andrew Meyer Timothy John Hobbs Emma Yeats University of Virginia University of Virginia Southern Methodist University Physics EIC Center at Jefferson Lab William A Oliver Fernanda Yepez-Lopez Gabriel Niculescu Virginia Commonwealth University University of Virginia James Madison University Mathematics Yelena Prok Abha Rajan *** Virginia Commonwealth University University of Virginia Assistant Professor Red: Undergraduate Blue: Graduate
3 8/13/19 Our project was characterized from the very beginning by a great response from students and young researchers:
4 8/13/19 https://www.uwlax.edu/ceu/current/ 3 contributions submitted to CEU@ Annual Fall Meeting of DNP!
5 8/13/19 A truly interdisciplinary effort Quantum Information Data Science 3D Structure of the proton Education Outreach
6 8/13/19 Main Questions we want to answer � What are the necessary steps for an unbiased extraction of the 3D structure and mechanical properties of the proton? � GPDs and Compton Form Factors are the observables � Probing nuclear physics with gravitational waves
7 8/13/19 Wigner Theory Data Analysis Modeling the Visualization Wigner Function Modeling the Communication Exploring issues Cross Section Outreach/ in common with Extracting pedagogical Atomic Physics Data Observables Research Tool Management from Data Dissemination Error Analysis Webpage/github Code sharing opensourcing
8 8/13/19 Wigner Theory Modeling the Wigner Function Exploring issues in common with Atomic Physics
9 8/13/19 Deeply Virtual Compton Scattering ep → e ' γ ' p ' e’ -(k-k’) 2 =Q 2 large scale q’ k’ k p e p’
10 8/13/19 � We focus on GPDs (one projection of the Wigner distribution which is directly observable in DVCS) � The observables are the Compton Form Factors which are convolutions of GPDs with known kernels
Flowchart/roadmap from data/observables to GPDs Generate Wigner/GPD forms @ Q o 2 H i=u,d,.. ({a}), E i=u,d,.. ({a}),… Constraint 1 Nucleon Form Factors 1st partial parameter fix F 1 , F 2 ,G A , G V Constraint 2 Solve pQCD Evolution PDFs u(x,Q 2 ), u(x,Q 2 ), … Δ u(x,Q 2 ), Δ d(x,Q 2 ), G(x, Q 2 ) 2 nd partial parameters fix Data DVCS,TCS,DVMP Evaluate Observables Evaluate d 5 σ DVCS A LU, A LL, A UT,… CFFs Evaluate Fit Result Estimator
12 8/13/19 u,d quarks and gluon GPDs …to Fourier transform: H(x,b) from PQCD evolved GPD H(x,0,t) parameter variations used for input for Data Analysis H(x,0,t) Fixed t H i = u,d,... ( x, ξ , t ) = N G { a } ( x, ξ , t ) R { β } ( x, ξ , t ) x { a } = M X , m, M Λ { β } = α , α 0 , p (parametrization from J. O. Gonzalez et al., arXiv:1206.1876, PRC88(2013) )
13 8/13/19 Working on Connection with Atomic Physics/Quantum Information with O. Pfister and C. Gonzalez Arciniegas
14 8/13/19 Experimental Results from UVA C. Gonzalez Arciniegas https://pages.shanti.virginia.edu/Femtography/ files/2019/06/Seminar.pdf
15 8/13/19 Data Analysis Modeling the Cross Section Extracting Observables from Data Error Analysis
16 8/13/19 y r a n i l p i c s i d r a e t t a n d I h t i w t r o f f e e c n e Machine Learning i c s correlated work Analytic Development: Precise Formulation of Cross Section
17 8/13/19 • What physics questions drive our choice? • What type of Machine Learning? • What kinematical domains are sensitive to which GPDs/Compton Form Factors • With what uncertainty?
Flowchart/roadmap from data/observables to GPDs Generate Wigner/GPD forms @ Q o 2 H i=u,d,.. ({a}), E i=u,d,.. ({a}),… Constraint 1 Nucleon Form Factors 1st partial parameter fix F 1 , F 2 ,G A , G V Constraint 2 Solve pQCD Evolution PDFs u(x,Q 2 ), u(x,Q 2 ), … Δ u(x,Q 2 ), Δ d(x,Q 2 ), G(x, Q 2 ) 2 nd partial parameters fix Data DVCS,TCS,DVMP Evaluate Observables Evaluate d 5 σ DVCS A LU, A LL, A UT,… CFFs Evaluate ML TensorFlow (logo Fit Result Estimator
19 8/13/19 Introducing the complete Modeling the Cross Section formalism where all kinematical dependences are precisely known and approximations are under control
20 8/13/19 Rosenbluth separation for Bethe-Heitler contribution d 5 σ BH ⇥ � � ⇤ = Γ unpol F 2 1 + τ F 2 + B BH τ G 2 M ( t ) A BH 2 dx Bj dQ 2 d | t | d φ d φ S t 2 h d 5 σ I Γ unpol = A I ( F 1 < e H + τ F 2 < e E ) + B I G M < e ( H + E ) + C I dx Bj dQ 2 d | t | d φ d φ S Q 2 ( � t )
21 8/13/19 G d 5 σ BH ⇥ � � ⇤ = Γ unpol F 2 1 + τ F 2 G E + B BH τ G 2 2 G M 2 G M G A A BH M ( t ) 2 dx Bj dQ 2 d | t | d φ d φ S t 2 h i d 5 σ I Γ unpol A I ( F 1 < e H + τ F 2 < e E ) + B I G M < e ( H + E ) + C I G M < e e = H dx Bj dQ 2 d | t | d φ d φ S Q 2 ( � t )
22 8/13/19 Extracting Observables from Data/Error Analysis The goal is to single out what measurements have the greatest impact on determining the Compton Form Factors. To address this challenge we present two analysis methods to identify both the type of experiments (beam/target polarization) and the kinematical domain providing the best constraints.
23 8/13/19 Based on Supervised Learning…
24 8/13/19 Femtography Imaging with Neural Networks (FINN) Strategy: 1. A fully connected neural network maps input kinematic data to a vector of eight form factors (see diagram). 1. 2. Use a code developed by our Data Analysis Team to evaluate the cross sections and in terms of the CFFs. We translate the x-sec. code into TensorFlow � Automatically differentiable � At variance with other efforts we can train CFF extraction network with backpropagation and variants of stochastic gradient descent.
25 8/13/19 Understanding the Error/Systematic bias: example Re H Q 2 Large error Small error x Bj present data kinematical coverage
26 8/13/19 The biggest challenge we have is limited data, which we can solve with a combination of regularization, Monte Carlo generation and interpolation between phi values for each kinematic range. Larger datasets from future DVCS/TCS experiments will ease these engineering challenges.
27 8/13/19 Based on unsupervised learning….
28 8/13/19 Ultimately, we want to understand features of the GPDs behavior Small t Large t diquark correlations Reggeized diquark model
29 8/13/19 Using UMAP (Uniform Manifold Approximation and Projection) to cluster GPDs from the Wigner Theory team into 2D Can be extended to 3D for virtual reality!
… similarities with CTEQ analysis Principal Component Analysis https://arxiv.org/pdf/1803.02777.pdf
31 8/13/19 Visualization Outreach/ pedagogical Research Tool
32 8/13/19 Outreach tool
33 8/13/19 Stab at visualizing UMAP
34 8/13/19 Communication Data Management Plan Dissemination Webpage/github Code sharing Open-sourcing See Pete Alonzi’s talk on Monday
35 8/13/19 Summary � We developed a strategy and specifically ML tools for analyzing deeply virtual exclusive experiments within a truly interdisciplinary effort � Visualization tool for outreach � Impact on Education: many undergradauate students from various departments involved in research, CEU presentations � Several manuscripts in progress � Website and model for both fostering and regulating community interactions
36 8/13/19 Future Work Issues with data: situations where data are scarce, how to add in dynamically new sets of data, Including lattice QCD results/constraints Fourier transforms More work to come on Wigner Distributions and Quantum Information Developing Unsupervised Learning tool Visualization tools for research and more animations/movies for outreach tool
37 8/13/19
Recommend
More recommend
