Using Sparse Gaussian Processes for Predicting Robust Inertial Confinement Fusion Implosion Yields

Peter Hatfield, S. J. Rose, R. H. H. Scott, Ibrahim Almosallam, Stephen Roberts, M. J. JarvisIEEE Transactions on Plasma Science 2019 · DOI · arXiv

Abstract

Here, we present the application of an advanced sparse Gaussian process-based machine learning algorithm to the challenge of predicting the yields of inertial confinement fusion (ICF) experiments. The algorithm is used to investigate the parameter space of an extremely robust ICF design for the National Ignition Facility, the “Simplest Design”; deuterium–tritium gas in a plastic ablator with a Gaussian, Planckian drive. In particular, we show that: 1) GPz has the potential to decompose uncertainty on predictions into uncertainty from lack of data and shot-to-shot variation; 2) it permits the incorporation of science-goal-specific cost-sensitive learning (CSL), e.g., focusing on the high-yield parts of parameter space; and 3) it is very fast and effective in high dimensions.

Introduction

Inertial Confinement Fusion (ICF) involves using high-power lasers to compress fuel capsules to achieve nuclear ignition. Despite significant efforts at facilities like the National Ignition Facility (NIF), matching experimental yields with theoretical predictions remains a major challenge. The complexity of the physics involved necessitates the use of "surrogate models"—machine learning emulators that can quickly explore vast parameter spaces where full-scale radiation-hydrodynamics simulations are too computationally expensive.

Traditional machine learning models often struggle with two critical aspects of ICF research: quantifying uncertainty and prioritizing high-yield "success" regions. Standard models typically treat all errors equally, but in fusion, accuracy near the "ignition cliff" is far more valuable than in low-yield failure regimes. Furthermore, experiments exhibit inherent "shot-to-shot" variation (aleatoric uncertainty) which must be distinguished from uncertainty arising from a lack of training data (epistemic uncertainty).

Methodology: Sparse Gaussian Processes (GPz)

This study utilizes GPz, a sparse Gaussian Process framework designed for efficiency and sophisticated uncertainty modeling. While standard Gaussian Processes offer elegant mathematical properties, their $O(n^3)$ scaling makes them impractical for large datasets. GPz employs a sparse framework with $O(nm^2)$ scaling, enabling it to handle thousands of data points and high-dimensional spaces rapidly.

Key innovations in the GPz approach include:

  • Heteroscedastic Noise Modeling: Unlike models assuming constant noise, GPz predicts a unique variance for every point in the parameter space.
  • Cost-Sensitive Learning (CSL): By weighting the loss function (e.g., $w = \sqrt{Y}$), the model is forced to prioritize accuracy in high-yield regimes.
  • Uncertainty Decomposition: The framework mathematically separates epistemic uncertainty (lack of data) from aleatoric uncertainty (intrinsic shot-to-shot variation).

Experimental Setup and Datasets

The researchers focused on the "Simplest Design," a robust ICF configuration using deuterium-tritium gas within a plastic shell. The primary dataset consisted of 5,000 simulations from the HYADES radiation-hydrodynamics code, covering a 5D parameter space including peak temperature, pulse width, gas radius, shell thickness, and gas density. To evaluate scalability, a 10-dimensional analytic "mock-yield" model with 100,000 data points was also tested.

Results and Discussion

GPz demonstrated exceptional speed, training in approximately 5 seconds compared to 80 seconds for DJINN (a specialized neural network). In high-dimensional tests, GPz was up to 70 times faster at making predictions.

Predictive Accuracy and Cost-Sensitive Learning

The application of Cost-Sensitive Learning (CSL) significantly improved performance in the high-yield regime (logY 15–17). As shown in Figure 1, the GPz CSL model exhibits tighter alignment with the ground truth in the critical high-yield region compared to the standard GPz Normal or the DJINN neural network. In fusion research, a model that is highly accurate about a potential breakthrough is far more valuable than one that is slightly wrong about a total failure.

Comparison of GPz Normal, GPz CSL, and DJINN models against Hyades ground truth. The CSL approach improves accuracy in high-yield regions.

Uncertainty Decomposition

A major advantage of the GPz framework is its ability to decompose uncertainty. Figure 2 illustrates how the model maps predicted yield alongside total uncertainty and the specific source of that uncertainty. This allows researchers to distinguish between regions where more data is needed (high epistemic uncertainty) and regions that are inherently unstable or chaotic (high aleatoric uncertainty). This capability is a game-changer for experimental design, identifying where additional simulations or experiments will be most impactful.

Heatmaps showing predicted yield, total uncertainty, and the source of uncertainty across the parameter space.

Conclusion

The GPz algorithm provides a powerful tool for ICF research, offering fast, accurate predictions with sophisticated uncertainty quantification. By identifying stable, high-yield designs that are robust to manufacturing perturbations, this approach can guide more effective experimental campaigns. Future work will focus on extending these models to 2D and 3D simulations and utilizing transfer learning to integrate real-world experimental data from the NIF.

Summary updated

Abstract

Here, we present the application of an advanced sparse Gaussian process-based machine learning algorithm to the challenge of predicting the yields of inertial confinement fusion (ICF) experiments. The algorithm is used to investigate the parameter space of an extremely robust ICF design for the National Ignition Facility, the “Simplest Design”; deuterium–tritium gas in a plastic ablator with a Gaussian, Planckian drive. In particular, we show that: 1) GPz has the potential to decompose uncertainty on predictions into uncertainty from lack of data and shot-to-shot variation; 2) it permits the incorporation of science-goal-specific <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">cost-sensitive learning (CSL)</i> , e.g., focusing on the high-yield parts of parameter space; and 3) it is very fast and effective in high dimensions.
BibTeX
@article{hatfield2019using,
  author = {Peter Hatfield and S. J. Rose and R. H. H. Scott and Ibrahim Almosallam and Stephen Roberts and M. J. Jarvis},
  title = {{Using Sparse Gaussian Processes for Predicting Robust Inertial Confinement Fusion Implosion Yields}},
  journal = {IEEE Transactions on Plasma Science},
  year = {2019},
  doi = {10.1109/tps.2019.2944416},
}
Figure