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

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.

💡 Research Summary

The paper presents a fully autonomous physical vapor deposition (PVD) platform that integrates machine learning, robotics, and in‑situ optical spectroscopy to address two major challenges in thin‑film fabrication: (i) the need for a closed‑loop cycle of deposition, characterization, and decision‑making without human intervention, and (ii) the detrimental impact of hidden parameters (e.g., substrate surface condition, residual gases, minute temperature drifts) on model reliability.

The hardware consists of an ultra‑high‑vacuum chamber equipped with a 72‑slot robotic sample carousel, an effusion cell for silver deposition, and a set of five p‑polarized lasers (443, 514, 689, 781, and 817 nm) mounted on a CNC linear rail. The lasers illuminate each sample at a 45° incidence angle, and both transmitted (Pt) and reflected (Pr) powers are recorded. From these measurements the reflected power ratio R = Pr/(Pr+Pt) and absorptivity A are computed in real time. Data acquisition is organized into 98‑second blocks, during which all five wavelengths are measured sequentially, allowing up to 72 samples to be processed consecutively without manual handling.

To capture the effect of hidden variables, the authors introduce a “calibration layer.” After an initial 1000 s of growth at a fixed temperature (875 °C), the reflected ratio at 689 nm (denoted Rc) is measured. Statistical analysis of 24 calibration samples shows a moderate linear correlation between Rc and the final reflected ratios after 5000 s, indicating that Rc encodes information about the current substrate and chamber state. Consequently, Rc is treated as an additional input feature for the machine‑learning models.

Two Gaussian Process Regression (GPR) models with radial‑basis‑function kernels are trained: one predicts reflected ratios R, the other predicts absorptivities A. The input vector is x = (T, t, λ, Rc), where T is the effusion‑cell temperature, t is growth time, λ is the wavelength, and Rc is the calibration measurement. The models output both a mean prediction (µ) and an uncertainty (σ).

Learning proceeds in two phases. First, a predefined grid of nine temperature points (820–880 °C in 7.5 °C steps) is used to generate an initial dataset (predefined learning). Then, an autonomous learning loop runs for eight iterations. In each iteration the GPR model evaluates the entire feasible space and selects the temperature that maximizes predictive uncertainty across all wavelengths and time points (Eq. 4). The selected temperature is used for a full growth run up to a temperature‑dependent maximum time tmax(T) (derived from an exponential fit to growth rate). After each run, the calibration layer is measured, the full optical dataset is collected, and the GPR models are updated with the new observations. This uncertainty‑driven exploration rapidly reduces the average model uncertainty to ≈0.032 after eight iterations, indicating convergence.

Following model convergence, the system enters the testing stage. Users specify target reflected ratios at one or more wavelengths. For each target, the GPR model searches the (T, t) space (given the measured Rc) to find the combination that minimizes a loss function based on the predicted mean and uncertainty. The system then executes the deposition with the suggested parameters and verifies the outcome. Across a series of random single‑wavelength targets, the autonomous platform achieves the desired reflected power ratios within 2.5 % of the target after an average of 2.3 deposition attempts.

Benchmarking shows that both the calibration layer and the uncertainty‑driven autonomous learning substantially lower prediction errors compared with a baseline that omits these components. The integration of high‑throughput robotics, real‑time spectroscopy, and probabilistic machine learning thus delivers a scalable, fully automated thin‑film workflow that dramatically cuts experimental time and labor.

The authors argue that the calibration‑layer concept is broadly applicable to other material systems where hidden variables are difficult to measure directly, and that GPR‑based active learning can be extended to multi‑objective optimization (e.g., simultaneously targeting optical, electrical, and structural properties). The work therefore represents a significant step toward autonomous materials discovery platforms capable of handling the intrinsic variability of complex deposition processes.