Hyperspectral imaging-based detection technology for immersion filling states in micro/nano structures

The immersion filling states of liquid medium in micro/nano structures directly affect the performance and quality of the devices, making precise characterization imperative. Traditional detection methods are limited by precision, efficiency, and applicability, making it challenging to meet the requirements for characterizing filling states at the micro- and nano-scale. To overcome these limitations, a technique for detecting filling states in micro/nano structures based on hyperspectral imaging and deep learning was proposed. By employing a Convolutional Neural Network (CNN) to integrate spatial-spectral multidimensional information of hyperspectral data, pixel-wise semantic segmentation is achieved, which provides an effective solution for high-precision and non-destructive detection of liquid filling states at the micro- and nano-scale. Leveraging a high-resolution medium detection system, synchronous in-situ acquisition of hyperspectral and microscopic RGB images was realized for immersion filling states on a silicon grating template with a period of 2 μm. After standard preprocessing including dark-field correction, spectral denoising, and normalization, a dataset containing 128 spectral channels and 708,400 pixel samples was constructed. Deep learning experiments were then conducted using a typical three-dimensional CNN (3D-CNN). The results demonstrate that the proposed technique effectively detects the filling states. The model achieved an overall accuracy (OA) of 95.15% on the test set, which is nearly 10% higher than the traditional method combining microscopic RGB images with K-means clustering (OA = 85.62%). Furthermore, the model exhibited good generalization and robustness when tested on new grating pattern specimens not involve in training.