Defect Detection and Error Source Tracing in Laser Marking of Silicon Wafers with Machine Learning

Abstract

Laser marking on wafers can introduce various defects such as inconsistent mark quality; under- or over-etching, and misalignment. Excessive laser power and inadequate cooling can cause burning or warping. These defects were inspected using machine vision, confocal microscopy, optical and scanning electron microscopy, acoustic/ultrasonic methods, and inline monitoring and coaxial vision. Machine learning has been successfully applied to improve the classification accuracy, and we propose a random forest algorithm with a training database to not only detect the defect but also trace its cause. Four causes have been identified as follows: unstable laser power, a dirty laser head, platform shaking, and voltage fluctuation of the electrical power. The object-matching technique ensures that a visible image can be utilized without a precise location. All inspected images were compared to the standard (qualified) product image pixel-by-pixel, and then the 2D matrix pattern for each type of defect was gathered. There were 10 photos for each type of defect included in the training to build the model with various labels, and the synthetic testing images altered by the defect cause model for laser marking defect inspection had accuracies of 97.0% and 91.6% in sorting the error cause, respectively

1. Introduction

The selection of the appropriate laser marking technique depends on the specific needs of the application and desired outcome. The following are some key considerations: marking depth and intensity; material sensitivity; speed; and throughput and traceability requirements [1]. The laser marking of wafers can introduce various defects. Typical types include inconsistent mark quality (variations in depth), under- or over-etching, and misalignment. Thermal damage is another concern: excessive laser power or poor cooling can cause burning or melting of the surface, as well as warping or cracking of the wafer. Contamination from redeposited debris may also occur. Some lasers (e.g., CO₂) can induce microcracks or chipping in Si, whereas shorter-wavelength UV/excimer lasers (193 nm) have been shown to avoid subsurface microcracks [2]. The inspection technologies are listed as follows: Machine vision and optical AOI: Automated optical inspection (AOI) uses high-resolution cameras and image processing to detect visible flaws on the marked wafer. AOI systems can detect alignment errors, missing or irregular characteristics, surface scratches, and particulate contaminants [3]. Line- and area-scan cameras (sometimes with patterned lighting) allow fast full-wafer scans. For example, line-scan cameras can be used to inspect defects on uneven wafer surfaces accurately. In practice, dedicated vision tools are used for code reading and pattern checks; for instance, wafer ID readers and OCR cameras follow the SEMI T7/BC412 code standards to verify the legibility of DataMatrix or QR marks. Three-dimensional confocal profilometry: Optical profilometry or confocal microscopy provides three-dimensional measurements of mark topography [4]. Confocal laser scanning systems (such as Nikon’s NEXIV series) can capture the true height and position of the engraved characters. These 3D metrology tools measure the mark depth, width, and contrast to ensure that they meet the specifications. For example, Nikon’s VMZ-K3040 confocal system inspects wafer ID codes immediately after marking, providing real-time feedback on mark geometry and enabling high-throughput QC. Optical and scanning electron microscopy: Traditional optical microscopes (bright-field or dark-field) can be used for detailed visual inspection of laser marks, verifying edge quality, and looking for small surface defects. In research and failure analysis, digital holographic microscopy has been used to inspect the high-aspect-ratio features of wafers [5]. For nanoscale defects (e.g., very fine cracks or debris), scanning electron microscopy (SEM) can provide high-magnification views. SEM is a standard inspection mode (along with bright-field and dark-field imaging) for wafer defect review. Acoustic/ultrasonic methods: ultrasonic techniques are used for subsurface or internal defects. Scanning Acoustic Microscopy (SAM) (also known as C-SAM) employs focused ultrasound to image the hidden features. SAM is widely used in microelectronics QA because it can detect voids, delaminations, or cracks inside bonded wafers and packages [6]. In semiconductor testing, SAM can reveal internal cracks and assess bond integrity (e.g., die attach or flip-chip underfill). Although not typically inline, SAM provides a powerful non-destructive check for microcracks beneath the surface. Inline monitoring and coaxial vision: Some laser marking systems include integrated cameras or sensors for in-process monitoring. For example, inline “through-the-laser” cameras can track the beam position and mark formation in real time [7]. These systems can adjust the laser to fly or flag misfires immediately. An example is the RAYLASE RAYSPECTOR, a coaxial monitoring unit that provides a live video of the laser process and even performs code “OK/NOK” checks during the marking. Inline spectrometers or pyrometers can also be used to monitor plume emissions or heat, although they are more commonly used in other laser processing contexts. A femtosecond laser was used to manufacture an anti-reflective subwavelength structure (ASS) with ultrahigh transmittance on the surface of infrared window materials directly. [8] proposed a design, manufacturing, and characterization method that can produce an ultrahigh-performance infrared window by a femtosecond laser Bessel beam. The ablation characteristics and intrinsic mechanism of alumina ceramics under the ablation of the long focal condition CW laser were comprehensively revealed through systematic process comparison experiments, high-speed shadow imaging, and theoretical simulation to ensure the design of ceramic-based high-energy laser protection materials [9].

Quality control in high-volume fabs is highly automated. Laser marking stations are typically integrated with wafer handlers and robots (e.g., FOUP loaders and conveyors) to eliminate manual handling [10]. For example, Nikon’s system automatically transfers wafers from a pod to a confocal meter without human contact. Many marking tools include built-in visual alignment tools [11]. For example, Pentamaster’s PM93 handler uses a granite anti-vibration base and advanced machine vision alignment, such that each wafer is perfectly positioned before marking. After marking, these systems capture a live video feed for the operator and perform an automatic post-mark inspection to verify code accuracy and quality [12]. Artificial intelligence (AI) is playing an increasingly important role. Modern vision systems use deep learning neural networks to improve defect detection and code reading. Keyence and Cognex, for example, offer “smart” vision tools that can be trained on large datasets to distinguish true defects from allowable variations [13]. These AI-based inspectors can learn complex patterns (even on transparent or multilayer wafers) and achieve a higher accuracy in picking up scratches, stains, or alignment errors. Cognex’s In-Sight D900 uses deep learning to automatically screen many wafers for anomalies (even amid complex layer backgrounds), passing ambiguous cases to a human reviewer [14]. Similarly, AI-based segmentation (e.g., SolVision) has been demonstrated to find very fine scratches and polishing defects on wafer surfaces that rule-based AOI might miss [15]. In summary, the combination of automated handling, high-speed vision sensors, and AI algorithms enables real-time sorting of wafers and rapid feedback on marking quality. These examples demonstrate the industrial trend toward fully automated vision-guided marking lines with feedback. Fabs can enforce marking quality to meet SEMI standards and minimize scrap by combining metrology (confocal or camera), AI software, and robot handling. However, these state-of-the-art technologies do not possess an error source tracing function, which is an improvement of this study.

2. Materials and Methods

A random forest classifier algorithm was chosen because of its powerful ensemble machine learning, which has been used primarily for classification tasks. Random forest is widely used because of its robustness, accuracy, and ability to handle complex high-dimensional data. It operates by constructing a multitude of decision trees during training and makes predictions based on the majority vote of those trees. It works as a flow function: Ensemble of decision trees: Random Forest builds multiple decision trees using different random subsets of the training data and features. Each tree is trained independently and may view the data differently. Bootstrapping: Each tree is trained on a bootstrapped sample, meaning that it is built from a random sample of data selected with replacement. This ensured the diversity among the trees. Feature randomness: during the tree-building process, at each split, only a random subset of features is considered, further increasing the diversity and reducing the correlation between trees. Voting mechanism: For classification, each tree votes for a class, and the class with the most votes is the final prediction. For regression analysis, the average of all tree predictions was used [16].

The key steps of this algorithm are as follows: Randomly select subsets of data and features for each tree. A decision tree is built for each subset. A new data point passes through all trees and collects their predictions. Aggregate predictions (majority vote for classification, mean for regression). Advantages of it are it reduces overfitting: By averaging the predictions of many trees, random forest reduces the risk of overfitting that is common with single decision trees; and handles missing data: it can maintain accuracy even when some features are missing. Feature importance: random forest can provide measures of feature importance, helping identify the variables that are most influential in the data. Versatility: it works well for both classification and regression problems and can handle large datasets with high dimensionality.

When using the random forest algorithm to solve regression problems, we use the mean squared error (MSE) to determine how the data branch from each node.

$$MSE={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N{\left(fi-yi\right)}^2$$

where N is the number of data points, fi is the value returned by the model, and yi is the actual value of the data point i.

A random forest python code is applied; the selective parameters are set as follows. The number of decision trees in the forest: 1000, maximum depth of each tree: 4, to prevent overfitting, minimum number of samples required to split an internal node: 2, minimum number of samples required to be at a leaf node: 1, number of features considered at each split: sqrt, using bootstrap samples while building trees and random seed: 42 to ensure reproducibility. In total, 1000 pieces of a tree from the iris databases with 4 various classes were employed to test the functioning and accuracy of the code. The 4 classes of the random forest decision boundaries are illustrated in Figure 1 and the classifications have an accuracy of 89.0 and kappa = 0.853; the entire computing took around 20 min on a regular CPU desktop computer with 64 GB RAM.

The specifications of the experimental camera and laser marking machine are listed in

Table 1. Specification of argus microscope.

Item	Specification
Frame Rate	60 fps
Image Resolution	1920 × 1080 pixel
Sensor	1/2.8” SONY
Pixel Size	2.9 × 2.9 μm
Picture	SD card 38M
Light	Ring Light 52D LED
Measurement	Via software
Output Interface	HDMI + USB

and

Table 2. Specification of MR. CARVE M1 Laser Marking Machine.

Laser Power	10 W
Repeat Accuracy	≤0.001 mm
Marking Depth	0.012–0.2 mm
Marking Precision	≤0.001 mm
Marking Speed	≤10,000 mm/s
Laser Wavelength	1064 nm
Marking Area	70 × 70 mm
Marking Line Width	0.001–0.05 mm

, respectively. Photographs of the microscope and laser marking machine are shown in Figure 2.

The scanned picture is set as 8 bits grayscale in 1920 × 1080 pixels; this setting implies the brightness at each pixel could have 256 difference levels to reflect the marking status. Firstly, the digital number (DN) of pixels in the upper left and lower right of the image was identified and defined as the region of interest (ROI) by the Canny edge detection approach since it offers sharper and more continuous edges than other approaches [17]. The OpenCV python code was modified to perform this task with the lower threshold set as 95 and the higher threshold as 215 for this kind of imagery with a sharp edge. The locations of these two pixels define the orientation and location of each scanned image, and the resampled DN of pixels within the ROI and imagery coordinate axis are illustrated in Figure 3. The row and column number of the DN in the matrix that is larger than the background (dark color of wafer) is the criterion. We applied this approach to overcome the problem of mis-location or orientation while placing wafers for inspection.

After resampling 800 × 600 DN from the ROI, three checkpoints were employed to ensure the alignment accuracy. In this case, it is the last pixel of @ and the two dots of i in the image. Only the imagery-passed alignment checking can be submitted for further handling steps. The first arrangement of this process is to scan an image of “the marker sample”, which is an image of the qualified laser marking on the wafer. Then, the height of the microscope lens and on-site illumination remain unchanged throughout the routine, and wafers can be placed by hand or conveyor into the FOV of the microscope. The image of the inspected wafer was then compared to the marker sample by pixel overlapping, and the cause of the defect was identified in near-real time if it was not a qualified product.

3. Results

Four laser configurations were applied during the testing procedure: laser speed, power, frequency, and path. Four examples from the various settings are shown in Figure 4. This indicates that both the laser speed and power are the primary controls on the marking results. The effects of various frequencies and paths cannot be distinguished visually.

For machine examination purposes, all the testing pictures were transformed into grayscale, and then the image portion that contained characters was obtained using the edge detection technique. The pixel value within the selected portion was then compared to the standard product at the exact location to obtain the brightness difference and then this difference was contoured to create charts, as shown in Figure 5.

The testing dataset contained 100 pictures of laser marking on wafers with various settings, 70% of which were randomly chosen for model training, and the remaining 30% served as model validation. The predicted and observed brightness values and regression results are listed in Figure 6, where the R² value of 0.94 denotes a very high agreement of this model prediction outcome and suggests that 25% difference values are the minimum acceptable product threshold. From the three samples shown in Figure 7, the clearest image has the lowest pixel value differencing index of 15 and the two products with such index values greater than 25 are ambiguous then been identified as defects. The ink density shown in Figure 6 also affects the pixel value differencing index, when the actual ink density exceeded than 25, the disagreement between the observed and predicted values is increased then jeopardizes the quality of engraving characters. Therefore, the pixel value differencing index of 25 is a reasonable threshold for defect detection at this configuration.

By examining the detailed distribution of pixel value differencing, as shown in Figure 8, a pattern of geometry variation can be observed. With this corresponding match for each type of defect, error source tracing becomes possible.

A shaking table for simulating the laser marking procedure under the effect of an earthquake and its products is shown in Figure 9. The shaking intensity, frequency, and many other factors can lead to various outcomes. The pixel grid overlap ratio method was applied to represent the degree of alteration, as shown in Figure 10. The overlap ratio is the percentage of the detected character pixel grid in the test image relative to that in standard product images. For perfect matching, 100% produces a lower blur ratio in the marking. When the laser power or frequency was increased, the width of the marking increased, and thus, the overlap ratio declined.

Four types of defective patterns acquired from 100 marking samples were introduced to generate 409 synthetic images of laser marking on wafers with various external causes. Because the degree of influence was randomly assigned, the ground truth information was validated by human inspection. There were 104 cases of improper alignment of marks, 97 cases of missing marks, 103 cases of excessive or insufficient marking depth, and 105 cases of blurred or illegible marks to validate the accuracy of the error tracing analysis. The external causes of laser marking defects could be unstable laser power, a dirty laser head, platform shaking, and voltage fluctuation of the electrical power. Unstable laser power causes excessive or insufficient marking depth, dirty laser heads induce missing marks, platform shaking causes improper alignment of marks, and voltage fluctuation of the electrical power creates blurred or illegible marks. Another 591 images gathered from the qualified product and reoriented randomly are blended to the system validation schema. A total of 1000 synthetic images were inspected by the proposed method; with 25 as the pixel value differencing index, 579 images passed the examination standard, and 421 images had been ruled as a defect product. The confusion matrix of this classified result is shown in

Table 3. The confusion matrix of classification result.

	Product
Result	570	9
	21	400

. The accuracy of defect classification is (570 + 400)/(591 + 409) = 97.0, and the F₁ score is (2 × 570)/(2 × 570 + 21 + 9) = 0.974. This outcome reveals the high accuracy in inspecting the defect laser marking on wafers with this inexpensive hardware and near-real-time handling process.

The error causes determined by the proposed methods are listed in

Table 4. The confusion matrix of error source tracing result.

Product	Improper Alignment of Marks	Missing Marks	Excessive or Insufficient Marking Depth	Blurred or Illegible Marks
Result
Improper alignment of marks	94	1	4	9	108
Missing marks	2	96	2	1	101
Excessive or insufficient marking depth	3	0	93	3	99
Blurred or illegible marks	5	0	4	92	101
	104	97	103	105	409

; there are four kinds of defects that have been detected. Four different phenomena have been reported, such as improper alignment of marks, missing marks, excessive or insufficient marking depth, and blurred or illegible marks. The number of classified results is listed in the corresponding table. The random forest algorithms with the training model can trace the error source with an overall accuracy of 375/409 = 91.69% and kappa = [409 × 375 − 41,831]/[409² − 41,831] = 0.889. This technique also detects the defect product with an accuracy of 97% using an inexpensive hardware configuration, and the high kappa value suggests the capability of separating the error caused by the photo of the wafers.

4. Discussion

Defect detection in the laser marking of silicon wafers relies on a combination of high-resolution imaging, laser-based scattering/reflection analysis, and real-time process monitoring. These quality control measures are essential to ensure the accuracy, durability, and traceability of wafer markings while also safeguarding the functional integrity of semiconductor devices. The new barcode laser marking on the wafers cannot be detected by traditional OCR methods but can be inspected by this pixel-matching technique.

Random forest algorithms have the disadvantage of being computationally intensive: training many trees can be resource-intensive, especially with large datasets and many features; less interpretability: while they are more interpretable than some black-box models, random forests are less transparent than a single decision tree because of their ensemble nature. Because defect checking of laser marking on a wafer is a supervised classification, the defect model of one routine is only suitable for a single case. This means that when the marker character is altered, the model must be retrained. Therefore, there is no need to introduce complex algorithms that require longer processing time and high-end hardware. Nevertheless, the proposed technique offers a low-cost and high-efficiency approach to inspect the defect product of laser marking with approximately 97.0% accuracy, and offers the cause of the error with 91.6% overall accuracy. The F₁ score of the defect inspection is 0.974, which means an almost perfect accuracy and recall; the 0.889 kappa value in the error cause tracing has demonstrated almost perfect agreement in separating the error cause. This finding ensures the defect model constructed by random forest algorithms with 100 accurate imagery laser markings on wafers could offer high accuracy for defect detection via unpretentious pixel overlapping methods.