In the double-side polishing of silicon carbide (SiC) substrates, wafer breakage and fragment incidents are serious challenges that equipment developers and process engineers must face. Silicon carbide has high hardness and high brittleness, while the double-side polishing process takes place inside a closed space between the upper and lower polishing plates, making it impossible to
directly observe the wafer condition in real time.
Once wafer breakage occurs, not only may the entire batch of wafers be scrapped, but the polishing plates may also be scratched and the carrier gears damaged, resulting in equipment downtime lasting several hours or even several days.
Therefore, how to achieve accurate early warning at the initial stage of wafer breakage and improve detection accuracy to more than 99% has become a key indicator of the intelligence level of double-side polishing equipment.
Why Is Fragment Detection in Double-Side Polishing So Difficult?
Fragment detection in double-side polishing machines is difficult mainly due to the following factors:
Enclosed processing space
The wafers are held inside carrier gears between the upper and lower polishing plates. The entire process is enclosed, making it difficult to apply traditional machine vision directly.
Numerous interference sources
Polishing slurry can easily contaminate optical sensor windows. The revolution and rotation of the carrier gears cause the wafer positions to change continuously. Bubbles, polishing pad debris, and other factors may also generate interference signals.
Diverse breakage modes
There may be instantaneous complete wafer breakage, or gradual fracture propagation starting from edge microcracks. The latter is more difficult to detect at an early stage.
At present, mainstream detection methods each have their own advantages and limitations. Acoustic emission sensors are sensitive to brittle fracture but are easily affected by environmental noise. Torque monitoring can reflect overall load changes, but it is not sensitive enough to the breakage of a single wafer. Optical transmission detection can provide high accuracy, but it is limited by slurry contamination and installation space.
A single-sensor solution is therefore difficult to meet both high detection rate and low false alarm rate requirements at the same time.
Multimodal AI Fusion: Listening to Sound, Monitoring Load, and Sensing Pressure
To achieve more than 99% early-warning accuracy, it is necessary to break through the limitations of single-sensor detection and move toward multimodal information fusion.
An AI-based fragment detection system usually consists of the following core modules:
Multi-source sensing layer
Complementary sensor combinations are deployed at key positions of the equipment.
For example, acoustic emission sensors can be installed near the polishing plate to capture high-frequency fracture signals. High-frequency pressure sensors can be integrated into the upper polishing plate air circuit to detect instantaneous impact. Optical fiber transmission sensors may also be arranged at the edge of the polishing plate as auxiliary verification.
Feature extraction layer
Raw data from different types of sensors need to be preprocessed and engineered into useful features.
For acoustic emission signals, time-frequency domain features need to be extracted. For torque signals, sudden slope changes and amplitude variations are key concerns.
Traditional threshold-based methods are often insufficient for complex operating conditions, while deep learning models can automatically learn feature representations of different breakage patterns.
Fusion decision layer
This is the core for achieving high accuracy.
A multimodal fusion model can be used to fuse feature vectors from different sensors at either the feature level or decision level. For example, an attention-based neural network can dynamically evaluate the confidence of different sensors under current operating conditions and assign different fusion weights.
When an abnormal peak appears in the acoustic emission signal and a slight torque jump occurs at the same time, the system can determine that there is a high probability of a wafer breakage event.
Key Technical Paths to Achieve >99% Accuracy
Construction of a high-quality dataset
The performance of an AI model depends heavily on the quality and coverage of training data.
It is necessary to systematically collect different types of breakage samples, including acoustic emission waveforms and torque curves under different material types, such as conductive and semi-insulating SiC substrates, different wafer thicknesses, and different breakage modes.
At the same time, a large amount of normal-process interference data should also be collected, such as polishing slurry flow fluctuations, carrier gear reversing impacts, bubble bursts, and other disturbances.
A balanced design of positive and negative samples is crucial for reducing the false alarm rate.
Deep optimization of time-series models
Fragment signals are essentially time-series data and often contain precursor characteristics.
Time-series models such as LSTM or Transformer can be used to capture weak precursor signals within tens of milliseconds before breakage occurs, such as low-frequency acoustic signals generated during microcrack propagation and the accumulated effects of small friction fluctuations.
This type of “advanced perception” capability is beyond the reach of traditional threshold-based methods.
Edge deployment and real-time inference
The early-warning system must complete signal acquisition, feature extraction, model inference, and decision output within milliseconds.
By deploying a trained lightweight model on an edge computing unit at the equipment side, data transmission latency can be avoided, ensuring that the machine can stop in time before wafer fragments cause further damage.
Model compression and quantization technologies can significantly improve inference speed while maintaining detection accuracy.
Continuous learning and adaptive updating
Differences in wafer physical properties between batches, aging of polishing pads, and changes in ambient temperature and humidity can all affect the statistical distribution of sensor signals.
By establishing an online incremental learning mechanism, the system can adaptively adjust its decision boundaries based on newly collected data and maintain an accuracy level of more than 99% throughout the equipment lifecycle.
Engineering Challenges and Countermeasures
Imbalanced samples
Wafer breakage is a very low-probability event. Normal samples may number in the millions, while breakage samples may only number in the hundreds.
Generative adversarial networks can be used to synthesize scarce breakage samples, or special loss functions designed for few-shot learning can be adopted to effectively reduce model bias.
Trade-off between false alarms and missed detections
The cost of missed detection is much higher than that of false alarms. However, excessive false alarms may lead to frequent machine stoppages and affect production efficiency.
Reasonable thresholds can be set according to process characteristics. A secondary confirmation mechanism can also be added after an early warning, such as reducing rotational speed for rapid re-evaluation, so as to avoid unnecessary downtime.
Multi-source data synchronization
Different sensors have very different sampling frequencies. Acoustic emission signals can reach the MHz level, while torque signals may only be at the kHz level.
The time alignment accuracy of the data directly affects the fusion result. At the hardware level, unified clock synchronization is required. At the software level, high-precision interpolation and alignment algorithms need to be designed.
Future Outlook
As SiC substrates continue to develop toward larger diameters and thinner thicknesses, the risk of breakage will further increase, and the requirements for detection systems will become even more stringent.
In the future, fragment early-warning technology for double-side polishing machines will show the following trends:
Richer sensing dimensions
In addition to existing sensors, technologies such as vibration spectrum analysis and infrared thermal imaging can be explored to capture precursor features from more physical dimensions.
More intelligent fusion models
Advanced architectures such as graph neural networks can be introduced to model spatial correlations and logical dependencies among different sensors, thereby improving discrimination capability under complex operating conditions.
Equipment self-protection and self-recovery
After an early warning is triggered, the system should not only realize rapid shutdown, but also minimize secondary damage to the polishing plates through methods such as reverse air blowing and coordinated rotational speed control.
Multi-machine collaborative intelligence
When multiple machines in a factory accumulate a sufficient amount of breakage data, a cloud-based shared early-warning model can be built to enable rapid iteration of breakage patterns and global optimization.
Conclusion
The AI-based fragment detection and early-warning system for double-side polishing machines represents a key technological transition from “post-event handling” to “pre-event prevention.”
By deeply integrating multimodal sensor fusion with deep learning models, improving early-warning accuracy to more than 99% can not only significantly reduce wafer breakage losses and equipment downtime, but also remove a key obstacle to the realization of fully automated unmanned polishing production lines.
The continuous evolution of this technology will become an important support for SiC substrate processing to move toward higher efficiency and lower cost.