Abstract
Differentiable, Multi-Fidelity Simulation for Integrated Inertial Fusion Energy Design
Archis Joglekar
Pasteur Labs
Designing a viable inertial fusion energy (IFE) power plant requires reasoning across coupled subsystems, from laser pulse shaping and target implosion physics, through chamber engineering and tritium breeding, to techno-economic viability. Addressing these subsystems in isolation leads to designs that are locally optimal but globally inconsistent.
We present a fully differentiable, end-to-end simulation framework for IFE plant design built on JAX and the Tesseract composition framework. The target physics layer is a 1D Lagrangian radiation-hydrodynamics solver coupling compressible hydrodynamics, sector ray-traced laser deposition, flux-limited thermal conduction, and electron-ion equilibration, all differentiable with respect to laser pulse parameters and model coefficients. The plant layer is a differentiable re-implementation of the LLNL Generalized Economics Model (GEM), spanning laser driver, chamber engineering, thermal cycle, and cost of electricity.
The solvers are composed as Tesseract components, each exposing a vector-Jacobian product endpoint. The architecture supports decomposition at disciplinary boundaries so that each node can be independently developed, validated, and swapped. Gradients of plant-level COE propagate back through the full composed chain to individual laser pulse parameters. We demonstrate end-to-end sensitivities of COE with respect to laser and target parameters at ignition-relevant scales, and compare against gradients from GEM's built-in empirical gain scaling. Because the full chain is differentiable, the same machinery supports uncertainty propagation without sampling: a single Jacobian evaluation yields sensitivities of COE to internal model parameters such as the flux limiter coefficient, translating physics-model uncertainty directly into economic uncertainty.
The modular Tesseract interface makes fidelity swapping straightforward: replacing the empirical gain curve with the rad-hydro solver at the target physics boundary changes both the predicted optimal design point and the gradient landscape used to reach it. We compare optimization trajectories and gradient quality across fidelity levels, and discuss how fidelity mismatches propagate through the coupled system.
We present a fully differentiable, end-to-end simulation framework for IFE plant design built on JAX and the Tesseract composition framework. The target physics layer is a 1D Lagrangian radiation-hydrodynamics solver coupling compressible hydrodynamics, sector ray-traced laser deposition, flux-limited thermal conduction, and electron-ion equilibration, all differentiable with respect to laser pulse parameters and model coefficients. The plant layer is a differentiable re-implementation of the LLNL Generalized Economics Model (GEM), spanning laser driver, chamber engineering, thermal cycle, and cost of electricity.
The solvers are composed as Tesseract components, each exposing a vector-Jacobian product endpoint. The architecture supports decomposition at disciplinary boundaries so that each node can be independently developed, validated, and swapped. Gradients of plant-level COE propagate back through the full composed chain to individual laser pulse parameters. We demonstrate end-to-end sensitivities of COE with respect to laser and target parameters at ignition-relevant scales, and compare against gradients from GEM's built-in empirical gain scaling. Because the full chain is differentiable, the same machinery supports uncertainty propagation without sampling: a single Jacobian evaluation yields sensitivities of COE to internal model parameters such as the flux limiter coefficient, translating physics-model uncertainty directly into economic uncertainty.
The modular Tesseract interface makes fidelity swapping straightforward: replacing the empirical gain curve with the rad-hydro solver at the target physics boundary changes both the predicted optimal design point and the gradient landscape used to reach it. We compare optimization trajectories and gradient quality across fidelity levels, and discuss how fidelity mismatches propagate through the coupled system.
No video available