Highly accurate classification of material types from spectroscopic ellipsometry heatmap measurements using deep learning

Abstract

Artificial intelligence (AI) applied to spectroscopic ellipsometry (SE) opens further possibilities for rapid and reliable optical metrology. Here, we present a deep learning model based on a convolutional neural network that classifies six thin-film material categories (i.e., amorphous, crystal, organic, metal, transparent conductive oxide, and transparent materials) from SE heatmaps measured over wide ranges of incident angles. Such classification is essential for developing AI models tailored to each material category. To construct a reliable model, an ultra-large dataset of 30 million hypothetical optical functions, generated from parameterized dielectric function models, was used for pre-training, followed by fine-tuning with real-material datasets. Validation with 45 experimental samples yielded an average absolute accuracy of 86.7% and a top-2 accuracy of 95.6%, with some categories reaching 100% in top-2 accuracy. Moreover, the analysis is completed within only a few seconds, highlighting its suitability for rapid and practical applications. These results demonstrate the effectiveness of machine learning in achieving practical material classification and underscore its potential to advance AI-driven optical metrology.