Egemen Kolemen, Princeton University
Abstract:
Fusion promises to be the ultimate green energy source of the future as it is abundant clean and greenhouse-emission free without the intermittency and location restrictions of solar and wind energy or fission’s safety and waste issues. While our current knowledge of plasma physics and technical capabilities is sufficiently mature for us to attempt to build fusion power reactors the path to economic competitiveness lies with compact high-energy-density fusion reactors. This requires operation at simultaneously physics parameters that are close to the edge of plasma instabilities and the technical possibilities of materials engineering and nuclear operation which is challenging. Artificial Intelligence and Machine Learning (AI/ML) help us tackle some of these fusion challenges. I will talk about some of our recent accomplishments in application of AI/ML to fusion reactors: 1) Robust plasma state prediction even when there is sensor failure 2) Finding the minimal set of diagnostics needed to operate a reactor 3) Fusing data from multiple diagnostics to obtain new physics insights 4) Prediction of plasma evolution by combining experimental data and simulations 5) Reinforcement learning control that achieves high performance fusion reactor operation without instabilities.
Bio:
Egemen Kolemen in an Associate Professor at Princeton University's Mechanical & Aerospace Engineering jointly appointed with the Andlinger Center for Energy and the Environment and the Princeton Plasma Physics Laboratory (PPPL). He is the director of the Program in Sustainable Energy, recipient of the David J. Rose Excellence in Fusion Engineering Award and the American Nuclear Society’s Technical Accomplishment Award, and an ITER Scientist Fellow. His research combines engineering and physics analysis to enable economically feasible fusion reactors. He currently leads research on machine learning, real-time diagnostics and control at KSTAR, NSTX-U and DIII-D. He directs liquid metal divertor and low temperature diagnostics labs. On the theoretical side, his group develops software for stellarator optimization and economical analysis of fusion reactor.
Summary:
Focus: generating energy from nuclear fusion
Fuse small atoms (e.g. Hydrogen) into larger ones
E.g. Deuterium + Tritium => Helium + Energy
Fusion in the sun
Gravity compresses the gases, so they can’t get away
Crushes them, forcing them to fuse
On Earth need another way to keep gases in place
Use strong magnetic fields
Electrons and ions (plasma) fly around but eventually hit the magnet
Tokamak:
Torus/donut shape with magnets on the outside
Charged plasma cycles inside
Reactor
Steady progress in power generation over the past 50 years
ITER reactor: international collaboration on fusion reactor design
Tokamak design
Goal: 500 MW of fusion power (10x input)
~50 private companies
Tokamak Challenge: collapse in stability of the plasma’s motion pattern
Stable: clean donut pattern of plasma inside to Tokamak
Unstable (~1s time): donut shape starts curling out of symmetry
Unstable mode growth: (<<1s): twist fully out of shape, hit the side of the tokamak, collapsing plasma and damaging the machine
To control shape of plasma
Must model its behavior
Monitor its state
Suggest adjustments to magnetic field to maintain its shape stably
Role for ML: monitoring plasma state
Challenge:
Experimental reactors have many diagnostic sensors
Industrial reactors will have far fewer
What is the state of the plasma?
Infer 6 1-D profiles
Use differential equations of plasma flow
Fit the state of the equations to the measurements from the diagnostics
Equations are complex and non-linear
Use simplified equations to fit it
Use an ML surrogate model of the original (non-simplified) physics model to speed it up
EQNet: faster and more robust
Captures the fully detailed physics
Much faster than simplified physics
Neural surrogate also provides a gradient
Enables much faster model inverses
Linear State Space System (PertNet): used in control
Surrogate is used in real-time for online plasma control
CAKE => RTCAKENN
Uses Keras2C: https://github.com/PlasmaControl/keras2c
Individual sensors can fail
Traditionally: many if-then-elses to decide how to adjust for each sensor failure
Neural nets discover correlations between different sensors
If some sensors become noisy or fully fail, the neural net can predict their measurements based on other sensors
Using ML to improve robustness to channel noise and loss
Channel: direction of measurement through the plasma
Given measurements through some channels, can reconstruct it based on the others
Measurements can be for different quantities (e.g. temperature vs density)
Can use this to find the minimal set of diagnostics needed
Multi-Modal ML for Synthetic super-resolution
Given low-res measurements can reconstruct high-res images
E.g. from Khz-frequency measurements to Mhz-frequency
Use-case infer fine-scale physical structures in plasma that cannot be directly imaged from sensors
Replaying fusion shots
Can relate real measurements to predicted reactor behavior
Predicting plasma behavior via simulations
Challenge: simulations trained on smaller fusion reactors are not very effective for predicting for larger reactors (extrapolation)
Approach: combine simulations with data about the larger reactors
Experiment: use data from initial (low-confinement) experiments on DIII-D, predict for high-confinement experiments using simulations and the data
This works ok but still not great; simpler methods for combining data + simulations most effective given the predictive noise (more complex non-linear methods overfit)
Realtime Adaptive ML Plasma Model via Reservoir Computing
Reservoir: randomly connected layers + final trained linear layer
Fast to train, can be done on CPU in realtime
Thus, easy to adapt to new datasets
Enables fast adaptive control
Keras2C: https://github.com/PlasmaControl/keras2c
Python is challenging for real-time operation
Library converts Keras neural network code to C code
Tools make it possible to achieve much larger confinement / power output in two different test reactors: DIII-D and KSTAR
Combining controller and measurement models
Use reinforcement learning
RL requires a cheap, accurate simulator of plasma behavior
Instead of simulations use data-based models instead as the RL environment
This makes it possible to control the reactor more effectively
LLM as operational copilot
GPT model takes in siloed data: turns into vectors
Train model to answer fusion-related questions based on similarity of vectors of questions and text in databases
Good for training/assisting people who are not yet experts
Uncertainty Quantification Toolbox: https://github.com/uncertainty-toolbox/uncertainty-toolbox