Human-in-the-Loop Meta Bayesian Optimization for Fusion Energy and Scientific Applications

Aarne Lees; Ejaz Rahman; Ricardo Luna Gutierrez; Riccardo Betti; Sahand Ghorbanpour; Soumyendu Sarkar; Varchas Gopalaswamy; Vineet Gundecha

arxiv: 2605.00068 · v1 · submitted 2026-04-30 · 💻 cs.LG · cs.AI· physics.plasm-ph

Human-in-the-Loop Meta Bayesian Optimization for Fusion Energy and Scientific Applications

Ricardo Luna Gutierrez , Sahand Ghorbanpour , Ejaz Rahman , Varchas Gopalaswamy , Riccardo Betti , Vineet Gundecha , Aarne Lees , Soumyendu Sarkar This is my paper

Pith reviewed 2026-05-09 20:42 UTC · model grok-4.3

classification 💻 cs.LG cs.AIphysics.plasm-ph

keywords human-in-the-loopmeta Bayesian optimizationinertial confinement fusionscientific discoveryfew-shot optimizationinterpretable machine learningenergy yield optimization

0 comments

The pith

Human-in-the-loop meta Bayesian optimization outperforms standard methods for inertial confinement fusion energy yield and related scientific tasks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents HL-MBO as a way to fold expert human judgment into a meta-learned Bayesian optimization loop so that scarce, expensive experiments can be chosen more effectively. It builds a surrogate model that learns across tasks and then uses an expert-informed acquisition function to pick the next candidate while also explaining why that candidate was chosen. The authors demonstrate the approach on the problem of maximizing energy yield from inertial confinement fusion shots, where each trial is rare and costly. They further test it on molecular property optimization and on finding materials with higher superconducting critical temperatures. A reader would care if the method truly reduces the number of physical trials needed to reach better outcomes in these domains.

Core claim

HL-MBO integrates expert knowledge with few-shot, uncertainty-aware machine learning through a meta-learned surrogate model and an expert-informed acquisition function that recommends experiments and supplies interpretable explanations; this combination is shown to outperform existing Bayesian optimization baselines on inertial confinement fusion energy yield optimization as well as on benchmarks for molecular optimization and critical temperature maximization in superconducting materials.

What carries the argument

The meta-learned surrogate model paired with an expert-informed acquisition function that recommends candidate experiments and supplies interpretable explanations of its suggestions.

If this is right

Fewer physical experiments are required to reach target energy yields in inertial confinement fusion.
The same framework improves performance on molecular property optimization and superconductor critical-temperature search.
Interpretable explanations allow experts to trust and iteratively refine the suggestions in high-stakes settings.
The approach is designed for domains where data are scarce and each trial is expensive.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Success on the three reported tasks suggests the meta-learning component helps the surrogate transfer useful structure across different but related optimization problems.
The human-in-the-loop design could be extended to other experimental sciences where domain experts hold knowledge that is difficult to encode in advance.
If the explanations prove reliable, they might allow experts to correct the model mid-loop and further reduce wasted trials.

Load-bearing premise

Expert knowledge can be folded into the acquisition function without introducing systematic bias or degrading the uncertainty estimates that Bayesian optimization relies on.

What would settle it

A side-by-side run of HL-MBO versus standard Bayesian optimization on the same sequence of real inertial confinement fusion shots, with the result measured by whether HL-MBO reaches higher energy yields after the same number of trials.

Figures

Figures reproduced from arXiv: 2605.00068 by Aarne Lees, Ejaz Rahman, Ricardo Luna Gutierrez, Riccardo Betti, Sahand Ghorbanpour, Soumyendu Sarkar, Varchas Gopalaswamy, Vineet Gundecha.

**Figure 1.** Figure 1: Refinements proposed in HL-MBO over standard Bayesian Optimization. HL-MBO trains a meta-surrogate using related source tasks and integrates a preference model to incorporate expert knowledge. During online black-box optimization the meta-surrogate and preference model are used in conjunction with an AF to optimize the target function (scientific experiment), providing explanations to support decision-maki… view at source ↗

**Figure 2.** Figure 2: (a) Visual representation of the expected experimental output (TNPs’ Mean) and the associated uncertainty of the prediction for both candidate points proposed. (b) Feature attribution using SHAP and LIME on the ICF optimization task over 5 features, for the two candidate points proposed by HL-MBO. Each bar represents contribution with respect to the BO metrics (AF, Expected Output, Uncertainty Estimation).… view at source ↗

**Figure 4.** Figure 4: Results of the comparison between different hypothesis. Expert and Adversarial represent the best-case and worst-case scenarios, respectively, for sampling in the preference learning initialization. Random, represents uniform sampling. value xp that yields the highest output yp from the candidate pairs {x1, x2} affects HL-MBO’s performance. We evaluate the performance of HL-MBO on different experts’ accu… view at source ↗

**Figure 5.** Figure 5: Results showing the accuracy of experts in selecting the better point from each presented pair during preference model construction. 50% accuracy indicates random choice. 5 Related Work Human-aided and Explainable BO. While no prior work has integrated few-shot meta-learning into human-aided explainable BO as proposed in this paper, recent efforts have advanced both explainable and human-in-the-loop BO. … view at source ↗

**Figure 1.** Figure 1: Illustration of an surrogate predictions, uncertainty and AFs value. Such visualizations provide ICF practitioners with insights to select the candidate most likely to succeed. The numbering represent the order in which samples were taken. B.1 HL-MBO’s Surrogate Predictions To understand the superior performance achieved by our method, we examine the predictions made by its surrogate model and assess how w… view at source ↗

**Figure 3.** Figure 3: MBO’s meta-learned surrogate prediction against NAP’s surrogate prediction. MBO’s predictions are more accurate, smoother and interpretable. C No-Harm Guarantee A critical design requirement for human-in-the-loop optimization is robustness to imperfect expert knowledge. As shown in CoExBO Adachi et al. [2024], we the decay parameter γ provides a no-harm guarantee: even with adversarial expert preference… view at source ↗

**Figure 2.** Figure 2: HL-MBO’s meta-learned surrogate predictions: (top) without only one context point; (bottom) after three context points; (right) the optimization target function. We can observe that our approach achieves quick adaptation with high sample efficiency, closely approximating the true function in just three samples. This property is ideal for the limited ICF experiments possible on a shot day. B.2 HL-MBO vs NAP… view at source ↗

read the original abstract

Inertial Confinement Fusion (ICF) holds transformative promise for sustainable, near-limitless clean energy, yet remains constrained by prohibitively high costs and limited experimental opportunities. This paper presents Human-in-the-Loop Meta Bayesian Optimization (HL-MBO), a framework that integrates expert knowledge with few-shot, uncertainty-aware machine learning to accelerate discovery in data-scarce, high-stakes scientific domains. HL-MBO introduces a meta-learned surrogate model with an expert-informed acquisition function to recommend candidate experiments. To foster trust and enable informed decisions, HL-MBO also provides interpretable explanations of its suggestions. We show HL-MBO outperforms current BO methods on ICF energy yield optimization, as well as benchmarks in molecular optimization and critical temperature maximization for superconducting materials.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

HL-MBO folds expert input into meta-BO for ICF optimization but tests the loop with simulated rather than live experts, so the reported gains stay preliminary.

read the letter

HL-MBO combines a meta-learned surrogate with an expert-informed acquisition function and built-in explanations to suggest experiments in data-scarce settings like inertial confinement fusion. It reports better energy-yield results than standard Bayesian optimization on ICF tasks, plus gains on molecular optimization and superconductor critical-temperature benchmarks. The concrete architecture for this domain, including the interpretability layer, is the clearest addition over prior meta-BO work. The framing is practical: each ICF shot is expensive, so any method that safely uses limited expert knowledge to cut the number of trials is worth examining. The explanations could help experimental teams decide whether to trust a recommendation. The soft spots sit in the human-in-the-loop validation. The experiments appear to rely on retrospective or simulated expert labels instead of live sessions, which leaves the central claim under-supported. Without live input it is hard to check whether the expert term improves recommendations or simply injects bias and harms uncertainty calibration in the few-shot regime. Meta-learning transfer from the other two domains to ICF also lacks detailed ablation or regularization checks. The abstract-level performance claims would need stronger baselines, statistical tests, and component breakdowns before the gains can be attributed confidently to the human component. This paper is aimed at applied researchers who build optimization tools for high-cost physics or materials experiments. A reader already working on Bayesian methods for scientific discovery would pick up the specific acquisition and explanation design. I would send it for peer review. The target application is important and the integration idea is concrete enough that referees can usefully press for tighter experiments and clearer evidence on the live-expert case.

Referee Report

2 major / 0 minor

Summary. The manuscript introduces Human-in-the-Loop Meta Bayesian Optimization (HL-MBO), a framework combining a meta-learned surrogate model with an expert-informed acquisition function to recommend experiments in data-scarce, high-stakes domains. It targets inertial confinement fusion (ICF) energy yield optimization, with additional benchmarks on molecular optimization and critical temperature maximization for superconducting materials. The approach includes interpretable explanations for suggestions and claims outperformance over standard Bayesian optimization methods.

Significance. If the empirical claims are substantiated, the work could meaningfully advance Bayesian optimization for scientific applications with costly experiments and limited data, such as fusion energy research, by integrating human expertise in a few-shot, uncertainty-aware manner. The emphasis on interpretability is a constructive feature for trust in high-stakes settings.

major comments (2)

[Abstract] Abstract: The headline claim that HL-MBO outperforms current BO methods on ICF energy yield optimization (as well as the two other benchmarks) is presented without any experimental details, baselines, number of trials, statistical tests, or ablation results, rendering the central empirical contribution unevaluable from the text.
[Method (expert-informed acquisition function)] Method section on expert-informed acquisition function: No analysis or experiments quantify potential bias, inconsistency, or miscalibration introduced by the human expert input in the few-shot ICF regime; this is load-bearing because the outperformance claim rests on the acquisition function reliably improving recommendations while preserving uncertainty calibration.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for their constructive comments, which have helped us improve the clarity and robustness of our work. Below we respond to each major comment in turn.

read point-by-point responses

Referee: [Abstract] Abstract: The headline claim that HL-MBO outperforms current BO methods on ICF energy yield optimization (as well as the two other benchmarks) is presented without any experimental details, baselines, number of trials, statistical tests, or ablation results, rendering the central empirical contribution unevaluable from the text.

Authors: We acknowledge that the abstract is necessarily concise and therefore omits the full experimental details. The manuscript body provides these details, including the experimental protocols, baselines, trial counts, statistical analyses, and ablation studies. To directly address the concern, we have revised the abstract to briefly reference the primary baselines and the scale of the evaluations. revision: yes
Referee: [Method (expert-informed acquisition function)] Method section on expert-informed acquisition function: No analysis or experiments quantify potential bias, inconsistency, or miscalibration introduced by the human expert input in the few-shot ICF regime; this is load-bearing because the outperformance claim rests on the acquisition function reliably improving recommendations while preserving uncertainty calibration.

Authors: We agree that quantifying the impact of human expert input is important for substantiating the claims. The original manuscript describes the expert-informed acquisition function and includes qualitative discussion of its integration but does not contain quantitative experiments on bias, inconsistency, or calibration effects. In the revised manuscript we have added a sensitivity analysis that perturbs expert inputs to simulate bias and inconsistency in the few-shot regime and reports the resulting effects on recommendation quality and uncertainty calibration. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical outperformance claims rest on benchmarks, not self-referential fits or derivations

full rationale

The paper introduces HL-MBO as a meta-learned surrogate plus expert-informed acquisition function and reports outperformance on ICF yield optimization plus molecular and superconductivity benchmarks. No equations, derivations, or first-principles results are presented in the provided text. Performance claims are framed as empirical comparisons against existing BO methods rather than any quantity that reduces by construction to fitted parameters or self-citations. The central premise (human-in-the-loop integration in few-shot regimes) is evaluated via experiments whose validity is independent of the method's own outputs. This is the common case of a self-contained empirical contribution with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the framework implicitly assumes expert knowledge can be encoded without loss of calibration.

axioms (1)

domain assumption Expert knowledge can be reliably translated into an acquisition function that improves few-shot optimization without introducing unquantified bias.
Core premise of the human-in-the-loop component stated in the abstract.

pith-pipeline@v0.9.0 · 5457 in / 996 out tokens · 30598 ms · 2026-05-09T20:42:26.860963+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

[1]

Looping in the hu- man: Collaborative and explainable bayesian optimization

Masaki Adachi, Brady Planden, et al. Looping in the hu- man: Collaborative and explainable bayesian optimization. In International Conference on Artificial Intelligence and Statistics, pages 505–513. PMLR, 2024. J. Delettrez, R. Epstein, M. C. Richardson, P. A. Jaanimagi, and B. L. Henke. Effect of laser illumination nonuni - formity on the analysis of ti...

work page 2024
[2]

Meta-learned bayesian optimization for energy yield in inertial confinement fusion, 2024

Betti, Avisek Naug, Paolo Faraboschi, and Soumyendu Sarkar. Meta-learned bayesian optimization for energy yield in inertial confinement fusion, 2024. Kexin Huang, Tianfan Fu, Wenhao Gao, Yue Zhao, Yusuf

work page 2024
[3]

Therapeutics data com - mons: Machine learning datasets and tasks for drug discov- ery and development

Roohani, Jure Leskovec, Connor W Coley, Cao Xiao, Ji - meng Sun, and Marinka Zitnik. Therapeutics data com - mons: Machine learning datasets and tasks for drug discov- ery and development. Proceedings of Neural Information Processing Systems, NeurIPS Datasets and Benchmarks , 2021. Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimizatio...

work page 2021
[4]

Tung Nguyen and Aditya Grover

Associates, Inc., 2017. Tung Nguyen and Aditya Grover. Transformer neural pro - cesses: Uncertainty-aware meta learning via sequence modeling. In Proceedings of the 39th International Confer- ence on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16569–16594. PMLR, 17–23 Jul 2022. Marco Tulio Ribeiro, Sameer Singh, and Car...

work page arXiv 2017

[1] [1]

Looping in the hu- man: Collaborative and explainable bayesian optimization

Masaki Adachi, Brady Planden, et al. Looping in the hu- man: Collaborative and explainable bayesian optimization. In International Conference on Artificial Intelligence and Statistics, pages 505–513. PMLR, 2024. J. Delettrez, R. Epstein, M. C. Richardson, P. A. Jaanimagi, and B. L. Henke. Effect of laser illumination nonuni - formity on the analysis of ti...

work page 2024

[2] [2]

Meta-learned bayesian optimization for energy yield in inertial confinement fusion, 2024

Betti, Avisek Naug, Paolo Faraboschi, and Soumyendu Sarkar. Meta-learned bayesian optimization for energy yield in inertial confinement fusion, 2024. Kexin Huang, Tianfan Fu, Wenhao Gao, Yue Zhao, Yusuf

work page 2024

[3] [3]

Therapeutics data com - mons: Machine learning datasets and tasks for drug discov- ery and development

Roohani, Jure Leskovec, Connor W Coley, Cao Xiao, Ji - meng Sun, and Marinka Zitnik. Therapeutics data com - mons: Machine learning datasets and tasks for drug discov- ery and development. Proceedings of Neural Information Processing Systems, NeurIPS Datasets and Benchmarks , 2021. Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimizatio...

work page 2021

[4] [4]

Tung Nguyen and Aditya Grover

Associates, Inc., 2017. Tung Nguyen and Aditya Grover. Transformer neural pro - cesses: Uncertainty-aware meta learning via sequence modeling. In Proceedings of the 39th International Confer- ence on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16569–16594. PMLR, 17–23 Jul 2022. Marco Tulio Ribeiro, Sameer Singh, and Car...

work page arXiv 2017