Human-AI collaborative autonomous synthesis with pulsed laser deposition for remote epitaxy
📝 Abstract
Autonomous laboratories typically rely on data-driven decision-making, occasionally with human-in-the-loop oversight to inject domain expertise. Fully leveraging AI agents, however, requires tightly coupled, collaborative workflows spanning hypothesis generation, experimental planning, execution, and interpretation. To address this, we develop and deploy a human-AI collaborative (HAIC) workflow that integrates large language models for hypothesis generation and analysis, with collaborative policy updates driving autonomous pulsed laser deposition (PLD) experiments for remote epitaxy of BaTiO $_3 $/graphene. HAIC accelerated the hypothesis formation and experimental design and efficiently mapped the growth space to graphene-damage. In situ Raman spectroscopy reveals that chemistry drives degradation while the highest energy plume components seed defects, identifying a low-O $_2$ pressure low-temperature synthesis window that preserves graphene but is incompatible with optimal BaTiO $_3$ growth. Thus, we show a two-step Ar/O $_2$ deposition is required to exfoliate ferroelectric BaTiO $_3$ while maintaining a monolayer graphene interlayer. HAIC stages human insight with AI reasoning between autonomous batches to drive rapid scientific progress, providing an evolution to many existing human-in-the-loop autonomous workflows.
💡 Analysis
Autonomous laboratories typically rely on data-driven decision-making, occasionally with human-in-the-loop oversight to inject domain expertise. Fully leveraging AI agents, however, requires tightly coupled, collaborative workflows spanning hypothesis generation, experimental planning, execution, and interpretation. To address this, we develop and deploy a human-AI collaborative (HAIC) workflow that integrates large language models for hypothesis generation and analysis, with collaborative policy updates driving autonomous pulsed laser deposition (PLD) experiments for remote epitaxy of BaTiO $_3 $/graphene. HAIC accelerated the hypothesis formation and experimental design and efficiently mapped the growth space to graphene-damage. In situ Raman spectroscopy reveals that chemistry drives degradation while the highest energy plume components seed defects, identifying a low-O $_2$ pressure low-temperature synthesis window that preserves graphene but is incompatible with optimal BaTiO $_3$ growth. Thus, we show a two-step Ar/O $_2$ deposition is required to exfoliate ferroelectric BaTiO $_3$ while maintaining a monolayer graphene interlayer. HAIC stages human insight with AI reasoning between autonomous batches to drive rapid scientific progress, providing an evolution to many existing human-in-the-loop autonomous workflows.
📄 Content
Autonomous and robotic laboratories are transforming materials synthesis by combining highthroughput experimentation and computation with machine learning to optimize processes 1,2 . The prevailing paradigm imbues automated synthesis and characterization laboratories with decisionmaking driven by artificial intelligence (AI) to realize desired molecular and material properties while minimizing costly human intervention 3 . However, human-in-the-loop (HITL) workflows have been shown to increase efficiency in industry 4 and in materials characterization, such as Xray phase mapping 5 and atomic force microscopy (AFM) 6 , by bounding or initializing optimization procedures with human intuition within an automated experimental loop 7 . This improvement can be rationalized as arising from the incorporation of expert prior knowledge rather than the typically uninformative priors assumed by standard optimization algorithms. However, existing HITL workflows still suffer from significant drawbacks, with most being limited to an interventionist approach in which humans either seed initial experiments, modify parameter spaces, or inject knowledge at predefined points, limiting their utility. To leverage the rapidly developing capabilities of modern AI agents, more tightly coupled, collaborative workflows are required across all stages of the experimental process, from hypothesis generation and experimental planning to results interpretation that go beyond existing HITL workflows.
Here, we evolve HITL into human-AI collaborative (HAIC) workflows that couple human expertise, large language models (LLMs), and autonomous systems through mixed-initiative, between-batch loops. This approach is amplified by using LLMs alongside autonomous systems as “co-scientists” 8,9 to help generate hypotheses, plan experiments 10 , and analyze data, especially when retrieval-augmented generation 11 (RAG) grounds the model in relevant scientific corpora 12 .
This flexibility is particularly suited to thin-film synthesis, where large parameter spaces with sparse prior data make defining robust success metrics challenging, especially when exploring new applications or materials systems.
Notably, thin-film synthesis poses a different set of challenges than materials discovery and requires a tightly coupled collaborative workflow. Success metrics may require complex measurements with bespoke analysis, defined iteratively rather than a priori, and are subject to significant data complexity (noise, unknowns, intangibles). Further, general mechanistic and experimental databases for thin-film synthesis are lacking, and literature-based recipes are often too unreliable to predict performance metrics 13,14 , making the incorporation of prior knowledge difficult. As data accumulates, the workflow must evolve to accommodate process improvement 15 in all but the simplest scenarios to be effective. Addressing these challenges is essential to broaden the success and applicability of autonomous platforms, especially for thin-film synthesis.
Our work applies HAIC to autonomous synthesis, addressing key challenges in the emerging field of remote epitaxy (RE) of complex oxides by pulsed laser deposition (PLD). In RE, a twodimensional interlayer, such as graphene, is placed on a single-crystal substrate, enabling the epitaxial growth of single-crystalline films that can be exfoliated as thin membranes and integrated on arbitrary substrates 16 . However, PLD of complex oxides such as BaTiO3 (BTO) requires growth conditions that destroy monolayer graphene interlayers, so successful PLD RE has relied on bilayer graphene to retain film transferability 17,18 . This poses a serious challenge for RE in PLD (and molecular beam epitaxy) because the transmitted electrostatic potential of the substrate through graphene is small (~10-20 meV) and short-ranged (~2 Å) 19 . Consequently, >1 graphene layers weaken or eliminate remote film alignment, yielding poorly oriented or polycrystalline films 20 , depending on the substrate’s ionicity/polarity 21,22 . Comprehensive synthesis studies are essential to understand the limitations of oxide RE with PLD and aid the development of alternative strategies to enable remote epitaxy for arbitrary materials.
We deploy a HAIC strategy by combining state-of-the-art LLMs with an autonomous PLD system equipped with in situ diagnostics to understand how complex-oxide growth conditions drive graphene damage during RE, using BTO as a test case. LLM-assisted hypothesis generation, experimental design, and iterative process refinements enabled the autonomous campaign to efficiently map the parameter space and identify the BTO growth regime that minimizes graphene damage. Targeted in situ Raman experiments further elucidate the interplay of chemical interactions and ballistic damage in graphene defect generation during BTO deposition, providing a mechanistic basis for O2-mediated RE with PLD. We also show that a two-step growth sequence m
This content is AI-processed based on ArXiv data.