Search

In the manufacturing process of electrical devices, ensuring the cleanliness of technical surfaces, such as direct bonded copper substrates, is crucial. An in-line monitoring system for quality checking must provide sufficiently resolved lateral data in a short time. UV hyperspectral imaging is a promising in-line method for rapid, contactless, and large-scale detection of contamination; thus, UV hyperspectral imaging (225–400 nm) was utilized to characterize the cleanliness of direct bonded copper in a non-destructive way. In total, 11 levels of cleanliness were prepared, and a total of 44 samples were measured to develop multivariate models for characterizing and predicting the cleanliness levels. The setup included a pushbroom imager, a deuterium lamp, and a conveyor belt for laterally resolved measurements of copper surfaces. A principal component analysis (PCA) model effectively differentiated among the sample types based on the first two principal components with approximately 100.0% explained variance. A partial least squares regression (PLS-R) model to determine the optimal sonication time showed reliable performance, with R2cv = 0.928 and RMSECV = 0.849. This model was able to predict the cleanliness of each pixel in a testing sample set, exemplifying a step in the manufacturing process of direct bonded copper substrates. Combined with multivariate data modeling, the in-line UV prototype system demonstrates a significant potential for further advancement towards its application in real-world, large-scale processes.

Hyperspectral imaging and reflectance spectroscopy in the range from 200–380 nm were used to rapidly detect and characterize copper oxidation states and their layer thicknesses on direct bonded copper in a non-destructive way. Single-point UV reflectance spectroscopy, as a well-established method, was utilized to compare the quality of the hyperspectral imaging results. For the laterally resolved measurements of the copper surfaces an UV hyperspectral imaging setup based on a pushbroom imager was used. Six different types of direct bonded copper were studied. Each type had a different oxide layer thickness and was analyzed by depth profiling using X-ray photoelectron spectroscopy. In total, 28 samples were measured to develop multivariate models to characterize and predict the oxide layer thicknesses. The principal component analysis models (PCA) enabled a general differentiation between the sample types on the first two PCs with 100.0% and 96% explained variance for UV spectroscopy and hyperspectral imaging, respectively. Partial least squares regression (PLS-R) models showed reliable performance with R2c = 0.94 and 0.94 and RMSEC = 1.64 nm and 1.76 nm, respectively. The developed in-line prototype system combined with multivariate data modeling shows high potential for further development of this technique towards real large-scale processes.

To correctly assess the cleanliness of technical surfaces in a production process, corresponding online monitoring systems must provide sufficient data. A promising method for fast, large-area, and non-contact monitoring is hyperspectral imaging (HSI), which was used in this paper for the detection and quantification of organic surface contaminations. Depending on the cleaning parameter constellation, different levels of organic residues remained on the surface. Afterwards, the cleanliness was determined by the carbon content in the atom percent on the sample surfaces, characterized by XPS and AES. The HSI data and the XPS measurements were correlated, using machine learning methods, to generate a predictive model for the carbon content of the surface. The regression algorithms elastic net, random forest regression, and support vector machine regression were used. Overall, the developed method was able to quantify organic contaminations on technical surfaces. The best regression model found was a random forest model, which achieved an R2 of 0.7 and an RMSE of 7.65 At.-% C. Due to the easy-to-use measurement and the fast evaluation by machine learning, the method seems suitable for an online monitoring system. However, the results also show that further experiments are necessary to improve the quality of the prediction models.