📄 chemistry

A modular active learning framework for autonomous atomic layer deposition

This paper presents a closed-loop autonomous platform integrating an automated PE-ALD tool, in situ characterization, and a Gaussian process-based active learning framework that efficiently maps high-dimensional synthesis-property relationships for TiO2 deposition, significantly outperforming random sampling in predictive accuracy and enabling targeted process optimization with far fewer experiments.

Original authors: Maher Alghalayini, Tim Kodalle, Edward S. Barnard

Published 2026-09-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Maher Alghalayini, Tim Kodalle, Edward S. Barnard

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Making the world's smallest electronic components requires building layers of material so thin they are measured in atoms. One of the most precise ways to do this is a technique called atomic layer deposition, where chemicals are introduced into a vacuum chamber one by one, reacting with a surface to create a film that is perfectly uniform. This process is vital for making everything from faster computer chips to more efficient batteries. However, finding the right settings to make these films is incredibly difficult. The process depends on dozens of variables, such as gas pressure, temperature, and the timing of chemical pulses. These variables interact in complex, unpredictable ways, meaning that changing one setting can unexpectedly ruin the result. Traditionally, scientists find the right combination by changing one variable at a time, a slow and laborious method that often misses the best settings because it cannot see how the variables work together.

To solve this, researchers at Lawrence Berkeley National Laboratory built a system that can think for itself. They created a fully automated machine, which they named ALDBot, capable of running experiments without human help. This machine is connected to a smart computer program that uses a strategy called active learning. Instead of guessing or testing settings randomly, the program builds a mental map of how the process works. After each experiment, it analyzes the results and decides exactly which setting to try next to learn the most new information. The goal is not just to find a single "best" setting, but to understand the entire landscape of possibilities, identifying the safe zones where the film grows perfectly and the dangerous zones where it fails.

The researchers tested this system by trying to deposit a thin film of titanium dioxide, a common material used in electronics and coatings. They started with a smaller, four-dimensional version of the problem, meaning they varied only four settings at a time while keeping the others fixed. They compared their smart system against a standard approach where settings are chosen completely at random. The results were clear: the active learning system learned much faster. It reached a high level of accuracy in predicting the film's growth rate using far fewer experiments than the random method. While the random approach kept stumbling over the same uninformative settings, the smart system systematically targeted the gaps in its knowledge, quickly mapping out the "window" where the process works perfectly.

Encouraged by this success, the team expanded the challenge to the full complexity of the machine, which involves twelve different variables. This is a massive increase in difficulty; mathematically, exploring all combinations of twelve variables would require trillions of experiments, a number impossible to achieve in a human lifetime. Even with a budget of only two hundred experiments, the system managed to build a useful model of the entire space. While it could not map every single detail with perfect precision, the model was accurate enough to make reliable predictions. More importantly, the system could explain why it made those predictions. By analyzing the data, it ranked the twelve variables by importance, telling the researchers which settings mattered most and which had almost no effect. This insight allowed them to simplify the problem, focusing only on the critical knobs that truly control the outcome.

To prove the system was working correctly, the researchers used the model to select a specific set of settings that lay right on the edge of the safe operating zone. They ran the machine with these settings and produced a film of titanium dioxide. Chemical analysis confirmed that the film was pure and exactly what the model predicted. This demonstrated that the system could not only navigate a complex, high-dimensional space but also guide a human to a successful result with high confidence, even when the data was sparse. The study shows that by combining a robotic lab with a learning algorithm, scientists can explore difficult chemical processes much faster and more efficiently than ever before. This approach does not just find one good answer; it builds a deep understanding of the process, offering a powerful new way to design materials for the future.

Problem Statement
Optimizing gas-phase thin-film deposition processes, such as Atomic Layer Deposition (ALD), remains a significant experimental challenge due to the high dimensionality of the synthesis parameter space and the complex, nonlinear interactions among process parameters. Conventional optimization relies on iterative, experience-driven "one-parameter-at-a-time" approaches, which are slow, labor-intensive, and often fail to capture parameter interactions or efficiently map the "ALD window" (the range of conditions where self-limiting growth occurs). While Machine Learning (ML) and automation have advanced materials discovery, fully autonomous closed-loop platforms for gas-phase deposition are scarce, and to the best of the authors' knowledge, no ALD-specific Materials Acceleration Platforms (MAPs) have been reported to date.

Methodology
The authors present a modular, closed-loop Materials Acceleration Platform (MAP) integrating three core components:

Autonomous Hardware (ALDBot): A custom-built, fully automated Plasma-Enhanced ALD (PE-ALD) tool with full Python-based programmatic control. It executes PE-ALD cycles for TiO2 on silicon substrates, managing 12 distinct process parameters (e.g., precursor dose time, plasma duration, pressure, flow rates). The system performs up to 200 experiments per substrate to maximize throughput, discarding the first 9 cycles of each run to ensure steady-state growth rate measurement.
In Situ Characterization: The system utilizes real-time spectroscopic ellipsometry to monitor film thickness continuously during deposition, extracting the growth rate (Å/cycle) for each experimental condition without breaking vacuum.
Active Learning Framework: A Gaussian Process (GP) model serves as the surrogate model to learn synthesis-property relationships. The framework employs an information-theory-based acquisition function (Total Correlation) to sequentially select the next experiment. This function targets regions of maximum epistemic uncertainty (unexplored or poorly understood areas of the parameter space) rather than exploiting known optima, aiming to map the entire synthesis landscape efficiently.

The workflow operates in a closed loop: ALDBot executes an experiment $\rightarrow$ ellipsometry extracts growth rate $\rightarrow$ the GP model is updated $\rightarrow$ the acquisition function selects the next most informative parameter set $\rightarrow$ the cycle repeats.

Key Contributions

Autonomous ALD-MAP: The paper reports what is, to the best of the authors' knowledge, the first fully autonomous, closed-loop MAP specifically for ALD, demonstrating the feasibility of integrating automated synthesis, in situ characterization, and active learning for gas-phase deposition.
Modular Framework: The system is designed to be modular, allowing the deposition tool, characterization method, and ML model to be swapped or adapted for different material systems or deposition techniques (e.g., sputtering).
High-Dimensional Exploration: The authors successfully applied this framework to both a 4-dimensional (4D) subspace and the full 12-dimensional (12D) synthesis parameter space of PE-ALD TiO2.

Results

4D Subspace Performance: In a 4D parameter space (100 experiments), the active learning framework significantly outperformed random sampling. It achieved lower Root Mean Square Error (RMSE) (~0.25 Å/cycle vs. ~0.35 Å/cycle for random) and faster convergence in information gain. The GP model successfully interpolated the ALD window and predicted growth rates across the entire subspace with high accuracy, despite no experiments being performed at the specific "window" coordinates during training.
12D Full Space Performance: In the full 12D space (200 experiments), while complete mapping required more data than the budget allowed, the active learning strategy still provided valuable insights. The resulting high-uncertainty model enabled the extraction of physically interpretable outputs, such as parameter importance rankings via SHapley Additive exPlanations (SHAP).
Validation: Independent experimental sweeps on the most influential parameters (precursor dose time and process pressure) and a non-influential parameter (plasma purge time) confirmed that the model's predictions fell within the 95% confidence intervals of the experimental data.
Material Verification: Depositions guided by the model's identified "ALD window" produced phase-pure TiO2, confirmed by X-ray Photoelectron Spectroscopy (XPS) with an atomic ratio of 33.6% Ti and 66.4% O.

Significance and Claims
The authors claim that the integration of automated experimentation, in situ characterization, and active learning constitutes a powerful strategy for the autonomous exploration of gas-phase deposition processes. They demonstrate that this approach can efficiently map synthesis-property relationships and identify optimal process windows with significantly fewer experiments than random sampling or traditional methods.

Crucially, the paper argues that even when the parameter space is too large for complete mapping within a given experimental budget (as in the 12D case), the trained models yield physically interpretable insights (e.g., feature importance) that guide targeted optimization. The framework is presented as broadly applicable beyond the specific TiO2/PE-ALD system studied, offering a pathway to accelerate the development of functional materials in gas-phase deposition.

More like this