A Machine Learning Approach Capturing Hidden Parameters in Autonomous Thin-Film Deposition

TL;DR

The paper tackles the difficulty of achieving a fully autonomous ML-driven thin-film deposition pipeline in physical vapor deposition (PVD) due to hidden parameters from substrates and chamber conditions. It introduces a calibration layer and Gaussian Process Regression to model optical outputs, enabling an autonomous learning and testing loop within a high-throughput, in-situ measurement setup. The approach achieves silver films with optical properties within $2.5\%$ of targets on average after $2.3$ attempts, and demonstrates superior predictive accuracy and uncertainty quantification compared to a non-calibrated baseline. This work advances autonomous material discovery by combining hardware automation, active learning, and target-driven optimization, with broad applicability to other thin-film systems.

Abstract

The integration of machine learning and robotics into thin film deposition is transforming material discovery and optimization. However, challenges remain in achieving a fully autonomous cycle of deposition, characterization, and decision-making. Additionally, the inherent sensitivity of thin film growth to hidden parameters such as substrate conditions and chamber conditions can compromise the performance of machine learning models. In this work, we demonstrate a fully autonomous physical vapor deposition system that combines in-situ optical spectroscopy, a high-throughput robotic sample handling system, and Gaussian Process Regression models. By employing a calibration layer to account for hidden parameter variations and an active learning algorithm to optimize the exploration of the parameter space, the system fabricates silver thin films with optical reflected power ratios within 2.5% of the target in an average of 2.3 attempts. This approach significantly reduces the time and labor required for thin film deposition, showcasing the potential of machine learning-driven automation in accelerating material development.

Figures (4)

• Figure 1: A self-learning autonomous physical vapor deposition system for silver thin film growth. (a) The autonomous learning cycle incorporates (i) identification of the growth condition with the highest model uncertainty, (ii) sample growth, and (iii) updating the model with new data. (b) The autonomous testing cycle incorporating (i) prediction of the optimal growth condition that minimizes the loss function, (ii) sample growth, and (iii) assessment of success. (c) Schematic illustration of the autonomous deposition setup featuring robotic sample handling and in-situ optical characterization.
• Figure 2: Calibration layer accounts for variations in optical properties of silver films. (a-c) A linear fit between $\mathscr{R}_{443}$, $\mathscr{R}_{689}$, $\mathscr{R}_{817}$ at 5000 seconds and $\mathscr{R}_{689}$ at 1000 seconds. The moderate R-squared values indicate that the variances of $\mathscr{R}$'s at later times can be partially captured by $\mathscr{R}$ at 1000 seconds. (d) Translucent planes of time required to reach specified $\mathscr{R}_{443}$ values for given effusion cell temperatures and $\mathscr{R}_c$'s. $\mathscr{R}_c$ is negatively correlated with the time required to reach a given $\mathscr{R}_{443}$.
• Figure 3: Illustration of the data acquisition process for each iteration in the autonomous learning stage: the calibration layer is first grown to determine the $\mathscr{R}_c$, followed by measuring optical properties at all 5 wavelengths in cycles.
• Figure 4: Performance of autonomous learning in the fabrication of silver thin films. (a) Evolution of model predicted uncertainty, averaged over all $\mathscr{R}_{\lambda}$'s, during autonomous learning. (b) Convergence of the maximum uncertainty during autonomous learning. (c) Comparing the model prediction errors, defined as the differences between all measured and predicted $\mathscr{R}_{\lambda}$'s. The error distribution when adopting the calibration layer and autonomous learning is benchmarked against the case without these techniques.