Human-in-the-Loop Meta Bayesian Optimization for Fusion Energy and Scientific Applications
Researchers combine expert knowledge with meta-learning to slash experimental costs in fusion research.
A new framework called Human-in-the-Loop Meta Bayesian Optimization (HL-MBO) integrates expert knowledge with few-shot, uncertainty-aware machine learning to accelerate discovery in data-scarce, high-stakes scientific domains like Inertial Confinement Fusion. It recommends candidate experiments, provides interpretable explanations, and outperforms current Bayesian optimization methods on ICF energy yield optimization as well as benchmarks in molecular optimization and critical temperature maximization for superconducting materials. The paper was accepted at IJCAI 2026.
- Outperforms standard Bayesian optimization by ~40% on ICF yield optimization tasks
- Integrates meta-learned surrogate model with expert-informed acquisition function for experiment recommendations
- Provides interpretable explanations to build human trust and support informed decision-making
Why It Matters
HL-MBO could slash experimental costs and accelerate breakthroughs in fusion energy, materials science, and other resource-intensive fields.