Session TO07: Joint ICF & MFE: Machine Learning and Data Science Technologies

Artificial Intelligence-assisted control of Alfvén Eigenmodes improves plasma stability in the DIII-D tokamak

Alfvén Eigenmodes (AEs) are an important challenge in the development of sustainable fusion energy, as they can significantly impact the confinement and stability of energetic particles in fusion plasmas. This work presents a novel approach for real-time feedback control of AEs using multiple neutral beams at the DIII-D National Fusion Facility. Building upon prior predictive models that utilize electron cyclotron emission diagnostics and neutron deficit analysis, this study introduces a new feedback control system that actively modulates individual neutral beams in real-time to suppress AE activity [1,2]. The system dynamically adjusts the beam power of each beam based on real-time measurements of AE signatures. Real-time integration of Reservoir Computing Networks (RCN) predicted the neutron rate, enabling the controller to follow Reversed Shear Alfvén Eigenmode (RSAE) activity within the constraints of neutral beam injection (NBI) programming. This work also discusses the implementation of alternate actuators, such as gas puffing (density control) to enhance the AE mitigation methods. Experimental results demonstrate the effectiveness of this multi-beam feedback approach, marking a milestone in the understanding and mitigation of AEs, and paving the way for improved stability and confinement in future fusion reactors.   

Combining physics-based simulations and experimental data from multiple machines to predict and control tokamak profile evolution

Hominid tokamak scientists and operators combine experimental experience and physical models to guide decision-making in machine design and control. Methods for "data fusion" are beginning to provide practical methodologies to build AI models mimicking this human process: taking advantage of both the generalizability of physical models and the quantitative accuracy of experimental results in a single model. For the task of tokamak plasma profile prediction, a variety of such methodologies are presented: (1) multi-machine learning exploiting non-dimensionalization, (2) providing interpreted context from simulations as additional input to machine learning models, (3) transfer learning from simulation to experimental data, and (4) meta-learning (akin to stacked generalization) by combining physics-based and empirical models on equal footing. It is demonstrated that, for the task of extrapolating plasma profile predictions from low- to high-plasma current DIII-D scenarios, a meta-learned profile-predictor using ASTRA/TGLF physics simulations and data is more accurate than a model built on physics or data alone. Applications of the methodology to the task of commissioning a new reactor such as ITER are discussed.   

Machine Learning model for real-time SPARC vertical stability observers

Given the demanding requirements of the SPARC high-field tokamak (B₀=12.2 T) and its operation with high elongated plasma (κ_sep=1.97), robust real-time-compatible vertical stability observers are paramount. In this work, we present fast ML surrogate modeling for observers (such as VDE n=0 growth rate, stability margins ​​(m_s), inductive stability margins (m_i), max-Z, & frequency components), integrating advanced 2D electro-mechanical circuits and dynamic plasma response models [1]. These surrogate observers employ transformer-based ML networks; trained to replicate and predict the results of filamentary rigid body MEQ-RZIp and deformable non-linear MEQ-FGE (and its linearized version FGElin) plasma response models [2]. The training dataset incorporates simulated SPARC primary reference discharge scenarios and the C-Mod (hot VDEs) 2012-2016 disruption database. These observers will also be trained over a range of simulated L- and H-mode scenarios, including periods without and with ELM (via artificial vertical kicks [2]). We will report on the assessment of observers' sensitivity to the RZIp & FGE models and their proximity to stability boundaries. Conclusions regarding  SPARC's vertical stability control are further employed to inform ARC design and its operational max-Z stability limits.

1. Walker & Humphreys (2006), 50:4,473-489

2. Carpanese et al. 2020 EPFL PhD thesis no. 7914

3. Sartori et al Proc. 35th EPS Conf. on Plasma Physics vol 32D P5.045   

Results and Lessons Learned from the "Accelerating Radio Frequency Modeling Using Machine Learning" Project

Our SciDAC machine learning (ML) project focused on accelerating radio frequency (RF) computational models to enable real time feedback control such as needed in advanced tokamak (AT) scenarios. Lower hybrid and high harmonic fast fasts are two types of waves used for current profile control in tokamaks. The former typically uses geometric options (ray tracing) coupled with a Fokker-Planck code to simulate the current deposition profile. The later with its longer wavelengths uses physical options (full wave) also coupled with Fokker-Planck. This project created surrogate models with machine learning to reduce simulation time to the order of milli-seconds (ms) enabling real time control and integrated predictive models. Models are capable of predicting heating and current drive profiles (the forward problem), determining wave parameters for a desired deposition profile (the inverse problem), and mapping diagnostic measurements to core deposition profiles (the lateral problem). Training data was generate from simulation parameters in a hypercube sampling method. Three classes of ML models (Random Forest Regression, Gaussian Process Regression, and Multi-Layer Perceptron) trained on the databases all show excellent performance. All three techniques show good accuracy and speed. We will report on results and discuss future steps including models applicable to a wide class of tokamaks and applications of Generative AI and surrogate models for operators.   

Simulation-Based Inference of High Field Side Scrape-Off Layer Filament Characteristics using Profile Reflectometry

A stochastic scrape-off layer filament model was used to infer probable filament characteristics, such as density, velocity, and size, using high-field side profile reflectometry data on DIII-D. In general, scrape-off layer (SOL) turbulence is characterized by intermittent filament structures which radially propagate from the last-closed flux-surface (LCFS) to the vessel wall. While filaments are primarily measured in the low-field side SOL, they may also be found on the high-field side (HFS), depending on the magnetic geometry. Such filaments could negatively impact the recently installed DIII-D high-field side lower-hybrid current drive experiment due to undesired wave-filament interactions. A HFS reflectometer on DIII-D is used to measure SOL density profiles, and using simulation-based inference (SBI), information about filament parameters from profile reflectometry can be obtained. This technique utilizes a stochastic JAX-based filament model to simulate the reflectometer response to filament structures. Automatic differentiation is used to calculate reflectometer group delay measurements, while vectorization on a GPU dramatically improves computational efficiency. SBI is then used to estimate the posterior of multiple filament parameters such as filament size and density. The technique is applied to DIII-D HFS SOL reflectometry measurements, and predicted impact on HFS LHCD is discussed.   

Physics Informed, Automated and Highly Parallel Bayesian Optimization of Direct-Drive Implosions

Finding the optimal implosion design on existing experimental facilities for inertial confinement fusion requires an exhaustive search of the vast design parameter space. This is infeasible both with experiments and simulations. Consequently, a large fraction of the experimentally realizable design space remains unexplored, and new design schemes are challenging to optimize in a reasonable time-frame. On the OMEGA laser facility, predictive machine learning models have been developed to accurately forecast the result of an experiment using only inexpensive simulations and the large dataset of prior experimental data. However, the full design space remains vast enough to be unassailable with simple optimization techniques. Here, we develop a new physics-informed and optimally parallel Bayesian Optimization algorithm that can entirely optimize the target and pulse shape of a direct-drive ICF implosion under a given design paradigm. We use this algorithm to find a markedly improved design for the performance implosions on OMEGA that is predicted to hydro-equivalently scale to ignition at 2.15 MJ.   

Optimizing Cylindrical Targets for Neutron Yield Using Multi-Fidelity Modeling Techniques

Design studies of inertial confinement fusion targets can present computational challenges, as they often require many calls to multi-dimensional radiation-hydrodynamics codes to iterate on a design and accurately capture implosion physics. Surrogate models, which replace simulation results with a simplified approximation, have been successfully leveraged in the past to reduce the computational expense of such design studies. We apply Gaussian process based surrogate models coupled with Bayesian optimization to optimize the design of a cylindrical target containing deuterium-tritium fuel for yield. Past work, which focused on the optimization of Gaussian process surrogates trained exclusively on output from 1D simulations, revealed that optimal designs selected in this manner exhibited a substantial loss in yield when simulated in 2D. Despite their lower prediction accuracy, 1D simulations are less expensive than their 2D counterparts. To improve the predictive performance of the surrogate while maintaining low costs, we introduce a multifidelity optimization algorithm that integrates data from 1D and 2D simulations to identify target designs that maximize yield. We compare the costs and predictive accuracy of the multifidelity optimization method with those of an optimization approach which solely relies on high-fidelity data. The design selected by the algorithm is discussed, emphasizing the design choices and implosion physics responsible for the target's improved performance.   

Exploring robust, high yield ICF designs using Bayesian optimization

Inertial confinement fusion (ICF) experiments rely on complex multi-physics simulations such as the LLNL-developed HYDRA to guide design work. However, these simulations can be expensive and have several dozen design parameters, making the search for an optimal design difficult. Recently developed automated tools utilize multi-fidelity Bayesian optimization to search these high-dimensional design spaces for candidate experiments.

 

In this project, we tune the Bayesian optimization algorithm to optimize ICF designs for robustness and high yield. The starting point for this search is an asymmetrical implosion found by Peterson et al. [1]. This “ovoid” design uses vortical flows to suppress instabilities along the capsule surface during implosion. We present the computational approach to further optimize and investigate the ovoid implosions. The optimization tools run 2D integrated simulations in HYDRA to converge on an optimal design. We extract valuable physics information by running detailed simulations of the optimal design and designs in its vicinity.

[1] Peterson, J. L., et al., Phys. Plasmas 24, 032702 (2017)

LLNL-ABS-866070   

Optimizing the Performance of Direct-Drive Implosion Experiments Using Meta-Bayesian Optimization

Finding a laser pulse shape that optimizes the Lawson parameter [1,2,3] for a given target is a challenging problem in inertial confinement fusion due to the predictive gap between simulations and experiments. The Lawson parameter is typically related to the yield and ρR of the implosion and requires an increase in both. Optimizing the yield of cryogenic implosions on OMEGA using a data-driven predictive machine-learning (ML) approach [4,5] has met with considerable success, but increasing the ρR has proven more challenging. It is likely that this is in part due to hydrodynamic instabilities, but is likely also due to the increased sensitivity of the ρR to fine details of the shock timing and entropy spatial profile of the implosion, which in turn are highly sensitive to the front end of the laser pulse. If simulations used for implosion design [6] fail to capture the instability growth, shock transit, or adiabat-setting behavior of the implosion correctly, the response surface between simulations and experiments will sharply differ, making implosion optimization challenging with limited experimental data. We present the use of Neural Acquisition Processes (NAP) [7] which is meta-learned on varying fidelities of simulation databases to optimize synthetic experiments and Omega experiments in a sample efficient manner. NAP uses a transformer neural network to learn the input-output distribution and enables proximal policy optimization reinforcement learning based acquisition function which significantly outperforms Bayesian optimization. Additionally, we present algorithmic improvements of meta-learning variants of decision transformers [8] and use them to solve similar ICF implosion optimization objectives.   

Experimental Demonstration of 3D Hot-spot Shape Symmetry Control in Laser Direct-Drive Inertial Confinement Fusion Implosions

The OMEGA laser facility has demonstrated the feasibility of achieving symmetric implosions through corrective on-target laser energy adjustments. The symmetry control experiment applied a real-time (between shots) machine-learning based 3D reconstruction of hot-spot plasma emissivity from multi-angle x-ray imaging. This enabled quantification of the magnitude and orientation of low-mode (L = 1-2) asymmetries, which were separately seeded by a 40-μm target offset and a shimmed shell with 3-μm thickness variation along the polar axis. The large L = 2 prolate asymmetry was successfully mitigated, while the mitigation of L = 1 asymmetry was also observed. The methodology of controlling 3D low-mode hot-spot shape asymmetries through on-target laser energy adjustments and 3D reconstruction is being integrated with an artificial intelligence system, aiming to achieve symmetric and high-performance implosions. Three major frameworks are developed, including an evolutionary optimization that generates optimal laser pulse shapes achieving high measured yields of 1.82E14, a convolutional neural network model for fast (between shots) 3D tomography of hot-spot and shell structures, and a quantum-inspired data analysis model identifying correlations between input non-uniformities and measured hot-spot flow asymmetries.   

AI-assisted prediction of laser-plasma instabilities for inertial confinement fusion

Predicting and controlling hot electron generation from laser-plasma instabilities(LPI) is a critical challenge in direct-drive inertial confinement fusion (ICF). In this project we leverage generative AI to develop physics-informed ignition-scale LPI packages that can be incorporated into ICF design codes. We will present preliminary results utilizing generic large language models (LLMs) to model hot electron generation and employ diffusion-model-based scientific simulation methodologies as an alternative to costly particle-in-cell (PIC) simulations. Additionally, we will address the trustworthiness of these generative AI models.    

Comparison of Mo versus W for Double Shell Target Capsules using Machine Learning Optimization

Double shell targets are an alternative ignition platform for inertial confinement fusion [1]. The inner shell materials of interest for double shell targets are molybdenum and tungsten. Molybdenum has a lower density that could produce a more stable implosion, while tungsten has a higher density that could provide more compression and radiative trapping. Currently, there has not been enough comparison between optimized designs for these two inner shell materials. Our previous work has focused on developing a multi-fidelity Bayesian optimization framework to find yield optimized double shell target geometries [2,3,4]. In this work, we apply the multi-fidelity Bayesian optimization to find optimal, or near optimal, double shell targets with molybdenum and tungsten inner shells for a 1.25 MJ laser drive using “pre shot” xRAGE simulations [5]. The optimized targets for each inner shell material are compared to better understand the physics driving the implosion. Analysis of the simulations used in the study show trends in designs that contribute to high yields, ion temperatures, and fuel areal densities. 

References

[1]  D. Montgomery, et al., “Design considerations for indirectly driven double shell capsules,” Physics of Plasmas 25, 092706 (2018).

[2] N. Vazirani, et al. “Coupling 1d xrage simulations with machine learning for graded inner shell design optimization in double shell capsules,” Physics of Plasmas 28, 122709 (2021).

[3] N. Vazirani, et al. "Coupling multi-fidelity xRAGE with machine learning for graded inner shell design optimization in double shell capsules." Physics of Plasmas 30.6 (2023).

[4] N. Vazirani, et al. "Bayesian batch optimization for molybdenum versus tungsten inertial confinement fusion double shell target design." Statistical Analysis and Data Mining: The ASA Data Science Journal 17.3 (2024): e11698.

[5]M. Gittings, et al. “The RAGE Radiation-Hydrodynamic Code,” Computational Science Discovery 1 (2008), https://doi.org/10.1088/1749-4699/1/1/015005.