Real-time monitoring of temperature distribution on silicon wafer using hybrid neural network-based regression model

Abstract

Thermal defects in semiconductor wafers introduce variations in surface temperature and thermal gradients, potentially degrading product quality and reducing yield during manufacturing. Real-time detection of localized heating and thermal imbalances is essential for process control, yet conventional temperature monitoring methods are limited. Point-based sensors offer sparse spatial coverage, and infrared thermal imaging is often unsuitable for extreme environments such as high-temperature or vacuum conditions commonly found in semiconductor processing. To address these limitations, this study presents a hybrid neural network (HNN) regression model that integrates a deep neural network (DNN) with convolutional neural network to reconstruct full-surface temperature distributions from localized temperature and gradient measurements. The model is trained on spatially distributed thermal data to learn complex relationships between localized inputs and overall thermal patterns. Experimental evaluation demonstrates the HNN model's superior predictive accuracy compared to a standard DNN model, achieving a 33.10 % reduction in root mean squared error and a 37.11 % reduction in mean absolute error. Additionally, 96.53 % of the predicted thermal maps reach a structural similarity index exceeding 0.9, indicating high-quality reconstructions. This approach enables real-time, thermal defect monitoring, enhancing stability and yield in advanced semiconductor manufacturing processes, and showcasing the potential of artificial intelligence in industrial applications.