Effective Enhancement for Printed Circuit Board Imaging in Near-Field Scanning Microwave Microscopy

Abstract

Near-field microwave microscopy (NSMM) is a promising technique for the non-destructive, high-resolution imaging of electrical and dielectric properties at the microscale. However, its performance is highly sensitive to the probe-to-sample distance, often requiring extremely close proximity, which limits its practical application in device manufacturing, especially in scenarios involving coatings and packaging. In this study, we propose a distance inversion method based on a dual-port symmetrical microwave probe to improve imaging performance at larger, safer scanning distances. This method utilizes the correlation between probe height and resonant frequency to compensate for distance-induced signal distortions. The experimental results demonstrate that even at a probe–sample distance of 80 µm, clear and distinguishable NSMM images of printed circuit boards (PCBs) can be obtained. The imaging resolution reached 13 µm. The defect structure with dimensions of 130 × 130 µm² on the PCB was successfully identified. The signal-to-noise ratio was significantly enhanced after applying the correction method. This approach not only improves the robustness and flexibility of NSMM in industrial scenarios but also extends its applicability to packaged or coated electronic devices, offering a valuable tool for advanced non-destructive testing.

1. Introduction

Near-field scanning microwave microscopy (NSMM) is a microscopic technique that uses evanescent waves to detect microstructures [1,2]. To obtain high-measurement sensitivity, the NSMM system can be built with the coaxial resonant cavity of high-quality factors [3]. The resonant frequency shift from the probe–sample interaction reflects the microstructure features of the sample [4,5,6,7]. In the circuits industry [8], including the manufacture and detection of printed circuit boards (PCBs), optical images are usually applied for inspection [9,10,11]. In recent years, deep learning and reinforcement learning have made significant progress in high-resolution imaging and intelligent control. For example, Xu et al. [12] proposed a reinforcement learning-based multi-agent system control strategy, which has potential applications for optimizing system parameters and improving computational efficiency. Jia et al. [13] implemented defect detection in photovoltaic modules using an improved VarifocalNet model, enhancing detection accuracy and demonstrating the potential of deep learning in non-destructive testing (NDT). Additionally, Luan et al. [14] and Yu et al. [15] applied deep learning to optimize signal processing in ultrasound microvessel imaging and denoising, improving imaging resolution and signal-to-noise ratio. These research advancements provide valuable references for the optimization of near-field scanning microwave microscopy (NSMM), particularly in intelligent imaging algorithms, noise suppression, and high-resolution imaging. The physical properties of PCBs, like conductivity and dielectric constant, can be obtained from the resonant frequency distribution in NSMM measurements, and defects can be recognized with NSMM imaging. Ben Martin and Ian Jeffrey developed the Cycle-GAN method to calibrate microwave scan images based on deep learning models [16]. Researchers have proposed denoising methods based on mathematics and statistics [17,18,19]. To control tip–sample distance precisely, various distance-regulation mechanisms have been developed. Among them, shear-force feedback and tuning-fork-based feedback are two widely used approaches [20,21,22]. The shear-force method utilizes lateral oscillations of the probe near the sample surface, detecting damping changes to maintain a fixed height. However, it suffers from limitations such as a low signal-to-noise ratio on rough surfaces, a slow response time, and reduced effectiveness under vacuum or cryogenic conditions. On the other hand, tuning-fork feedback relies on piezoelectric vibrations of a quartz tuning fork to detect tip–sample interactions. While it offers mechanical stability and avoids the need for optical detection, it is constrained by a low feedback bandwidth and lower sensitivity compared to shear-force systems. These limitations highlight the need for more robust and responsive distance-control techniques to further improve NSMM imaging resolutions and measurement stability.

We proposed a unique method to improve the image quality of circuit structures by leveraging the characteristic correlation between probe height and resonant frequency in microwave near-field scanning. By analyzing this relationship, we can effectively perform distance inversions to correct for topographical influences, thereby enhancing the accuracy and contrast of subsurface or surface imaging. This technique not only improves spatial resolution in complex circuit environments but also enables the more reliable characterization of buried defects, dielectric distributions, and conductive pathways. Its potential applications include the non-destructive testing of integrated circuits, high-frequency device inspection, and advanced material analysis in both research and industrial settings.

2. Materials and Methods

A schematic diagram of the NSMM system is shown in Figure 1a. It consists of ① a λ⁄4 coaxial resonant cavity with an extended probe, ② an X−Y−Z motorized platform, ③ an anti-vibration base, ④vector network analyzer, ⑤ the sample, ⑥ a light source, ⑦ CCD camera, and ⑧ a laptop. In this system, a CCD camera is applied to monitor the probe–sample distance, and electrical parameters of the resonant cavity can be obtained through VNA, including resonant frequency (fr), quality factors (Q), and S parameters [23]. Figure 1b shows a homemade NSMM system in our laboratory. In this system, the vector network analyzer (VNA, MS46122B co. Anritsu Atsugi Japan) is applied with the frequency range of 1MHz~8GHz. The VNA serves as the signal source, feeding energy into the coaxial resonant cavity and receiving the signals reflected from the sample surface. The X-Y-Z three-axis motor (Zaber, XLSM025A Vancouver Canada) has a minimum step size of 0.047625 µm. For the NSMM system, a sensitive response between the probe and the sample is important [24]. We used a probe with a prototype that was a quarter-wavelength coaxial cavity. As shown in Figure 1c, it is a dual-port symmetrical device, which has a stable operating mode before it becomes close to the sample; however, based on the perturbation theory, touching the sample will break the original stable resonant operation mode. We constructed a cartesian coordinate system on the surface of the sample, with the probe in the +z direction of the sample. The point scanning and surface scanning of the NSMM were obtained with motors using different motional modes.

For point scanning, the distance between the probe and the sample must be soft contact, which means the probe of the resonant cavity has just made physical contact with the sample but has not yet caused the sample to deform. Then, the sample moves to a specified position along the +z direction, and the resonant frequencies fr are recorded [25]. For surface scanning, the distance between the probe and the sample is fixed at a certain distance to ensure that the probe does not touch and destroy the surface of the sample while moving on the x−y plane. Usually, the distance is chosen at the soft contact position with an extra ∆z; here, the maximum value of ∆z is 100 µm to ensure that the measurement remains within the near-field region. Therefore, the results are obtained for the surface scanning with responses from the resonant cavity at each scan coordinate point of the sample surface.

Figure 2 shows the relationship between the resonant frequency and probe–sample distance in an NSMM point scanning experiment. In this point-scanning experiment, the curve reflects the response between the probe and the sample, and the horizontal interval represents the distance change between the probe and the sample. The NSMM image can be obtained with higher quality when the probe–sample distance falls to the rapidly growing part, and the mapping shows lower quality when this distance falls to the approximately linear part. Three intervals with equal ranges are marked with different colors on the curve in Figure 2. The experimental distance between the probe and the sample in region I (yellow, 65 < z < 85 µm) can strictly obey the requirement for non-destructive testing. However, the tip–sample distance in region II (green, 25 < z < 45 µm) and region III (blue, 0 < z < 20 µm) were expected to be greater in the near-field microwave testing. Higher quality for NSMM imaging can be obtained in these intervals of region II and region III. The application for NSMM is greatly expanded by drawing distinct images in region II or region III. All needed data can be obtained by non-destructive microwave testing. (e.g., in region I).

As the tip-to-sample distance changes, the coupling between the tip and the sample also changes, which, in turn, causes a shift in the resonant frequency of the cavity. According to the definition of the resonant frequency, as shown in Equation (1), the resonant frequency exhibits a power–law relationship with the tip-to-sample distance. Based on the discrete data obtained from point-by-point scanning, which records how the resonant frequency varies with the tip–sample distance, we fit the data to a power function and extracted the corresponding coefficients A and B.

$$fr=1/2\pi \sqrt{LC}=A\sqrt{d}+B$$

(1)

The function obtained from point-by-point scanning was used to derive its inverse function, denoted as fr⁻¹(d). By substituting the resonant frequency matrix acquired from surface scanning into fr⁻¹(d), a distance parameter matrix was obtained. Each element of this matrix was then subtracted by a fixed value so that the resulting distance values fell within the ranges 65 < z < 85 µm and 0 < z < 20 µm, corresponding to regions II and III, respectively. The adjusted distance matrix was subsequently substituted into the original power function fr(d), thereby correcting the surface scan image from the relatively distant region I to the more proximate regions II and III.

First, denoising filters are applied to reduce the increasing noise signal. The Gaussian filter and mean filter were first applied to the NSMM mapping in region I of Figure 2. Next, we backtracked from region I to region II or region III along the tip–sample curve to obtain the modified NSMM imaging.

There are three steps in this backtrack for NSMM mapping enhancement: (i) First, for isotropic materials, based on the principle that a general point of the point scanning experiment on the surface can represent the nature of the entire sample, the reverse function fr⁻¹(d) is obtained with numerical fitting. (ii) Second, each coordinate point in the raw NSMM image of region I is brought into this reverse function fr⁻¹(d) between the probe–sample distance and X-Y coordinates, and then this distance is reduced into region II and region III. This step can be obtained through subtraction. (iii) Third, the processed distance function is given into fr(d), and a modified image can be obtained. Therefore, the transformation from region I to region II is noted as a first-order modification, and the transformation from region I to region III is noted as a second-order modification.

3. Results

Figure 3a shows an optical photograph of “UESTC” letters hollowed on a steel mask. NSMM point scanning on this steel sample was applied to obtain the function. All figures were drawn using resonant frequency data. And we undertook data normalization, so all the labels of the color bar range from 0 to 1. The λ⁄4 coaxial resonant cavity works at around 2.25 GHz, and the diameter of the probe is 50 µm. In Figure 3b, the average distance from the probe to the steel mask is about 80 µm, which indicates that it is safe to keep the probe and the sample in soft contact. Although this tip–sample distance follows the requirement of non-destructive testing, it presents blurring and distortion in the original image, as shown in Figure 3b. In the NSMM imaging experiment, we reduced the tip–sample distance to the average of 36 µm, and a much more distinct steel sample was obtained, as shown in Figure 3c. Here, we intercepted the part of the image showing “TC” for comparison. According to the point scan function, the image after the first-order modification from the original NSMM distribution in Figure 3b is shown in Figure 3d. After the first-order modification, the tip–sample average distance was backtracked to region II from region. Similarly, Figure 3e shows the image after the second-order modification; therefore, the average tip–sample distance backtracked to region III. The image enhancement in Figure 3d,e is more significant than the original NSMM image in Figure 3b. The image in Figure 3e is even better than the NSMM image in Figure 3c, with a close tip–sample distance of 36 μm. This method indicates effective method enhancement in the surface mapping of NSMM with large and safe tip–sample distances.

Figure 4a shows an optical image of an octagonal inductance on a printed circuit board (PCB). The scanning area is inside the red box of Figure 4a. The base of the PCB is vinyl material, and it is covered with copper foil lines to form the coils. The tip diameter of the probe is 50 µm. Figure 4b shows the NSMM image at a corner of the sample with a tip–sample distance of 80 µm. Figure 4c is the image after the first-order modification, where the probe–sample distance of Figure 4c roughly backtracked to region II. After the second-order modification, the modified image is shown in Figure 4d with the tip–sample distance backtracked to region I. In the original NSMM image of Figure 4b, there is little difference in resonant frequency between the copper foil lines and the background of vinyl material. The boundary between the two materials in the image is blurred in Figure 4b. With the first- and second-order modification, the copper foil lines in the image of Figure 4c,d are much clearer, and the boundary between the copper and the vinyl material is more distinguished.

The data were extracted from red lines on each NSMM image in Figure 4b–d to quantize the enhancement of this method. Each value in the line was normalized according to Equation (2). The normalized denominator is the average value of these line data, and the normalized relationship between the resonant frequency and the position is shown in Figure 4e.

$$\widetilde{f_r}={\displaystyle \frac{fr\left(x^{\prime }\right)}{\mathrm{average}\left(fr\right)}}$$

(2)

In Figure 4e, the purple line represents the normalized original data from the line in Figure 4b; the green line represents the normalized first-order-modified data from the line in Figure 4c; and the yellow line represents the normalized second-order-modified data from the line in Figure 4d. Each curve has three peaks at the same x position, which correspond to the boundary between the copper lines and the vinyl base in the fr distribution. These three curves show a similar tendency from the same line marked on the surface of the fr distribution in Figure 4. There are five copper foil lines on this line, with scanning data corresponding to the five peaks on each curve in Figure 4e. The peak in the original NSMM curve shows the lowest difference, which indicates that the copper foil line and the background material are hardly distinguished in the original image. As shown in the other two curves of Figure 4e, the difference between the copper and vinyl after the first-order and second-order modification is much more obvious. The near-field microwave image of microstructures composed of complex materials can be effectively distinguished with this method.

4. Discussion

Figure 5a shows the photo of a defective printed circuit board (PCB), and the broken stripe on the PCB is shown by the red box. Figure 5b shows the raw NSMM imaging, and Figure 5c shows modified NSMM imaging with a second-order modification. The tip diameter of the probe is 10 µm. The boundaries at the defect stripe in Figure 5c can be recognized much more obviously than those in Figure 5b, especially for the defect on the PCB in red circles. As a result, it can be determined that the stripe has been broken. It indicates that circuit detection and circuit layout inspection can be performed using modified NSMM imaging.

The signal-to-noise ratio (SNR), an important indicator in image processing, was introduced to evaluate the performance of this method quantitatively. SNR is usually applied to represent the ratio of signal-to-noise in an image, and the expression of this is shown in Equation (3).

$$\mathrm{SNR}=10\times {\log }_{10}\left({\displaystyle \frac{\mathrm{Signal}\mathrm{power}}{\mathrm{Noise}\mathrm{power}}}\right)=10\times {\log }_{10}\left({\displaystyle \frac{\sum _1^{\mathrm{m}}\sum _1^{\mathrm{n}}{\mathrm{img}}_{\mathrm{processed}}^2}{\sum _1^{\mathrm{m}}\sum _1^{\mathrm{n}}{({\mathrm{img}}_{\mathrm{raw}}-{\mathrm{img}}_{\mathrm{processed}})}^2}}\right)$$

(3)

In Equation (3), the processed image supplies the signal, and the difference between the raw image and the processed image acts as the noise. Since the gray value of the images is generally in the range of 0~255, the image is normalized from the frequency range to the gray value range. In the enhancement experiments on “UESTC” mask NSMM imaging, the SNR between the raw image and the filtered image is 19.76 dB, the SNR of the filtered image and first-order-modified image is 33.03 dB, and the SNR of the filtered image and second-order-modified image is 19.28 dB. In the experiment on the octagonal PCB, the SNR between the raw image and the filtered image is 20.11 dB, and the SNR of the filtered image and second-order-modified image is 20.44 dB. In the mapping on the defective PCB, as shown in Figure 5, the SNR of the raw image and filtered image is 23.66 dB, and the SNR of the filtered image and second-order-modified image is 18.72 dB. These results show that the enhancement method in our work can perform quite effectively from both visual imaging and data perspectives.

5. Conclusions

We propose a unique method that enables accurate NSMM imaging at a large and safe scanning distance of up to 80 µm, avoiding potential sample damage typically associated with probe–sample distances of only a few microns. The NSMM images of a “UESTC” steel mask were modified and enhanced effectively with the scanning probe–sample distance at 80 µm. A blurred NSMM image of PCB was modified successfully into a distinguished image, and defects could be clearly recognized. This method can be used to detect internal defects in packaged circuits, such as potential electrical faults in wafer-level packaging (WLP) or chip-scale packaging (CSP). It is also suitable for the quality inspection of flexible electronic devices or printed sensors, particularly enabling non-contact imaging without damaging the protective layers.