Multi-Agent Design Assistant for the Simulation of Inertial Fusion Energy

TL;DR

The paper tackles the challenge of designing inertial confinement fusion capsules under extreme physics and uncertainty by proposing a Multi-Agent Design Assistant (MADA) that orchestrates planning, simulation, and machine-learning-based emulation. MADA builds a full-field physics emulator (Professor) from autonomous simulation runs and uses natural language to drive inverse design of capsule geometry, x-ray drive, and layer structure. The authors demonstrate interactive and autonomous design loops that produce high-fidelity emulation outputs and use visual feedback to drive optimization toward ignition, crossing the Meldner threshold in $T$–$\rho R$ space. This framework points toward AI-driven control and autonomous discovery for future IFE power plants.

Abstract

Inertial fusion energy promises nearly unlimited, clean power if it can be achieved. However, the design and engineering of fusion systems requires controlling and manipulating matter at extreme energies and timescales; the shock physics and radiation transport governing the physical behavior under these conditions are complex requiring the development, calibration, and use of predictive multiphysics codes to navigate the highly nonlinear and multi-faceted design landscape. We hypothesize that artificial intelligence reasoning models can be combined with physics codes and emulators to autonomously design fusion fuel capsules. In this article, we construct a multi-agent system where natural language is utilized to explore the complex physics regimes around fusion energy. The agentic system is capable of executing a high-order multiphysics inertial fusion computational code. We demonstrate the capacity of the multi-agent design assistant to both collaboratively and autonomously manipulate, navigate, and optimize capsule geometry while accounting for high fidelity physics that ultimately achieve simulated ignition via inverse design.

Figures (10)

• Figure 1: The Multi-Agent Design Assistant (MADA) acts both as an aid to the human ICF designer within an interactive environment and as an autonomous researcher capable of action under broad direction.
• Figure 2: Prompt-response pairs for several example queries to the agent system.
• Figure 3: Surrogate model produced by Professor, an ML-based multiphysics emulator. Professor is trained to predict full-field data (i.e. r-t plots of density, pressure, etc.) and time varying scalars (average fuel temperature and areal density) from input design parameters (e.g. layer thicknesses). The model was trained using results from 3000 simulations. The predictions are summarized using r-t plots and state-space plots (fuel temperature vs areal density).
• Figure 4: In (a) we show the iterative approach converging toward the burn region. For (b), (c), and (d), The blue curve represents the Meldner burn threshold; black traces show the T–$\rho R$ trajectory for a given parameter configuration. (a) shows the trajectory corresponding to the best-performing sample identified in the initial iteration, (b) shows the best sample in the final iteration, and (c) shows the plot corresponding to the globally best parameter configuration.
• Figure 5: User prompt initiating a batch optimization task. The user provides a YAML configuration file and instructs the agent to generate $20$ configuration and explain the sampling methodology.