Session GO10: ICF: Machine Learning

Live

Machine learning (ML) is a promising tool for predicting the performance of fusion experiments. Unfortunately, deep learning models require significant amounts of synthetic data to predict the performance of low shot rate experiments such as the National Ignition Facility (NIF). Here, we present an alternative approach based on shallow ML models trained exclusively on experimental data. We narrow our focus to predicting neutron yield, an important performance metric that is notoriously difficult to estimate. Although the dataset includes fewer than 150 shots, each with widely-varying experimental set-ups, we demonstrate a ML model that predicts the logarithm of neutron yield with an average error of ~10% using only a priori knowledge. Another model achieves an average of ~7% error utilizing X-ray diagnostic metrics in addition to a priori knowledge. These models can be used to roughly estimate the neutron yield of proposed NIF shots and identify the relative significance of shot parameters.   

Live

Statistical models of cryogenic implosions on OMEGA have been used to increase performance on the laser by relating the outputs of 1-D codes to previous experimental results. Here, we conduct a similar exercise on a synthetic dataset of 1-D and 3-D simulation codes and show that the results from a systematically perturbed simulation can be reproduced by a statistical model trained on 1-D codes. We also find connections between the inferred relationships in the synthetic data set and real data set that suggest the physical origins for degradation sources on OMEGA. Uncovering trends in the observables and comparing trends in measured data with synthetic data, has enabled to identify the dependencies of the fusion yield on the ion temperature asymmetries from the l$=$1 mode and on the laser beam to target ratio.   

Live

Magnetized Liner Inertial Fusion (MagLIF) is a magneto-inertial fusion concept relying on quasi-adiabatic heating of a gaseous D-D fuel and flux compression of a pre-imposed axially oriented magnetic field to reach fusion relevant plasma conditions. Calculations show that up to $\sim1000\times$ flux compression is possible, sufficient for trapping charged fusion products and reducing electron thermal conduction. However, physical mechanisms such as resistive diffusion and the Nernst effect may cause magnetic flux to leave the fuel, potentially causing performance degradation. As a result, quantifying fuel magnetization is critical for understanding performance. Recently it was shown that yield and time-of-flight measurements of primary D-D and secondary D-T fusion neutrons are sensitive to the magnetic field-fuel radius product ($BR$). Yet analysis of experimental data is time consuming, requiring significant user input, and is somewhat lacking in rigorous uncertainty quantification. We present a deep-learning approach based on the aforementioned diagnostics within a Bayesian framework that provides uncertainty quantification. We analyze several MagLIF experiments, showing an indication of the importance of the Nernst effect as laser preheat of the target is increased.   

Live

Experiments in inertial confinement fusion (ICF) and high energy density physics (HEDP) rely heavily on non-scalar diagnostics like images and spectra. Traditionally, comparing these data with simulations requires they are first `featurized' into scalar quantities so that a standard metric like chi-squared can be used. This process requires significant user input and has the potential to introduce bias and/or loss of information, reducing the utility of the diagnostics in constraining the parameters of interest. In this talk we will present the `Bayesian Super Postshot', which uses state-of-the-art deep learning to extract important features from an entire experimental dataset. This new capability produces the best possible constraints on unknown drive and physics parameters in ICF experiments by directly matching simulated X-ray images, FNADs, line-of-site resolved data and multiple scalar quantities. We will present results for recent high-performance ICF implosions at the National Ignition Facility and give a discussion of the new constraints on simulations that come from our full treatment of diverse experimental data types. Prepared by LLNL under Contract DE-AC52-07NA27344. LLNL-ABS-812146   

Live

Design improvements of the Bigfoot N180128 capsule has led to a series of proposed designs known as SQ-1 through SQ-5. In recent work by J. A. Gaffney and collaborators the degradation mechanisms of N180128 were inferred using a Bayesian-Super-Postshot analysis in an ensemble of approximately 100,000 2D Hydra calculations. The inferred posterior distributions constrain the flux asymmetries such as P2-swings caused by the growth of the gold bubble inside the hohlraum, among others. Assuming the degradation mechanisms are approximately the same for a scaled up version of N180128 and for the improved SQ designs we can estimate the statistical mean and variance in observables such as the neutron yield, DSR, and ion temperature, without having to resort to a new computationally expensive ensemble calculation. We use the concepts of arbitrary polynomial chaos expansions (aPCE) to create a set of orthogonal polynomials with respect to each probability distribution in the 4-dimensional input space. The roots of the polynomials determine the location of the desired Hydra input parameters. A modest set of Hydra runs totaling a few hundred points can lead to a reasonable estimate of the mean and variance of the observables for SQ-1 and a hydrodynamically scaled N180128 capsule.   

Live

The development of accurate reduced plasma models is crucial to enable predictive and computationally efficient multi-scale models for Inertial Confinement Fusion (ICF) and High-Energy-Density (HED) plasma dynamics. Computationally intensive fully kinetic plasma simulations play a pivotal role in our understanding of the complex nonlinear physics from first-principles, but distilling reduced plasma models from such simulations remains an outstanding challenge. Here we show how sparse regression techniques can be used to uncover reduced plasma physics models [in the form of interpretable partial differential equations (PDEs)] directly from the data of fully kinetic particle-in-cell (PIC) simulations. We introduce an integral formulation for the discovery of PDEs and demonstrate that it is critical to enable robust inference from noisy data associated with PIC and other particle-based approaches. We discuss how this methodology can complement traditional analytically derived reduced models, and bring important advantages to the discovery of kinetic-fluid closure models for ICF and HED plasmas.   

Live

Neural network surrogates of the computer simulators begin to play an important role in the uncertainty quantification of the predictions of the ICF experiments, as well as in the design optimization. In this paper, we address three big challenges in building reliable surrogate models: (i) simulation-experiment bias, which needs to be suppressed given sparse experimental data, (ii) incorporating multimodal diagnostic data to better constrain the surrogate model, and (iii) inferring unknown inputs for the indirect-drive, high-resolution capsule simulations, for which we build the surrogate model. The first challenge is addressed by partial retraining of the simulation-trained model to match the experimental data. The second challenge is addressed by compressing different types of diagnostic data into a small set of variables using an autoencoder network, which takes advantage of the correlations in the data. The third challenge is addressed by iteratively improving the surrogate model and the inferred capsule inputs. We demonstrate that our technique brings the error of scalar observable predictions to within the experimental error and corrects major errors in the predictions of the experimental X-ray images.