Matching parameter estimation for high power Inductively coupled plasma sources using Machine learning techniques

Abstract

Inductively coupled plasma�or ICP sources form a basis for multiple applications ranging from�semiconductor fabrication�to reliable heating systems for�tokamak�machines. To meet the functional requirements, ICP sources need efficient plasma formation utilizing the various input parameters. Operation of ICP sources is a complex and challenging task since it involves scanning a wide multi-dimensional parameter space involving filament bias, radio frequency (RF) power, gas pressure, matching parameters, and other system configurations. The foremost challenge is to maximize the coupling of RF power in the ion source for efficient plasma formation. Standard ICP sources use a matching network that consists of variable capacitors to compensate for plasma inductance to enable maximum power coupling. Identification of an accurate set of matching parameters for high power sources is a complex task and is generally driven by operator experience which is established after years of operations. Due to these challenges, recent developments in the area of machine learning can be utilized for identifying the underlying model function to make accurate predictions and explore an alternative approach to the existing Physics-electrical models developed for the estimation of matching parameters for�plasma sources. The present work attempts to perform a data-driven model discovery for the identification of appropriate matching parameters utilizing�machine learning algorithms. In this work, ROBIN, a high-power ICP source that operates with a 1MHz, 100 kW RF generator is considered which has been operational since 2011 and has generated a considerable database. This database can be utilized for training/developing data-driven models for the estimation of matching parameters for ensuring better power coupling. The paper describes the development of two data-driven regression models for predicting the coupling efficiency in terms of power factor (denoted by Cos) and the capacitor values based on input parameters utilizing well known algorithms such as�support vector machine, random forest and�neural networks. Emphasis has been laid on developing the models using parameters that are tuneable externally. Also, the effect of system configurations on parameter prediction is investigated. The developed machine learning-based models have achieved test accuracy scores of 0.93 and 0.91 for predicting Cos�and capacitor values respectively. The paper presents the training and optimization process for various machine and�deep learning algorithms�in detail.