Role of new surrogate models, acquisition functions, and LLMs in self-driving labs

Determine the role and effectiveness of new surrogate models, acquisition functions, and large language models in overcoming cold-start limitations and enabling few-shot or zero-shot Bayesian optimization workflows in self-driving laboratory systems.

Background

Self-driving labs typically rely on Bayesian optimization to maximize data efficiency, yet cold-start scenarios and rigid workflows limit early-stage performance and sustainability. Leveraging prior knowledge, flexible function-network workflows, and improved representations for categorical design spaces are proposed to accelerate convergence.

The paper highlights uncertainty about how novel surrogate architectures, acquisition rules, and LLM-based tools will contribute to resolving these challenges, motivating targeted research and benchmarking in realistic experimental campaigns.

References

Several limitations define the scope of this study. AssayBench-Loop is a retrospective benchmark focused on genome-wide screens with sufficient hit signal, and the policies evaluated here remain to be validated prospectively in a live experimental loop.

— Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens  (2609.11877 - Edwards et al., 10 Sep 2026) in Section Discussion, paragraph beginning “Several limitations define the scope of this study.”

The role of new types of surrogate models and acquisition functions, as well as LLMs in achieving this is yet to be determined.

— Perspective: Towards sustainable exploration of chemical spaces with machine learning  (2604.00069 - Sandonas et al., 31 Mar 2026) in Subsubsection 'Toward autonomous discovery in self-driving labs', Open challenges

While this direction may improve sample efficiency, important open questions remain regarding safety and reliability, and these methods still require thorough validation in real-world hardware settings.

— A Decade of Bayesian Optimization for Controller Tuning and Robot Learning: Tutorial, Review, and Future Prospects  (2609.09403 - Stenger et al., 8 Sep 2026) in Section 2, subsection “Further Variants” (Sec. 2.10)

Whether these strategies transfer to a live, multi-objective experimental loop in acidic OER, where the goal is to discover novel catalysts across a coupled activity–stability space, remains an open question.

— AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution  (2609.30133 - Jenewein et al., 24 Sep 2026) in Section 2, Introduction

The exact method is open. Three paths are plausible:

— Decision-Focused Active Learning for Scale-Aware Critical-Materials Recovery  (2609.09413 - Srinivas et al., 8 Sep 2026) in Section 3.4, “Candidate method family”; Section 5, “Exploratory synthetic design probe”; Section 6, “Proposed prospective campaign”

The approach also extends to batch constrained MOBO, and contextual policies remain open, particularly under the low-data conditions typical of autonomous materials design campaigns.

— Portfolio-Based Constrained Multi-Objective Bayesian Optimization for Materials Design  (2609.19550 - Sinha et al., 17 Sep 2026) in Section 4, Discussion