Co4ICF: Co-evolving Physics-Informed Surrogate and RL-based Pulse Optimizer for Inertial Confinement Fusion

2026-07-11 · Source: Artificial Intelligence · Field: Science & Research — Artificial Intelligence & Machine Learning, Engineering & Applied Sciences · Depth: Expert, quick

Summary

The Co4ICF framework addresses a critical failure mode in Inertial Confinement Fusion (ICF) optimization, where offline-trained surrogates produce unreliable predictions when optimizers explore out-of-distribution input regions. Co4ICF integrates a physics-informed surrogate with a PPO-based pulse optimizer in a co-evolving loop. The surrogate continuously fine-tunes itself using policy-induced trajectories, thereby correcting extrapolation errors as the optimizer shifts the input distribution. This evolving surrogate then serves as a rapid environment for the optimizer. In the 1D MULTI environment, Co4ICF achieved a 146.1% normalized yield compared to the current laser design baseline. Furthermore, without any 2D training, the optimized pulse demonstrated a 246.9% normalized yield when directly evaluated in 2D-MULTI, confirming its cross-fidelity performance. The authors also released a large-scale MULTI-IFE simulation dataset for future benchmarking.

Key takeaway

For AI Scientists and Research Scientists optimizing complex physical systems like Inertial Confinement Fusion, you should recognize that traditional offline surrogates are vulnerable to out-of-distribution inputs. Implementing a co-evolving framework, such as Co4ICF, can mitigate these extrapolation errors by continuously adapting the surrogate to the optimizer's exploration. This approach promises substantial performance gains, as demonstrated by the 246.9% normalized yield in 2D-MULTI, and offers a robust strategy for high-fidelity simulation and optimization.

Key insights

Co4ICF co-evolves a physics-informed surrogate and an RL optimizer to mitigate out-of-distribution errors, significantly boosting Inertial Confinement Fusion yield.

Principles

• Offline surrogates fail on OOD inputs.
• Iterative fine-tuning corrects extrapolation errors.
• Co-evolution enhances optimizer robustness.

Method

Co4ICF couples a physics-informed surrogate with a PPO optimizer. The surrogate iteratively fine-tunes on policy-induced trajectories, correcting extrapolation errors, while the optimizer queries this evolving surrogate as a fast environment.

In practice

• Achieved 146.1% normalized yield in 1D MULTI.
• Attained 246.9% yield in 2D-MULTI without 2D training.
• Utilize the released MULTI-IFE dataset.

Topics

• Inertial Confinement Fusion
• Reinforcement Learning
• Physics-Informed Surrogates
• Out-of-Distribution Detection
• PPO Algorithm
• Co-evolutionary AI

Best for: AI Scientist, Research Scientist

Related on AIssential

• Offline Reinforcement Learning for Plasma Control in Nuclear Fusion: Codebase and Benchmark — RL4F provides a standardized offline RL benchmark for multi-actuator, long-horizon nuclear fusion plasma control.
• Sci-CoE: Co-evolving Scientific Reasoning LLMs via Geometric Consensus with Sparse Supervision — Sci-CoE improves LLM scientific reasoning by co-evolving solver and verifier through sparse-to-unsupervised learning.
• Non-intrusive Learning of Physics-Informed Spatio-temporal Surrogate for Accelerating Design — PISTM combines physics constraints and Koopman autoencoders for generalizable spatio-temporal surrogate modeling.
• Co-Learning Port-Hamiltonian Systems and Optimal Energy-Shaping Control — Co-learning pH system models and optimal EB-PBC policies via alternating optimization ensures stability and robustness in control.
• Towards Data-Efficient Cross-Device Generalization of Grad-Shafranov Equilibria via Transfer Learning Neural Operator — Neural operator pretraining enables data-efficient, real-time, cross-device generalization of Grad-Shafranov equilibria for magnetic confinement fusion.
• Learning Design Skills as Memory Policies for Agentic Photonic Inverse Design — SkillPCF uses memory-policy learning to accumulate design knowledge for efficient photonic inverse design.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.