Active learning-driven uncertainty reduction for in-flight particle characteristics of atmospheric plasma spraying of silicon

TL;DR

This paper introduces an active learning framework using Gaussian Processes to improve prediction accuracy of in-flight particle characteristics in atmospheric plasma spraying of silicon, reducing uncertainty and enhancing efficiency.

Contribution

It is the first to apply active learning with Bayesian optimization in thermal spray, significantly improving prediction accuracy with limited experimental data.

Findings

52.9% reduction in RMSE after optimization

8.5% increase in R2 score

Three-fold improvement in prediction accuracy at test points

Abstract

In this study, the first-of-its-kind use of active learning (AL) framework in thermal spray is adapted to improve the prediction accuracy of the in-flight particle characteristics and uses Gaussian Process (GP) ML model as a surrogate that generalises a global solution without necessarily involving physical mechanisms. The AL framework via the Bayesian Optimisation was utilised to: (a) reduce the maximum uncertainty in the given database and (b) reduce local uncertainty around a contrived test point. The initial dataset consists of 26 atmospheric plasma spray (APS) parameters of silicon, aimed at ceramic matrix composites (CMCs) for the next generation of aerospace applications. The maximum uncertainty in the initial dataset was reduced by AL-driven identification of search spaces and conducting six guided spray trails in the identified search spaces. On average, a 52.9% improvement (error reduction) of RMSE and an R2 increase of 8.5% were reported on the predicted in-flight particle velocities and temperatures after the AL-driven optimisation. Furthermore, the Bayesian Optimisation around a contrived test point to predict the best possible characteristics resulted in a three-fold increase in prediction accuracy as compared to the non-optimised prediction. These AL-guided experimental validations not only increase the informativeness of the limited dataset but is adaptable for other thermal spraying methods without necessarily involving physical mechanisms and underlying mechanisms. The use of AL-driven optimisation may drive the thermal spraying towards resource-efficiency and may serve as the first step towards fully digital thermal spraying environments.