NEURAL NETWORKS FOR MODELING AND CONTROL OF PARTICLE ACCELERATORS - PowerPoint PPT Presentation

NEURAL NETWORKS FOR MODELING AND CONTROL OF PARTICLE ACCELERATORS Auralee Edelen 2016-02-02 Advisors: Sandra Biedron and Stephen Milton

Published Work • Edelen, A. et al. Neural Networks for Modeling and Control of Particle Accelerators. Submitted to IEEE Transactions in Nuclear Science, Jan. 2015. • Edelen, A., et al. Initial Experimental Results of a Machine Learning-Based Temperature Control System for an RF Gun. Paper for the 6th International Particle Accelerator Conference (IPAC), Richmond, VA, May 3-8, 2015. • Morin, A., et al. Trajectory Response Studies at the Jefferson Laboratory Energy Recovery Linac and Free Electron Laser. Paper for the 16th Annual Directed Energy Symposium, Huntsville, AL, March 10-14, 2014.

Control Challenges Fermilab JLAB SLAC LBNL Visualization Group Fermilab www.tka-architects.com

Inspiration from Operators Model learning Prediction Planning Diagnostic Optimization Analysis Learning Control Control Room Photo: Reidar Hahn, FNAL

A trip to the zoo … Intelligent Control Biological Sciences and Psychology Artificial Intelligence Nonlinear Control (inspiration!) Adaptive Control Machine Mathematical Learning Optimal Control Optimization Regression Robust Control Classification Gradient descent Clustering Conjugate gradient Stochastic Control Dimensionality reduction Newton method Quasi-Newton methods System Identification Learning Theory Simulated annealing Supervised Learning Model-free Methods Evolutionary algorithms Unsupervised Learning Swarm intelligence Model-based Methods Reinforcement Learning Ensemble Methods Reactive search optimization Computational Support Vector Machines Fuzzy Logic Statistics Neural Networks Decision Trees ICA, PCA

Many Failures Early On à So Why Try Again Now? In general: greater theoretical understanding + J. Schmidhuber increased computational capability IBM, ANL + advantageous co-developments in related fields + feedback from a wider variety of application attempts B. Rhoads, UCSB à greater overall technological maturity Shutterstock Google But still difficult in the context of nonlinear control à à we need R&D !

Central Focus: Let’s develop and test some AI-based solutions for control problems in accelerators! • Explore the tools and techniques • Examine some real-world problems, focusing on process control • Need to test on an actual machine; not just in simulation • Have at it!

Some Tools

Neural Networks x 1 w 1 w 2 x 2 • What are they . f y . . w n x n • How do they learn? a neuron (“node”) • When are they useful? x 1 x 2 . . • What are the disadvantages? . x n a neural network à How can we use these things in particle accelerators?

Learning Paradigms

Model Predictive Control Basic concept: use a predictive model to assess the outcome of possible future actions

Model Predictive Control Measured Variables Reference Trajectory u m (k – 1)… u m (k – N m ) y r (k)… y r (k + N p ) N m previous measurements N p future time steps predicted Optimization of Controlled Variable Trajectories N c future time steps controlled Predicted Outputs y p (k)… y p (k + N p ) ! ! ! ! ! ! ! ! ! ! + ! − ! ! ! + ! Cost Function ! ! ! Plant Model (output variable targets) Constraints ! ! ! ! ! Solver ! !" ! ! , ! ! ! ! + ! − ! ! , !"# ! + ! ! Future Inputs ! ! ! ! ! ! (controllable variable targets) u cv (k)… u cv (k + N c – 1) ! ! ! ! ! ! !" ! ∆ ! , ! ! ! ! + ! − ! ! ! + ! − 1 ! ! ! ! ! ! (movement size) u cv (k) Plant

Reinforcement Learning

Real-World Problems

At Fermilab … RF electron gun at the Fermilab Accelerator Science and Technology (FAST) facility — Long, variable time delays — Tight tolerances — Recursive behavior — Two controllable parameters Photo: P. Stabile High-intensity RFQ for the PIP-II Injector Experiment (PXIE) — Time delays — Large, dynamic frequency response — Tight tolerances — Coupling — Recursive behavior — Three controllable parameters Photo: J. Steimel

At Fermilab … RF electron gun at the Fermilab Accelerator Science and Technology (FAST) facility — Long, variable time delays — Tight tolerances — Recursive behavior — Two controllable parameters FAST!RF!Gun!Parameters Photo: P. Stabile !!Gun!Parameters !!Type Photoinjector !!Number!of!cells 1 ½ !!RF!Mode TM 010, π !!Loaded!Q ~11,700 !!RF!Frequency 1.3 GHz !!Frequency!Shift 23 kHz/°C !!Nominal!Operating!Parameters Photo: E. Harms !!Macropulse!Duration 1 ms !!Repetition!Rate 1 − 5 Hz !!Bunch!Frequency! 3 MHz !!Design!Gradient 40 − 45 MV/m !!Power!Source 5 MW Klystron

Gun Water Cooling System

Water Temperatures — Open Loop Temperature%[°C]% % Time%Elapsed%[minutes]% impulse response from a 20-second decrease in the heater power setting from 7 kW to 2.5 kW

Existing feed-forward/PI Control of the Gun Temperature Temperature%[°C]% % Time%Elapsed%[minutes]% 1-°C step change in temperature set point à Oscillation is NOT due to poorly tune PI gains!

Initial Solution • Neural network model • Model predictive control à Serves as a simple benchmark for future studies

Neural Network Modeling T01 ([ t - d 1 ], . . ., [ t - d 1 - n 1 ]) T01 ([ t-d 1 ], . . ., [ t-d 1 -n ]) model TOUT ([ t - d 2 ], . . ., [ t - d 2 - n 2 ]) T01 ([t-d 1 ], . . ., [t-d 1 -n]) valve ([ t - d 3 ], . . ., [ t - d 3 - n 3 ]) predicted next value of T02 heater ([ t - d 4 ], . . ., [ t - d 4 - n 4 ]) d - delay tme n - number of previous samples

Benchmark Controller

MPC Benchmark Controller Note: different horizontal and vertical scale than for PI loop ~5x faster settling time no more overshoot still needs work … (esp. T02-to-TCAV model)

MPC Benchmark Controller: Actions Requested by Controller Actual Read-backs Requested%Heater%Power%[kW] % Time%Elapsed%[minutes]% Requested%Control%Valve%Posi9on%[%%open] % % Time%Elapsed%[minutes]% %

Of course, there’s more to the story … . 3 Steady'State'Temperature'Difference'[°C]' TOUT − TIN 2.5 TCAV − TIN TOUT − TIN expected 2 1.5 1 0.5 0 0 0.5 1 1.5 2 Average RF Power [kW] Average'RF'Power'[kW]' ! !"# [ ℃ ] ! ! !" [ ℃ ] ! × ! !"#$ ! [ !"# ] ! !""# = !"#$% ! !""#$%& ! !"#"$%&' ! !"# ! ℃ !" ! !""# = ! !" ≈ ! !" !"# .

FAST Next Steps • Neural network model predictive control • Extension to direct resonance control • Reinforcement learning control

At Fermilab … RF electron gun at the Fermilab Accelerator Science and Technology (FAST) Facility — Long, variable time delays — Tight tolerances — Recursive behavior — Two controllable parameters Photo: P. Stabile High-intensity RFQ for the PIP-II Injector Experiment (PXIE) — Time delays — Large, dynamic frequency response — Tight tolerances — Coupling — Recursive behavior — Three controllable parameters Photo: J. Steimel

PXIE RFQ Constructed by LBNL PXIE!RFQ!Parameters !!RFQ!Design!Parameters !!RF!frequency 162.5 MHz !!Q-factor ~13,900 !!Loaded!Q ~7,000 !!Physical!Length 4.45 m (2.4 wavelengths) !!Vane-to-Vane!Voltage 60 kV !!Estimated!Power!Dissipation < 100 kW !!RF!Repetition!Rate pulsed − CW !!Beam!Parameters !!Current 0.5 − 10 mA (nominal 5 mA) !!Input!Energy 30 keV !!Output!Energy 2.1 MeV High-intensity RFQ for the PIP-II Injector Experiment (PXIE) — Time delays — Large, dynamic frequency response — Tight tolerances — Coupling — Recursive behavior — Three controllable parameters Photo: J. Steimel

PXIE RFQ 3-kHz max. freq. shift 0.1-°C water stabilization Vane channels 4"inner"channels" Wall channels 8"outer"channels" All images courtesy LBNL, D. Li, A. Lambert

Expected Frequency Response ANSYS simulation data courtesy A. Lambert, LBNL

Water System

Resonance&Control&System&Architecture&and&Component&Interfaces& ACNET&user&interface& CLX&Machine& & Opera&onal*Mode*Request* * Present*Opera&onal*Mode* Erlang& Resonance&Controller& Error*Read4backs* Data*Reading*and* Heartbeat* ACNET& Organiza&on* UDP&& LLRF& Opera&onal*Module* & Resonant*Frequency*Devia&on** Selec&on* RF*Amplitude,*Rep.*Rate,*Pulse* Length* CW*vs.*Pulsed*Indicator* Control*Calcula&on*in* Other*State*Informa&on*(TBD)* Corresponding*Module* * PLL*(“SEL”)*vs.*GDR*switch* Error4checking* (Op&onal)*feed4forward*amplitude* Protocol&Defini@on& Cryo5con& (from&protocol&file&and& Data*Organiza&on*and* * Cri&cal*Temperature*Sensors* protocol&compiler)& Sending* PLC& * Secondary*Temperature*Sensors* Flow*Valve*Read4backs* to#resonance#controller# Pressure*and*Flow*Readings* * from#resonance#controller# Flow*Valve*SeCngs* PLC*Cntrl.**vs.*Resonance*Cntrl.*