LLM Agents for Knowledge Discovery in Atomic Layer Processing

TL;DR

This work investigates using large language model (LLM) agents to enable knowledge discovery in materials science by interrogating a black-box system and extracting generalizable rules rather than optimizing predefined objectives. It employs two demonstrations—the Alien Market parlor game and a detailed Atomic Layer Processing (ALP) reactor model with limited probes—to assess the capabilities and limits of trial-and-error knowledge discovery. Results show that LLM agents can generate hypotheses and summarize system behavior, but progress is strongly path-dependent and contingent on sustained exploration and prompt design. The findings suggest AI-assisted discovery in data-poor regimes can support database construction and guide human-in-the-loop or multi-agent approaches for ALD/ALE chemistries in materials research.

Abstract

Large Language Models (LLMs) have garnered significant attention for several years now. Recently, their use as independently reasoning agents has been proposed. In this work, we test the potential of such agents for knowledge discovery in materials science. We repurpose LangGraph's tool functionality to supply agents with a black box function to interrogate. In contrast to process optimization or performing specific, user-defined tasks, knowledge discovery consists of freely exploring the system, posing and verifying statements about the behavior of this black box, with the sole objective of generating and verifying generalizable statements. We provide proof of concept for this approach through a children's parlor game, demonstrating the role of trial-and-error and persistence in knowledge discovery, and the strong path-dependence of results. We then apply the same strategy to show that LLM agents can explore, discover, and exploit diverse chemical interactions in an advanced Atomic Layer Processing reactor simulation using intentionally limited probe capabilities without explicit instructions.

Figures (5)

• Figure 1: The goal of probing an experimental system is to discover the rules mapping the input space to the output space. Asking specific questions limits discovery to very specific rules. In materials science, one interpretation of this figure could be to see input space as precursor chemicals, the rules as all possible synthesis pathways, and the output space as desired behavior of the synthesized materials (as this behavior may be synthesis-dependent).
• Figure 2: a) graph for the first example (Alien market) b) graph for the second example (ALP reactor) LLM-based components are shown in blue, while deterministic components are grey. c) schematic reactor layout.
• Figure 3: Initially, most models except for gpt-5 perform poorly at discovering the rules of the Alien market (orange squares). The performance markedly improves upon requesting the agent to perform a defined number of experiments (dark dots).
• Figure 4: Full concentration profiles (a, b, c), surface coverage profiles (e, f), mass deposition profiles (g) generated during an experiment pulsing B and C, generating gaseous reaction product E and solid S. The dotted line in g) shows the position of the QCM sensor. Sensor data available to the agent is displayed in d) and h).
• Figure 5: Illustration of the experimental path taken by the agent in Run 2 (D accessible) iteration 3. The agent discovers growth, etching and passivation.