Self-driving laboratories (SDLs) are behind an increasingly popular experimental paradigm that integrates robotics, remote data collection, and AI analysis to perform scientific experiments autonomously in real time. Prime candidates for SDL setups are laser-based, high-repetition-rate (HRR, 10+ Hz), high energy-density (HED) platforms where scientists are performing cutting-edge research on laser–plasma interactions and inertial fusion energy (IFE).
Current estimates suggest that an IFE facility would need to cycle through ten deuterium–tritium fuel targets per second to reach the energy yields required of a reliable, large-scale power plant. However, LLNL’s National Ignition Facility (NIF)—the only ICF facility to achieve net target gain so far—is a single-shot facility. Research teams from across the globe vie to reserve narrow time slots on the system, and engineers require hours of preparation to painstakingly set up the target and diagnostics for a reaction that lasts a tiny fraction of a second. Afterward, these teams run analyses on the hundreds of gigabytes of data produced by the experiment to glean scientific insights, benchmark simulation models, and help them better prepare for a subsequent shot by adjusting certain experimental parameters. With experiments occurring at a rate of ten times per second, humans can no longer drive the cycle.
Instead, researchers are increasingly looking at self-driven setups to autonomously accelerate the fine-tuning of experimental parameters—for instance, laser pulse shape, intensity and duration—without having to pause experimentation. High-repetition-rate facilities are not only necessary for the eventual goal of IFE as a reliable power source, but HRR experiments could also further scientific progress at a never-before-seen scale. One of the barriers to this future is the ability to aggregate and analyze terabytes of high-fidelity scientific data every second in order to rapidly reconfigure experimental parameters and meet physics-based objectives.
While AI is an obvious choice to accomplish such rapid data analysis and recommend parameter adjustments, researchers first have to ensure the machine learning (ML) models the AI uses are robust and will produce quality recommendations without compromising the laser driver. Researchers often use digital simulations of a laser system’s operation to generate synthetic data that is used to train ML models. However, these simulations, by definition, only approximate the laser’s ideal operating conditions; they thus lack the statistical randomness that is characteristic of real-life systems.
“Variability comes from many places, but especially the operating environment. For instance, performing an experiment with the exact same parameters could yield slightly different readings from one shot to the next due to changes in temperature, humidity, instrument alignment etc.” says Abhik Sarkar, researcher in the Center for Advanced Computing (CASC) and one of Sidekick’s developers. Without factoring in this natural system variability, an ML algorithm trained to reconcile different parameters might confidently recommend settings that ultimately prove non-optimal.
To address these challenges, a team of researchers—including four from LLNL—developed a tabletop experimental setup, Sidekick, which emulates the shot behavior of real-world HED facilities. Sidekick will help researchers determine their data aggregation strategies and instrument control techniques via EPICS, an open-source library of control systems software hosted by Argonne National Laboratory, without firing a single shot at the facility. Further, Sidekick’s hardware-in-the-loop approach incorporates real-world variability from a paired laser and photodetector which enables more thorough testing of the ML models involved.
Exhibiting Sidekick at the Supercomputing 2024 conference, the team engineered a simplified shot environment for pitting researchers’ human abilities for pulse shaping against those of a specially trained AI. Participants were asked to replicate a randomly selected pulse profile by altering five “synthetic laser” parameters. Given the range, precision and hard to generalize behavior of each parameter, users quickly realized the difficulty of the task. After many unsuccessful attempts by participants, the team’s AI model was finally given a chance to rapidly adjust shot parameters in response to observed shot profile changes—a method called closed-loop pulse shaping. This system’s repeated success at rapidly converging to the optimum parameter settings illustrated the value of AI in developing self-driven, HRR labs investigating HED science, potentially paving the way towards IFE.
Sidekick’s development team from LLNL includes Abhik Sarkar, Derek Mariscal, Mason Sage, and Aldair Gongora. Development partners include Scott Feister, instructor at California State University, Channel Islands, and NVIDIA.
