Applications of Machine Learning and Data Science to predict, design and improve laser-fusion implosions for inertial fusion energy

Despite many years of laser-driven inertial confinement fusion (ICF) research, there is not a clear path to the high energy gains required for inertial fusion energy.  Typical high gain designs for both direct and indirect drive ICF use low adiabat (i.e. entropy) laser pulse shapes to achieve high compression and high areal densities as indicated by multi-dimensional rad-hydro simulations. However, when fielded on the NIF and OMEGA laser facilities, these designs perform poorly and very differently from the simulation predictions. On the other hand, we have learned over the past 13 years of experiments at the NIF, how to empirically improve target designs leading to a ~3000x increase in target gain from 5x10-4 (~1kJ of fusion energy output) in the first NIF implosions in 2011 to 1.5x (~3MJ of fusion output) on December 5th 2022. A large increase in fusion yield (about 15x) was also achieved on OMEGA over the past ten years.   However, such designs use high adiabat pulse shapes leading to low target gains and they are not suitable for inertial fusion energy.

 The goal of the proposed research is to develop a data-enhanced predictive capability and real-time control and optimization for laser fusion using advanced machine learning (ML) algorithms trained on simulation databases and re-trained on experimental databases or trained directly on experiments with simulation support. This project will use the OMEGA experimental database as training data as well as the simulation databases from the LLE radiation hydrodynamic codes LILAC and DRACO. This work will build upon the early successes of a data-driven statistical model trained on experiments developed by the Rochester’s group [Nature 565, 581 (2019), Phys. Rev. Lett. 127, 105001 (2021)] and will focus on direct-drive laser fusion. The results and the ML framework will also be applicable to indirect drive laser fusion as well as other fusion schemes and possibly other fields of science.

The results will be used to identify a viable path towards high energy gains. This predictive capability will be based on AI/ML tools that combine the results of rad-hydro simulations and experimental data including three dimensional and time-resolved diagnostic measurements.

It will include these major AI/ML applications to laser fusion: (1) Deep Neural Networks (DNNs) for Transfer Learning including multimodal diagnostic data, (2) Statistical modeling using DNNs trained on experiments using basis variables from simulations, (3) Solutions to Inverse Problem for ICF implosions using DNNs/Generative AI (i.e. transformers), (4) Real-time Implosion Performance Optimization via Enhanced Gaussian Process Modeling and Reinforcement Learning to leverage prior knowledge unlike traditional Bayesian optimization, (5) ML-guided Real-time Symmetry Control.  This predictive capability will include qualified uncertainties and will be first applied to the design of DT-layered implosion experiments on the OMEGA laser.

This collaborative proposal brings together ICF physicists from the UR Laboratory for Laser Energetics (LLE) and AI/ML experts from the UR Department of Computer Science and from Hewlett Packard Enterprises.