Low-stress cutting control method for silicon carbide and gallium nitride epitaxial material

Abstract

The invention discloses a low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials, which comprises the following steps: step 1, constructing a blade-wafer-adsorption roller coupling model, step 2, integrating an MEMS real-time sensing system, step 3, constructing a reinforcement learning intelligent regulation and control system, step 4, designing a hardware execution module, and step 5, carrying out low-stress cutting on the silicon carbide and gallium nitride epitaxial materials. 5, integrating a closed-loop control system; according to the method, the stress concentration area is pre-judged through multi-body dynamics modeling, the stress is monitored in real time through the MEMS, learning and dynamic parameter adjustment are reinforced, the edge breakage rate is smaller than or equal to 0.5%, the maximum local stress fluctuation is smaller than or equal to + / -10%, and the packaging reliability requirement of a third-generation semiconductor device is met.

Description

Technical Field

[0001] This invention relates to the field of semiconductor material processing technology, specifically to a low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials. Background Technology

[0002] Third-generation semiconductor materials (SiC, GaN) have become core materials in power devices, RF front-ends, and Mini / MicroLEDs due to their advantages such as high temperature resistance, high breakdown field strength, and high frequency characteristics. As wafer sizes increase to 4-8 inches, their hard-brittle properties (SiC fracture toughness ~3-4 MPa·m) become more prominent. 1 / 2 GaN fracture toughness ~1.5-2.5 MPa·m 1 / 2 The anisotropy (4H / 6H structure of SiC, c-axis orientation of GaN) poses a severe challenge to precision cutting: Traditional mechanical cutting processes rely on fixed parameters (feed speed, cooling flow rate) and cannot dynamically adapt to differences in material properties, resulting in a chipping rate generally higher than 3% (chipping length exceeding 10μm), local stress fluctuations exceeding ±20%, and a tendency to induce latent cracks in wafers; at the same time, existing vacuum adsorption rollers have insufficient rigidity (bending deformation exceeding 5μm under an 8-inch span), low diamond blade installation accuracy (radial runout exceeding 5μm), and a lack of real-time stress monitoring mechanisms, further aggravating cutting damage.

[0003] While imported DISCO cutting equipment can achieve a certain level of precision control in current mainstream solutions, it suffers from three major drawbacks: First, its high cost necessitates urgent domestic substitution (reducing equipment procurement costs by more than 30% is a key industry demand); second, its low level of intelligence, relying on manual preset process parameters, makes it unable to handle the dynamic changes of multiple batches and multiple materials being cut; and third, its stress monitoring response is lagging (response frequency < 500Hz), making it difficult to predict edge chipping risks. Existing domestic technologies mostly focus on single hardware optimizations (such as blade material improvements) or static parameter adjustments, failing to form a closed-loop control system of "perception-decision-execution-feedback," thus unable to meet the cutting requirements of third-generation semiconductor mass production lines for "low stress, high precision, and high stability."

[0004] To address the aforementioned issues, this invention proposes a low-stress cutting control technology for 4-8 inch SiC / GaN epitaxial materials: by constructing a multibody dynamics model to reveal the stress concentration mechanism, integrating MEMS real-time sensing and reinforcement learning intelligent control, and combining high-rigidity hardware for collaborative execution, an integrated closed-loop system is formed, achieving core indicators such as edge breakage rate ≤0.5% and stress fluctuation ≤±10%, while reducing cooling energy consumption and equipment costs, providing key technical support for third-generation semiconductor wafer processing. Summary of the Invention

[0005] The purpose of this invention is to provide a low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials, comprising the following steps: Step 1: Construct a blade-wafer-adsorption roller coupling model. Using simulation software with multiphysics coupling analysis capabilities, input the material parameters of silicon carbide and gallium nitride, including Young's modulus, Poisson's ratio, and fracture toughness. Considering material anisotropy, simulate the contact force distribution under different cutting depths and rotational speeds, and output stress cloud maps, vibration modes, and critical edge chipping region predictions. Derive the empirical formula for cutting force and generate an initial process window, which includes the feed rate baseline and cooling flow rate baseline. Step 2: Integration of the MEMS real-time sensing system. Multiple sensing units are embedded at the bottom of the tool holder. These sensing units are MEMS pressure sensors. High-frequency noise is filtered out using an IIR low-pass filter, and pressure peak values ​​and slope abrupt changes are extracted. The data is transmitted via an SPI interface or I... 2 The C interface is connected to the main control FPGA to synchronously mark timestamps and position codes; Step 3: Build a reinforcement learning intelligent control system and construct a neural network controller. The state inputs include MEMS pressure signals, cumulative cutting depth, cooling temperature, estimated tool wear, and historical chipping records. The MEMS pressure signals include mean and standard deviation. The action outputs include feed rate adjustment, cooling flow rate adjustment, and emergency deceleration / pause triggering. Design a reward function. First, pre-train the model in a simulation environment, and then deploy a lightweight inference model on the actual platform to support online fine-tuning and iteration. Step 4: Hardware execution module design. The adsorption roller is made of high-strength nickel-plated aluminum alloy, with honeycomb reinforcing ribs and hollow reflux channels. The surface is set with a micro-pore array, and independent vacuuming is performed in each zone. Selective adsorption is achieved by controlling multi-zone solenoid valves through PLC. The blade has an electroplated diamond layer structure, built-in micro cooling channels, and adopts a double-end clamping and elastic compensation mechanism with a quick-release buckle structure. The adsorption roller actively compensates for wafer warpage by micro-deformation, and the blade synchronously adjusts the cutting angle. Step 5: Closed-loop control system integration, interconnecting various modules via bus to construct a "perception-decision-execution-feedback" closed loop; the hardware architecture includes a high-precision XYZ platform, a high-speed air bearing spindle, and an industrial PC+FPGA+PLC; data acquisition uses the DAQmx system, which automatically calculates the edge breakage rate using a microscopic CCD camera and image recognition technology.

[0007] Preferably, the empirical formula for the cutting force in step one is: In the formula, For feed rate, For depth of cut, For rotational speed, The angle of entry.

[0008] Preferably, in step three, the reward function is: In the formula, For pressure deviation, For the number of edge defects, The over-temperature penalty term is determined by weighting coefficients. , , Balanced multi-objective optimization.

[0009] Preferably, the simulation software in step one is any one of ANSYS, COMSOL Multiphysics, or ABAQUS.

[0010] Preferably, the number of sensing units in step two is 3-5, which are evenly distributed along the circumference of the bottom of the tool holder.

[0011] Preferably, the neural network controller in step three is a DQN neural network controller or a PPO neural network controller, and the lightweight inference model is accelerated by TensorRT.

[0012] Preferably, the micropore array in step four has a consistent diameter and uniform spacing, the adsorption roller compensates for wafer warpage by independently vacuuming in partitions, and the electroplated diamond layer of the blade contains diamond particles with a diameter of 2-5 μm.

[0013] Preferably, the bus in step five is an EtherCAT bus, the PLC is a Siemens S7-1500 PLC, and the image recognition technology is OpenCV image recognition technology.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves a chipping rate of ≤0.5% (chipping length ≤5μm) and a maximum local stress fluctuation of ≤±10% by using multibody dynamics modeling to predict stress concentration areas, MEMS real-time stress monitoring, and reinforcement learning dynamic parameter tuning, thus meeting the reliability requirements of third-generation semiconductor device packaging. Compared with fixed parameter cutting, this invention improves the average feed efficiency by more than 15%, reduces cooling energy consumption by about 20%, avoids wafer contamination caused by excessive spraying, and supports multiple materials (SiC / GaN / sapphire / SOI) and multiple sizes to meet the flexible production needs of mass production lines. The hardware of this invention adopts domestic design, which reduces the cost by more than 30% compared with imported equipment, and the mean time between failures (MTBF) is ≥500h, thereby reducing the equipment investment and maintenance costs of semiconductor manufacturing companies. Attached Figure Description

[0015] Figure 1 This is an overall framework diagram of the present invention; Figure 2 This is an example of the multibody dynamics modeling process and stress cloud diagram of the present invention; Figure 3 This is the architecture diagram of the reinforcement learning intelligent control system of the present invention. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1-3 This invention provides a low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials, comprising the following steps: Step 1: Construct a blade-wafer-adsorption roller coupling model. Using simulation software with multiphysics coupling analysis capabilities, input material parameters for silicon carbide and gallium nitride, including Young's modulus, Poisson's ratio, and fracture toughness. Considering material anisotropy, simulate the contact force distribution under different cutting depths and rotational speeds, and output stress contour maps, vibration modes, and critical chipping region predictions. Derive the empirical formula for cutting force and generate an initial process window, including a feed rate baseline and a cooling flow rate baseline. The empirical formula for cutting force is: In the formula, For feed rate, For depth of cut, For rotational speed, The angle of entry; the simulation software can be any one of ANSYS, COMSOL Multiphysics, or ABAQUS; Step 2: Integration of the MEMS real-time sensing system. Multiple sensing units are embedded at the bottom of the tool holder. These sensing units are MEMS pressure sensors. High-frequency noise is filtered out using an IIR low-pass filter, and pressure peak values ​​and slope abrupt changes are extracted. The data is transmitted via an SPI interface or I... 2 The C interface connects to the main control FPGA, synchronously marking timestamps and position codes; the number of sensing units is 3-5, evenly distributed along the circumference of the bottom of the tool holder; Step 3: Build a reinforcement learning intelligent control system and construct a neural network controller. State inputs include MEMS pressure signals, cumulative cutting depth, cooling temperature, estimated tool wear, and historical chipping records. The MEMS pressure signals include mean and standard deviation. Action outputs include feed rate adjustment, cooling flow rate adjustment, and emergency deceleration / pause triggering. Design a reward function. First, pre-train the model in a simulation environment, then deploy a lightweight inference model on the actual platform, supporting online fine-tuning and iteration. The reward function is: In the formula, For pressure deviation, For the number of edge defects, The over-temperature penalty term is determined by weighting coefficients. , , Balanced multi-objective optimization; the neural network controller is a DQN neural network controller or a PPO neural network controller, and the lightweight inference model is accelerated by TensorRT; Step 4: Hardware execution module design. The adsorption roller is made of high-strength nickel-plated aluminum alloy, with honeycomb reinforcing ribs and hollow reflux channels. The surface is equipped with a micro-pore array, and independent vacuuming is performed in each zone. Selective adsorption is achieved by controlling multi-zone solenoid valves through PLC. The blade has an electroplated diamond layer structure, built-in micro-cooling channels, and adopts a double-end clamping and elastic compensation mechanism with a quick-release buckle structure. The adsorption roller actively compensates for wafer warpage through micro-deformation, and the blade synchronously fine-tunes the cutting angle. The micro-pore array has a consistent diameter and uniform spacing. The adsorption roller compensates for wafer warpage through independent vacuuming in each zone. The electroplated diamond layer of the blade contains diamond particles with a diameter of 2-5μm. Step 5: Closed-loop control system integration. Modules are interconnected via a bus to construct a "perception-decision-execution-feedback" closed loop. The hardware architecture includes a high-precision XYZ platform, a high-speed air bearing spindle, and an industrial PC + FPGA + PLC. Data acquisition uses the DAQmx system, which automatically calculates edge chipping rate using a microscopic CCD camera and image recognition technology. The bus is an EtherCAT bus, the PLC is a Siemens S7-1500 PLC, and the image recognition technology is OpenCV image recognition technology.

[0018] Example: This embodiment of the invention takes the dicing of a 6-inch SiC epitaxial wafer as an example and includes the following steps: Step 1: Multibody dynamics modeling and process window preset.

[0019] (1) A three-dimensional FEM model of blade-wafer-adsorption roller was established using ANSYS: input the structural parameters of SiC 4H (Young's modulus 450 GPa, Poisson's ratio 0.18, fracture toughness 3.5 MPa·m). 1 / 2The cutting depth range was set to 5-15μm and the spindle speed to 30k-50krpm to simulate the cutting force distribution and stress state under different parameters. (2) The coupling effect of the rotational motion and feed motion (initial velocity 2mm / s) of the diamond cutting tool (50mm diameter, 20μm thickness) was analyzed by MBD simulation, and the critical chipping region (1-2mm range at the wafer edge) and stress cloud map (e.g.) were output. Figure 2 As shown), derive the empirical formula for cutting force: , Use a temperature range of 15°-30° and a preset cooling flow rate baseline of 30 ml / min.

[0020] Step 2: Deployment and calibration of MEMS sensing system.

[0021] (1) Embed three LPS27HHW sensors along the circumferential direction at the bottom of the tool holder, with a spacing of 120° and a sampling rate of 1kHz. Connect them to the Xilinx Kintex-7 FPGA via the SPI interface. (2) The sensor was calibrated using a standard pressure source (0-100 kPa), and the error was controlled within ±0.5 hPa; an IIR low-pass filter (cutoff frequency 500 Hz) was designed to filter out spindle vibration noise and extract the average pressure value. with standard deviation ,when A collapse warning is triggered when the pressure exceeds 5 hPa.

[0022] Step 3: Training and deploying reinforcement learning models.

[0023] (1) Constructing a PPO neural network in the PyTorch framework: The state space contains the pressure mean. (0-50kPa), cooling temperature (25-40℃), tool wear value (0-0.1mm), historical chipping count; motion space set to feed rate adjustment step ±0.01-0.2mm / s, cooling flow rate adjustment step ±1-3ml / min; reward function weight. , , The architecture diagram of the reinforcement learning intelligent control system is as follows: Figure 3 As shown.

[0024] (2) Generate 1000 sets of virtual cutting scenarios (including different stress fluctuations and edge breakage conditions) pre-trained models in the simulation environment. After 500 iterations, deploy a lightweight model on the actual platform (accelerated by TensorRT, with a sampling period of 10ms) to support online fine-tuning (update the strategy once every 10 wafers are cut).

[0025] Step 4: Hardware execution module assembly and debugging.

[0026] (1) Vacuum adsorption roller: 6061 aluminum alloy with nickel plating (thickness 5μm), honeycomb reinforcing ribs (spacing 5mm, height 3mm) and hollow return channel (diameter 8mm) are processed. The bending deformation test value under 8-inch span is 1.8μm; the micropore array (diameter 50μm, spacing 2mm) is vacuumed independently in 3 zones, and the PLC controls the solenoid valve to achieve an adsorption force of 12kPa (uniformity deviation ±3%). (2) Diamond cutting tool assembly: adopts a double-end clamping mechanism, and uses a laser interferometer to calibrate the radial runout to 2.5μm and the axial parallelism to 1.8μm. It has a built-in micro cooling channel (0.5mm in diameter) connected to the cooling system to ensure that the coolant reaches the cutting edge directly.

[0027] Step 5: Closed-loop control and cut verification.

[0028] (1) Interconnect the XYZ platform (resolution 0.1μm), high-speed spindle (up to 60krpm, runout 0.8μm), and Siemens S7-1500 PLC via EtherCAT bus, with a control delay of <1ms; (2) Cut 5 batches of SiC wafers (5 wafers per batch): The first batch used fixed parameters (feed speed 2mm / s, cooling flow rate 30ml / min) with a chipping rate of 1.8%; the second to fifth batches used reinforcement learning control to dynamically adjust the feed speed (1.8-2.3mm / s) and cooling flow rate (28-32ml / min), and the final chipping rate was reduced to 0.4%, stress fluctuation was controlled within ±8%, and it ran continuously for 72 hours without failure (MTBF=520h).

[0029] Multi-material compatibility testing: Replace with a 6-inch GaN epitaxial wafer (c-axis orientation, Young's modulus 340 GPa), and adjust the FEM model parameters and the weights of the reinforcement learning reward function. The cutting edge breakage rate was increased to 0.4%, and the cutting edge breakage rate was 0.5%, verifying the adaptability of the present invention to different hard and brittle materials.

[0030] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for controlling low-stress cutting of silicon carbide and gallium nitride epitaxial materials, characterized in that: Includes the following steps: Step 1: Construct a blade-wafer-adsorption roller coupling model. Using simulation software with multiphysics coupling analysis capabilities, input the material parameters of silicon carbide and gallium nitride, including Young's modulus, Poisson's ratio, and fracture toughness. Considering material anisotropy, simulate the contact force distribution under different cutting depths and rotational speeds, and output stress cloud maps, vibration modes, and critical edge chipping region predictions. Derive the empirical formula for cutting force and generate an initial process window, which includes the feed rate baseline and cooling flow rate baseline. Step 2: Integration of the MEMS real-time sensing system. Multiple sensing units are embedded at the bottom of the tool holder. These sensing units are MEMS pressure sensors. High-frequency noise is filtered out using an IIR low-pass filter, and pressure peak values ​​and slope abrupt changes are extracted. The data is transmitted via an SPI interface or I... 2 The C interface is connected to the main control FPGA to synchronously mark timestamps and position codes; Step 3: Build a reinforcement learning intelligent control system and construct a neural network controller. The state inputs include MEMS pressure signals, cumulative cutting depth, cooling temperature, estimated tool wear, and historical chipping records. The MEMS pressure signals include mean and standard deviation. The action outputs include feed rate adjustment, cooling flow rate adjustment, and emergency deceleration / pause triggering. Design a reward function. First, pre-train the model in a simulation environment, and then deploy a lightweight inference model on the actual platform to support online fine-tuning and iteration. Step 4: Hardware execution module design. The adsorption roller is made of high-strength nickel-plated aluminum alloy, with honeycomb reinforcing ribs and hollow reflux channels. The surface is set with a micro-pore array, and independent vacuuming is performed in each zone. Selective adsorption is achieved by controlling multi-zone solenoid valves through PLC. The blade has an electroplated diamond layer structure, built-in micro cooling channels, and adopts a double-end clamping and elastic compensation mechanism with a quick-release buckle structure. The adsorption roller actively compensates for wafer warpage by micro-deformation, and the blade synchronously adjusts the cutting angle. Step 5: Closed-loop control system integration, interconnecting various modules via bus to construct a "perception-decision-execution-feedback" closed loop; the hardware architecture includes a high-precision XYZ platform, a high-speed air bearing spindle, and an industrial PC+FPGA+PLC; data acquisition adopts the DAQmx system, which automatically calculates the edge breakage rate through a microscopic CCD camera and image recognition technology.

2. The low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials according to claim 1, characterized in that: The empirical formula for cutting force mentioned in step one is: In the formula, For feed rate, For depth of cut, For rotational speed, The angle of entry.

3. The low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials according to claim 1, characterized in that: In step three, the reward function is: In the formula, For pressure deviation, For the number of edge defects, The over-temperature penalty term is determined by weighting coefficients. , , Balanced multi-objective optimization.

4. The low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials according to claim 1, characterized in that: The simulation software in step one can be any one of ANSYS, COMSOL Multiphysics, or ABAQUS.

5. The low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials according to claim 1, characterized in that: The number of sensing units in step two is 3-5, which are evenly distributed along the circumference of the bottom of the tool holder.

6. The low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials according to claim 1, characterized in that: The neural network controller in step three is a DQN neural network controller or a PPO neural network controller, and the lightweight inference model is accelerated by TensorRT.

7. The low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials according to claim 1, characterized in that: The micropore array in step four has a consistent diameter and uniform spacing. The adsorption roller compensates for wafer warpage by independently vacuuming the partitions. The electroplated diamond layer of the blade contains diamond particles with a diameter of 2-5 μm.

8. The low-stress cutting control method for silicon carbide and gallium nitride epitaxial materials according to claim 1, characterized in that: The bus used in step five is the EtherCAT bus, the PLC is a Siemens S7-1500 PLC, and the image recognition technology is OpenCV image recognition technology.

Images