Machine learning enhanced real-time temperature measurement for millisecond-scale plasma processing using optical-interference contactless thermometry (OICT)

Real-time wafer temperature measurement is essential in plasma processing to enable feedback control and improve process repeatability. Accurate measurement of wafer surface temperature remains challenging in plasma environments, particularly for millisecond-scale processes that induce pronounced temperature gradients across the wafer thickness. The authors have developed an optical-interference contactless thermometry (OICT) for measuring transient temperature profiles at both the wafer surface and within the wafer during plasma processing. This study proposes a novel analysis method for OICT using machine learning (ML) to facilitate real-time temperature measurement across a wide range of processing conditions. An indirect ML modeling strategy was employed utilizing the fundamental physical relationship between the optical thickness and temperature–depth profile within the wafer. This ML framework enables temperature prediction at arbitrary depths within the wafer and at arbitrary time points during the processing. Through a comprehensive comparison of multiple ML models, multilayer perceptron (MLP) is identified as the most ideal model for OICT. Following simulation-based training, the MLP model demonstrated an average prediction temperature error of less than 0.1 K, with a maximum error under 1 K for the simulation data. Experimental demonstrations using an atmospheric pressure thermal plasma jet (TPJ) confirm that the proposed ML-prediction method enables real-time temperature measurement over a broad scan speed range of 100–1500 mm/s, with an analysis time of 0.2–0.6 s. This method effectively reproduces the temperature analysis results of OICT manual fitting, maintaining a temperature difference below 1.6 K throughout the entire wafer depth when the surface temperature peaks.