Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Performance of AI agents based on reasoning language models on ALD process optimization tasks

Angel Yanguas-Gil

TL;DR

This study tests whether reasoning large language models can autonomously optimize ALD processes without prior knowledge of growth-per-cycle dynamics. It employs a two-step agent that first reasons about the process and then outputs a structured JSON with dose times, interfacing with a simulated Langmuir-type ALD model that may include a non self-limited CVD component. Results show that reasoning models like o3 and GPT5 can converge to near-saturating dose times in most cases using about 10–15 experiments, but substantial run-to-run variability and occasional misclassification of non self-limited behavior are observed. The work demonstrates the potential and limits of model-based autonomous optimization in ALD, highlights the benefit of simple priors to improve data efficiency, and provides an open-source simulator to foster reproducibility and further improvement.

Abstract

In this work we explore the performance and behavior of reasoning large language models to autonomously optimize atomic layer deposition (ALD) processes. In the ALD process optimization task, an agent built on top of a reasoning LLM has to find optimal dose times for an ALD precursor and a coreactant without any prior knowledge on the process, including whether it is actually self-limited. The agent is meant to interact iteratively with an ALD reactor in a fully unsupervised way. We evaluate this agent using a simple model of an ALD tool that incorporates ALD processes with different self-limited surface reaction pathways as well as a non self-limited component. Our results show that agents based on reasoning models like OpenAI's o3 and GPT5 consistently succeeded at completing this optimization task. However, we observed significant run-to-run variability due to the non deterministic nature of the model's response. In order to understand the logic followed by the reasoning model, the agent uses a two step process in which the model first generates an open response detailing the reasoning process. This response is then transformed into a structured output. An analysis of these reasoning traces showed that the logic of the model was sound and that its reasoning was based on the notions of self-limited process and saturation expected in the case of ALD. However, the agent can sometimes be misled by its own prior choices when exploring the optimization space.

Performance of AI agents based on reasoning language models on ALD process optimization tasks

TL;DR

This study tests whether reasoning large language models can autonomously optimize ALD processes without prior knowledge of growth-per-cycle dynamics. It employs a two-step agent that first reasons about the process and then outputs a structured JSON with dose times, interfacing with a simulated Langmuir-type ALD model that may include a non self-limited CVD component. Results show that reasoning models like o3 and GPT5 can converge to near-saturating dose times in most cases using about 10–15 experiments, but substantial run-to-run variability and occasional misclassification of non self-limited behavior are observed. The work demonstrates the potential and limits of model-based autonomous optimization in ALD, highlights the benefit of simple priors to improve data efficiency, and provides an open-source simulator to foster reproducibility and further improvement.

Abstract

In this work we explore the performance and behavior of reasoning large language models to autonomously optimize atomic layer deposition (ALD) processes. In the ALD process optimization task, an agent built on top of a reasoning LLM has to find optimal dose times for an ALD precursor and a coreactant without any prior knowledge on the process, including whether it is actually self-limited. The agent is meant to interact iteratively with an ALD reactor in a fully unsupervised way. We evaluate this agent using a simple model of an ALD tool that incorporates ALD processes with different self-limited surface reaction pathways as well as a non self-limited component. Our results show that agents based on reasoning models like OpenAI's o3 and GPT5 consistently succeeded at completing this optimization task. However, we observed significant run-to-run variability due to the non deterministic nature of the model's response. In order to understand the logic followed by the reasoning model, the agent uses a two step process in which the model first generates an open response detailing the reasoning process. This response is then transformed into a structured output. An analysis of these reasoning traces showed that the logic of the model was sound and that its reasoning was based on the notions of self-limited process and saturation expected in the case of ALD. However, the agent can sometimes be misled by its own prior choices when exploring the optimization space.
Paper Structure (14 sections, 16 equations, 12 figures, 4 tables)

This paper contains 14 sections, 16 equations, 12 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Reasoning large language models
  4. ALD process optimization agent
  5. ALD process model
  6. Results
  7. Performance of base agent
  8. Performance of base agent without initial guess
  9. Performance of agents with memory
  10. Model search strategy and reasoning
  11. Discussion
  12. Conclusions
  13. Derivation of ALD model
  14. Prompts used for the reasoning model

Figures (12)

  • Figure 1: Scheme of our AI agent for ALD process optimization: the logic component generates queries and process the response of the AI component, which uses a reasoning model to determine the strategy for optimization and request additional experiments. Based on the AI component's response, the logic component sends new conditions to the simulated ALD reactor.
  • Figure 2: Sample of the prompt passed to the reasoning model as well as a sample of the structured JSON returned by the model during each iteration. The model response determines whether the agent concludes the optimization or continues requesting additional experiments
  • Figure 3: Precursor saturation curves of simulated ALD processes used to evaluate the process optimization by AI agents: A) Soft-saturating model, showing the impact of a second, lower reactivity reaction pathway ($k^a_1$ = 5 s$^{-1}$, $k^b_1$ = 1 s$^{-1}$, $k_2$ = 4 s$^{-1}$); B) Impact of CVD component for various values of the non self-limited CVD component ($k_1$ = 5 s$^{-1}$, $k_2$ = 4 s$^{-1}$). In both cases, the coreactant dose time was set to $t_2$ = 1 s.
  • Figure 4: Optimized precursor and coreactant dose times for AI agents based on o3 and GPT5 models for the fast/fast process with a GPC of 1Å/cycle. In all cases the agents converged to an optimal solution, albeit with run-to-run variability as evidenced by the scatter in the plot. The greyscale contour plot represents the growth per cycle as a function of the precursor and coreactant dose time
  • Figure 5: Relative error between the growth per cycle resulting from the optimization process by the AI agent and the saturated GPC of the ALD process : A) Agent built on top of o3 reasoning model; B) Agent built on top of GPT5 reasoning model
  • ...and 7 more figures