BROWSE
Detail
Journal Article
Thin-Film Thickness Estimation From Optical Images Using Neural Networks
Cited 0 time in 
Share 
Share
- Authors
- Eun-Hee Kim, Munyoung Lee, Minoh Jeong, Kyu-Sung Lee, Minki Kim
- Issue Date
- 2026-08
- Citation
- IEEE Transactions on Instrumentation and Measurement, v.75, pp.1-12
- ISSN
- 0018-9456
- Publisher
- IEEE
- Language
- English
- Type
- Journal Article
- Abstract
- This article presents an artificial intelligence (AI)-based metrology framework that predicts SiO2 thin-film thickness directly from microscope color observations and acquisition metadata. While conventional instruments, such as spectroscopic ellipsometers and $\alpha $ -step profilers, offer high accuracy, their cost, calibration overhead, and throughput limitations make rapid in-line deployment challenging. To address these issues, we propose a compact feed-forward neural network (FNN) that maps RGB statistics and metadata (e.g., magnification, illumination level, and microscope maker) to thickness without reconstructing the full optical pipeline. The network is trained on a curated dataset of 2520 measurements acquired under diverse conditions, with ellipsometry serving as ground truth. Our approach achieves about 2.6 % average error in the most data-rich setting and remains within 10 %–15 % error under sparse training, demonstrating favorable data–accuracy scaling. Then, leveraging the trained regressor, we implement a prototype application that converts a single microscope image into pixel-wise thickness maps and 3-D surface visualizations within seconds. This software shows close agreement with a commercial $\alpha $ -step profiler for most samples, indicating strong practical viability as a fast, noncontact auxiliary tool for semiconductor process monitoring. Finally, we discuss anticipated improvements, including flat-field illumination correction, radiometric/colorimetric calibration, and cross-device/material generalization, to further enhance robustness and transferability. Overall, this work provides a simple, low-cost, and noncontact route to quantitative thin-film metrology from commodity microscopy, serving as a practical bridge between high-fidelity optical instruments and real-time in-line process monitoring.
- Keyword
- Feedforward neural networks (FNNs), noncontact measurement, optical microscopy, thin-film thickness estimation
- KSP Keywords
- Average error, Colorimetric calibration, Cross-device, Feedforward neural network(FNN), Film thickness, High-fidelity, Illumination Level, Illumination correction, In-line deployment, In-line process, Optical Microscopy