Autonomous Sampling and SHAP Interpretation of Deposition-Rates in Bipolar HiPIMS

TL;DR

This paper employs autonomous Bayesian sampling and SHAP interpretability to analyze bipolar HiPIMS processes, linking process parameters to deposition rates and exploring optimization strategies for coating technologies.

Contribution

It introduces a novel workflow combining autonomous experimentation with interpretable machine learning to understand and optimize complex plasma deposition processes.

Findings

Minimal improvements in deposition rates from positive pulse adjustments

SHAP analysis links process variations to physical mechanisms like plasma ignition

Quenching of plasma reduces the effectiveness of positive pulses

Abstract

High-power impulse magnetron sputtering (HiPIMS) offers considerable control over ion energy and flux, making it invaluable for tailoring the microstructure and properties of advanced functional coatings. However, compared to conventional sputtering techniques, HiPIMS suffers from reduced deposition rates. Many groups have begun to evaluate complex pulsing schemes to improve upon this, leveraging multi-pulse schemes (e.g. pre-ionization or bipolar pulses). Unfortunately, the increased complexity of these pulsing schemes has led to high-dimensionality parameter spaces that are prohibitive to classic design of experi-ments. In this work we evaluate bipolar HiPIMS pulses for improving deposition rates of Al and Ti sputter tar-gets. Over 3000 process conditions were collected via autonomous Bayesian sampling over a 6-dimensional parameter space. These process conditions were then interpreted using Shapley Additive Explanations (SHAP), to deconvolute complex process influences on deposition rates. This allows us to link observed var-iations in deposition rate to physical mechanisms such as back-attraction and plasma ignition. Insights gained from this approach were then used to target specific processes where the positive pulse components were expected to have the highest impact on deposition rates. However, in practice, only minimal improve-ments in deposition rate were achieved. In most cases, the positive pulse appears to be detrimental when placed immediately after the neg. pulse which we hypothesize relates to quenching of the afterglow plasma. The proposed workflow combining autonomous experimentation and interpretable machine learning is broad-ly applicable to the discovery and optimization of complex plasma processes, paving the way for physics-informed, data-driven advancements in coating technologies.