Hypothesis-Driven Autonomous Materials Synthesis with Multimodal LLM Agents

This review is based on the arXiv preprint, version 1 of the paper dated 16 September 2026; its independent peer-review status is therefore not confirmed in the source text. The paper presents SynAgent, which connects multimodal large-language-model agents to an automated laboratory. The agent’s goal is not only to find a desirable sample, but also to construct an explicit, traceable, and revisable understanding of the relationship between synthesis conditions and outcomes from each experiment. The system was tested on (001)-oriented LiCoO₂ thin films.

SynAgent follows four steps: proposing conditions from its current understanding, autonomously conducting synthesis and measurement, analyzing the data, and rewriting the understanding after observing the outcome. The agent’s understanding is a natural-language document that records process claims. In each iteration, it either proposes an experiment to verify a claim or deliberately tests conditions where the hypothesis is expected to fail. Before execution, the hypothesis, proposed temperature, and numerical prediction are recorded; afterward, the result is assessed against the predicted success or failure category.

Another system feature is adaptive generation of analysis “skills.” The skill library was initially empty, and the main agent delegated analysis-code creation to a subagent when needed. In the first experiment, it created a skill to locate the 003 and 006 peaks in an X-ray diffraction pattern and calculate I(003)/I(006), the optimization target and crystallinity measure. After observing bright particles in a scanning electron microscopy image, the agent created a skill to quantify particles and background texture.

In the 18-experiment campaign, substrate temperature was the only controllable variable, selected from 200 to 750 °C in 10 °C increments; RF power, deposition time, and gas pressures remained fixed. The agent used GPT-5.5 and alternated verify and falsify modes, starting with verification. Of the 18 predictions, 15 were supported and 3 were refuted. Films at 250–630 °C had ratios below 0.1; ordered growth appeared at 650 °C, and the largest ratio, 61.6, was obtained at 670 °C.

Falsify proposals played a central role in changing the agent’s understanding. In the first iteration, a predicted success at 600 °C was refuted by a measured ratio of 0.001. The agent then predicted failure at 700 °C, but measured 35.6, overturning its initial view of uniform high-temperature degradation. In iteration 10, it predicted 0.2 at 650 °C, yet observed 54.9. The final map included no ordering below 630 °C, a 630–650 °C transition, a 650–690 °C ordered window, and an ordered but suboptimal 700–750 °C region.

The limitations are important: the demonstration covers only one material and one controllable variable, and each temperature was tested once; thus, small differences within 650–690 °C may reflect run-to-run variation. No controlled comparison with Bayesian optimization under matched experimental budgets was performed. Skills were validated only through system self-tests, and judgments that predictions were supported or refuted relied on the model’s reasoning. SEM alone also cannot establish particle composition or phase identity; the paper proposes further characterization and a 640 °C experiment.

Original arXiv paper