Universal Monte Carlo Event Generator Nobuo Sato Supported by - PowerPoint PPT Presentation

Universal Monte Carlo Event Generator Nobuo Sato Supported by Jefferson Lab Laboratory CHEP19, Adelaide research and development (LDRD19-13) 1 / 18

Partnership with computer scientists Y. Alanazi (ODU) M. P. Kuchera (Davidson College) Y. Li (co-PI) (ODU) T. Liu (JLab) R. E. McClellan (JLab) W. Melnitchouk (PI) (JLab) E. Pritchard (Davidson College) R. Ramanujan (Davidson College) M. Robertson (Davidson College) NS (co-PI) (JLab) R. R. Strauss (Davidson College) L. Velasco (Dallas) 2 / 18

The big picture hadrons as emergent phenomena of QCD quarks and gluons 3 / 18

The big picture hadrons as emergent phenomena of QCD nucleon structure quarks and gluons 3 / 18

The big picture hadrons as emergent phenomena of QCD nucleon structure quarks and gluons hadronization 3 / 18

Motivations A new era of nuclear physics has started with the JLab 12 GeV program 4 / 18

Motivations A new era of nuclear physics has started with the JLab 12 GeV program New tools based on Machine Learning (ML) to boost the discovery potential are needed 4 / 18

The goals 5 / 18

The goals Build a theory-free MCEG 5 / 18

The goals Build a theory-free MCEG Map out particles correlations without biases from approximated theory 5 / 18

The goals Build a theory-free MCEG Map out particles correlations without biases from approximated theory MCEG as a data storage utility 5 / 18

Nature e − P 6 / 18

Nature e − experimental P detector 6 / 18

Nature e − experimental P detector detector level events 6 / 18

Nature vertex level events e − experimental P detector detector level events 6 / 18

Nature vertex level events e − experimental P detector detector level events 6 / 18

Nature vertex level events experimental detector detector level events 7 / 18

Can we use ML to: Nature vertex level events experimental detector detector level events 7 / 18

Can we use ML to: Nature simulate vertex level events? vertex level events experimental detector detector level events 7 / 18

Can we use ML to: Nature simulate vertex level events? vertex level events simulate detector level events? experimental detector detector level events 7 / 18

Can we use ML to: Nature simulate vertex level events? vertex level events simulate detector level events? experimental detector simulate nature ? detector level events 7 / 18

Nature UMCEG vertex level vertex level events events experimental detector simulator detector detector level detector level events events 7 / 18

datacompression Nature UMCEG vertex level vertex level events events experimental detector simulator detector detector level detector level events events 7 / 18

Our strategy Event level ML training → GAN 8 / 18

Our strategy Event level ML training → GAN Use a dual GAN as the event generator ρ (particles | multiplicity ) × ρ (multiplicity) � �� � � �� � vectors generator multiplicity generator 8 / 18

Challenges Find optimal data representation → what is the image of an event ? 9 / 18

Challenges Find optimal data representation → what is the image of an event ? How to make the GAN to learn the features of the event ? → CNN 9 / 18

Challenges Find optimal data representation → what is the image of an event ? How to make the GAN to learn the features of the event ? → CNN How to escalate from low to higher multiplicities? 9 / 18

Our current work in progress Use Pythia as a training and validation tool 10 / 18

Our current work in progress Use Pythia as a training and validation tool Ignore detector effects 10 / 18

Our current work in progress Use Pythia as a training and validation tool Ignore detector effects Start with inclusive particle generator ρ (particles | multiplicity) → ρ (particles + X) 10 / 18

Pythia UMCEG vertex level vertex level events events detector level detector level events events 11 / 18

10 0 Pythia GAN 10 − 1 probabilities 10 − 2 10 − 3 10 − 4 10 − 5 L ¯ γ e − e + µ − µ + ν e ¯ ν e ν µ ¯ ν µ ν τ ¯ K 0 K 0 ν τ p n π + π − K + p n ¯ ¯ K − L Multiplicity generator 12 / 18

z ∈ N (0 , 1) Generator FC NN Pythia FC NN l + p → l ′ + X FC NN p x p y p x p y p z p z p x p y p x p y p z p z p i → k i p i → k i Features Transform Features Extension k x k y k x k y k z k i k j k T k 0 k z /k T k z k i k j k T k 0 k z /k T k x k y k x k y k z k i k j k T k 0 k z /k T k z k i k j k T k 0 k z /k T Discriminator FC NN MMD FC NN FC NN Vectors generator Wasserstein Loss MMD Loss 13 / 18

z ∈ N (0 , 1) Event image = l ′ Generator FC NN x,y,z Pythia FC NN l + p → l ′ + X FC NN p x p y p x p y p z p z p x p y p x p y p z p z p i → k i p i → k i Features Transform Features Extension k x k y k x k y k z k i k j k T k 0 k z /k T k z k i k j k T k 0 k z /k T k x k y k x k y k z k i k j k T k 0 k z /k T k z k i k j k T k 0 k z /k T Discriminator FC NN MMD FC NN FC NN Vectors generator Wasserstein Loss MMD Loss 13 / 18

z ∈ N (0 , 1) Event image = l ′ Generator FC NN x,y,z Pythia FC NN l + p → l ′ + X FC NN Feature extension: p x p y p x p y p z p z p x p y p x p y p z p z l ′ i · l ′ j , l ′ 0 , l ′ z /l ′ p i → k i p i → k i Features Transform T Features Extension k x k y k x k y k z k i k j k T k 0 k z /k T k z k i k j k T k 0 k z /k T k x k y k x k y k z k i k j k T k 0 k z /k T k z k i k j k T k 0 k z /k T Discriminator FC NN MMD FC NN FC NN Vectors generator Wasserstein Loss MMD Loss 13 / 18

z ∈ N (0 , 1) Event image = l ′ Generator FC NN x,y,z Pythia FC NN l + p → l ′ + X FC NN Feature extension: p x p y p x p y p z p z p x p y p x p y p z p z l ′ i · l ′ j , l ′ 0 , l ′ z /l ′ p i → k i p i → k i Features Transform T Features Extension k x k y k x k y k z k i k j k T k 0 k z /k T k z k i k j k T k 0 k z /k T k x k y k x k y k z k i k j k T k 0 k z /k T k z k i k j k T k 0 k z /k T Discriminator WGAN+MMD Butter, Plehn, FC NN MMD Winterhalder (’19) FC NN FC NN Vectors generator Wasserstein Loss MMD Loss 13 / 18

Validation 14 / 18

Validation Relevant observables for inclusive DIS Q 2 = − ( l − l ′ ) 2 Q 2 x bj = 2 P · ( l − l ′ ) 14 / 18

Validation Relevant observables for inclusive DIS Q 2 = − ( l − l ′ ) 2 Q 2 x bj = 2 P · ( l − l ′ ) x bj , Q 2 not included as features 14 / 18

10 − 1 Normalized Yield Normalized Yield 10 1 10 − 3 10 0 10 − 5 10 − 1 10 − 7 GAN GAN Pythia 10 − 2 Pythia 10 − 9 10 1 10 2 10 3 10 4 0 . 01 0 . 1 1 Q 2 (GeV 2 ) x bj 10 0 Error bands generated with GAN Normalized Yield 10 − 1 Pythia bootstrapped samples 10 − 2 10 − 3 10 − 4 10 − 5 0 5 10 15 20 25 30 p T (GeV) 15 / 18

Q 2 Pythia 10 2 Isocontours are in 10 1 agreement GAN x bj , Q 2 correlation is 10 2 learned without adding x bj · Q 2 feature 10 1 0 . 001 0 . 01 0 . 1 1 x bj 16 / 18

Summary and outook It is possible to train a GAN at the event level to build a MCEG 17 / 18

Summary and outook It is possible to train a GAN at the event level to build a MCEG The current design provides a blueprint for a generator with higher multiplicity 17 / 18

Summary and outook More work is needed, but the results are encouraging 18 / 18

Summary and outook More work is needed, but the results are encouraging A fully trained UMCEG will be a complementary tool to theory-based MCEGs such as Pythia 18 / 18