In‑situ Machine Learning Enables Calibration‑Free Composition Mapping in Magnetron Co‑Sputtering Labs

Global: In‑situ Machine Learning Enables Calibration‑Free Composition Mapping in Magnetron Co‑Sputtering Labs

A research team has unveiled a self‑driving laboratory that uses magnetron co‑sputtering to produce accurate composition maps of multi‑element thin films without the need for external calibration. The system integrates automation, in‑situ quartz‑crystal microbalance sensors, and machine‑learning algorithms to predict material distributions across a substrate in real time.

New SDL Architecture

Self‑driving labs have traditionally focused on solution‑based synthesis, which limits access to the expansive chemical space of inorganic compounds. By adopting magnetron co‑sputtering, the new platform expands experimental capabilities to include a broader range of materials while maintaining high levels of automation.

In‑situ Sensing and Machine Learning

Quartz‑crystal microbalance sensors positioned within the sputter chamber continuously record deposition rates. These sensor outputs are modeled as functions of sputtering pressure and magnetron power using Gaussian processes (GPs) trained through active learning, allowing the system to adapt to varying source conditions.

Hybrid Modeling Approach

The trained GPs are combined with a geometric model of the deposition flux, enabling interpolation of deposition rates from each source at any point on the substrate. This hybrid framework translates raw sensor data into spatial composition maps without resorting to time‑consuming ex‑situ analyses.

Active Learning Performance

Among several acquisition functions evaluated, a fully Bayesian GP implementation—Bayesian Active Learning MacKay (BALM)—demonstrated the strongest performance, learning the deposition rate for a single source after just 10 experimental runs.

Experimental Validation

The predicted composition distributions for co‑sputtered films were experimentally verified, confirming the accuracy of the in‑situ approach and its ability to reproduce expected material gradients across the sample.

Broader Impact

By eliminating extensive post‑deposition characterization and calibration steps, the framework dramatically accelerates materials discovery workflows. The method showcases the potential of machine‑learning‑guided self‑driving labs to expedite exploration of inorganic thin‑film chemistries.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.