A method and system for anomaly detection and in-situ repair of self-healing epitaxial growth

Abstract

The present application provides a self-healing epitaxial growth anomaly detection and in-situ repair method and system, wherein the self-healing epitaxial growth anomaly detection and in-situ repair method comprises: multi-dimensional state data construction; defect nucleus germination discrimination; adaptive repair protocol triggering; when the discrimination result is that there is a defect nucleus, automatically matching the corresponding intervention strategy according to the defect type label, suspending the current epitaxial growth main program, and triggering a reverse repair pulse sequence; in-situ reconstruction and verification; executing the reverse repair pulse sequence, physically removing or lattice reconstructing the region with the defect nucleus, immediately performing secondary data acquisition and re-executing the defect nucleus discrimination of step S2 to verify the defect elimination state after the reverse repair pulse sequence is executed, and if the discrimination result is that there is no defect nucleus, then the epitaxial growth main program is restored.

Description

Technical Field

[0001] This invention relates to the field of semiconductor technology, and in particular to a method and system for detecting and repairing anomalies in self-healing epitaxial growth. Background Technology

[0002] In epitaxial growth processes (such as MOCVD, MBE, etc.), the quality of the crystal directly determines the electrical and optical performance of the final semiconductor device. However, in actual production, the growth environment is extremely complex and is easily affected by factors such as trace impurity contamination, temperature drift, and airflow disturbance.

[0003] Existing technologies typically employ the following two strategies: Post-production inspection and scrapping: After the growth process is completed, the quality of the wafer is inspected by means of microscopes, X-ray diffraction and other means. If serious defects caused by the expansion of defect nuclei generated during the growth process are found, the wafer is scrapped directly. This method causes a great waste of materials and time and high production costs. Passive shutdown: When abnormal parameters are detected by sensors, the growth process is stopped directly. However, interrupting growth often leads to interface states on the wafer surface. Restarting growth makes it difficult to obtain a perfect single-crystal structure, which can also easily lead to product degradation or scrap.

[0004] The main drawback of existing countermeasures is the lack of a "real-time intervention" mechanism during the growth process, and the intervention methods are crude (such as direct shutdown), which cannot effectively eliminate or repair defects in situ at the early stage of defect germination (defect core stage), resulting in irreversible defect expansion. Summary of the Invention

[0005] Based on this, the purpose of the present invention is to provide a method and system for detecting and repairing abnormalities in self-healing epitaxial growth, which can effectively solve the shortcomings of the prior art.

[0006] A method for detecting and in-situ repairing abnormalities in self-healing epitaxial growth, comprising: S1, Construction of multidimensional data; After the main program of epitaxial growth starts, in-situ monitoring image sequences, temperature field thermal distribution data and gas flow dynamic curves in the epitaxial reaction chamber are collected in real time to construct a spatiotemporally aligned multidimensional growth state dataset. S2, Defective nucleus germination discrimination; The multidimensional growth state dataset is input into a pre-trained defect kernel identification model. The defect kernel identification model fuses image features and temporal parameter features based on an attention mechanism, and outputs the discrimination result of whether there is a defect kernel on the current growth surface and its defect type label. S3, Adaptive Repair Protocol Triggered; When the judgment result indicates the presence of a defective core, the corresponding intervention strategy is automatically matched according to its defect type label, the current epitaxial growth main program is paused, and the reverse repair pulse sequence is triggered; S4. In-situ reconstruction and verification; The reverse repair pulse sequence is executed to physically remove or reconstruct the lattice in the region where the defect nucleus exists. After the reverse repair pulse sequence is completed, a second data acquisition is immediately performed and the defect nucleus discrimination in step S2 is re-executed to verify the defect elimination status. If the discrimination result is that there is no defect nucleus, the epitaxial growth main program is resumed.

[0007] Compared with existing technologies, the beneficial effects of this invention are as follows: Through the closed-loop process of S1-S4, real-time monitoring, identification, repair, and verification are performed during epitaxial growth, endowing the equipment with a "self-healing" ability similar to that of a living organism. Unlike traditional post-detection methods, this method eliminates defects at the early stage of nucleus germination (micrometer-level stage), avoiding further evolution of macroscopic defects (such as dislocations, polycrystalline structures, and cracks), and significantly improving epitaxial yield. Moreover, the entire monitoring, repair, and verification process is completed in situ within the reaction chamber, without interrupting the main growth program for chamber cleaning or manual intervention. Growth resumes immediately after the repair pulse sequence is completed, avoiding secondary contamination (particles, oxidation) and interface defects caused by traditional shutdown and chamber opening, ensuring crystal continuity and epitaxial layer interface quality. The defect nucleus identification model, based on an attention mechanism that fuses multimodal data, can output defect type labels, providing a basis for fine matching of subsequent repair strategies and avoiding unnecessary damage to the growth surface caused by "one-size-fits-all" repair.

[0008] Furthermore, the reverse repair pulse sequence includes a chemical vapor etching sub-sequence, which specifically includes: While keeping the substrate temperature constant, the precursor source supply is cut off, and halogen-based etching gas is pulsed into the epitaxial reaction chamber for a duration of 0.5 to 3 seconds. The defect core region is converted into gaseous byproducts and discharged from the epitaxial reaction chamber by chemical reaction.

[0009] Furthermore, the reverse repair pulse sequence also includes a hydrodynamic shearing sub-sequence, which is specifically used for: By controlling the rotational speed and holding time of the carrier disk, the physical impurity particles attached to the surface are peeled off by utilizing the airflow boundary layer shear force and centrifugal force generated by the sudden change in rotational speed. The base rotational speed of the carrier disk is 500 rpm.

[0010] Furthermore, the disk rotation speed and holding time of the hydrodynamic shear subsequence are dynamically adjusted in stages according to the impurity particle size output by the defect kernel identification model: When the equivalent diameter of impurity particles D ≤ 1 μm, the microparticle removal mode is executed. In the microparticle removal mode, the peak rotation speed of the carrier disk is equal to 2 to 2.5 times the reference rotation speed, and the peak rotation speed is maintained for 1 to 2 seconds. When the equivalent diameter of impurity particles is 1μm < D ≤ 5μm, the medium particle removal mode is executed. In the medium particle removal mode, the peak rotation speed of the carrier disk is equal to 2.5 times to 3.5 times the reference rotation speed, and the peak rotation speed is maintained for 2 to 4 seconds. When the equivalent diameter D of the impurity particles is greater than 5 μm, the large particle removal mode is executed. In the large particle removal mode, the peak rotation speed of the carrier disk is equal to 3.5 times to 5 times the reference rotation speed, and the peak rotation speed is maintained for 4 to 6 seconds. The equivalent diameter D of the impurity particles is calculated by the defect kernel identification model based on particle edge detection and pixel equivalent calibration in the in-situ monitoring image sequence.

[0011] Furthermore, the disk rotation speed and holding time of the hydrodynamic shear subsequence are dynamically adjusted in stages according to the impurity particle adhesion level output by the defect kernel identification model: When the critical load Lc between the impurity particles and the substrate surface is less than 10mN, a low adhesion removal mode is executed. In the low adhesion removal mode, the peak rotation speed of the carrier disk is equal to 1.5 to 2 times the reference rotation speed, and the peak rotation speed is maintained for 0.5 to 2 seconds. When the critical load between the impurity particles and the substrate surface is 10mN≤Lc<100mN, the medium adhesion removal mode is executed. In the medium adhesion removal mode, the peak rotation speed of the carrier disk is equal to 2 to 3 times the reference rotation speed, and the peak rotation speed is maintained for 1.5 seconds to 4 seconds. When the critical load Lc between the impurity particles and the substrate surface is greater than or equal to 100 mN, a high adhesion removal mode is executed. In the high adhesion removal mode, the peak rotation speed of the carrier disk is 3 to 4.5 times the reference rotation speed, and the peak rotation speed is maintained for 3 to 6 seconds. The adhesion level is inferred by the defect core identification model based on the morphological characteristics of impurity particles and the airflow fluctuation response characteristics; the critical load Lc is calibrated through offline scratch testing and is used to establish the mapping relationship between the adhesion level and online characteristics.

[0012] Furthermore, the fluid dynamics shear subsequence also includes a step mode, which specifically includes: The carrier disk speed is controlled to jump from the reference speed to the peak speed, which is 1.5 to 5 times the reference speed. After the peak speed is maintained for 0.5 to 6 seconds, it gradually drops back to the reference speed at a fixed slope of 100 rpm / s to 300 rpm / s.

[0013] Furthermore, the fluid dynamics shear subsequence also includes oscillation modes, specifically: The carrier disk rotation speed is controlled to oscillate rapidly between the reference speed and the peak speed. The peak speed is 1.5 to 5 times the reference speed, the frequency is 0.5 Hz to 5 Hz, the amplitude is 50% to 100% of the difference between the peak speed and the reference speed, and the oscillation duration is 2 to 8 seconds.

[0014] Furthermore, the fluid dynamics shearing subsequence also includes a progressive mode, which specifically includes: The carrier disk speed is controlled to increase from the reference speed to the peak speed at a fixed slope of 100rpm / s-500rpm / s. The peak speed is 1.5 times to 5 times the reference speed. After the peak speed is maintained for 0.5 seconds to 6 seconds, it is restored to the reference speed at the same or different slope of 100rpm / s-500rpm / s.

[0015] Furthermore, the pre-training process of the defect kernel identification model includes: A defect kernel sample library containing different impurity types and different temperature drift degrees was constructed. Micrometer-level feature annotation was performed on the samples, a convolutional neural network was trained to extract image texture features, and a recurrent neural network was combined to analyze the temporal abnormal fluctuations of temperature and flow data.

[0016] On the other hand, the present invention also provides a system for performing the above-described method for abnormal detection and in-situ repair of self-healing epitaxial growth, comprising: Multimodal data acquisition module: integrates an in-situ imaging unit, an infrared temperature measurement unit, and a mass flow meter to acquire multidimensional status data in real time; Edge computing analysis module: used to run the defect kernel identification model, identify defect types in real time and generate repair decision instructions; Actuator control module: Used to receive instructions and precisely control the speed of gas valves and motors, and execute reverse repair pulse sequences. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the abnormal detection and in-situ repair method for self-healing epitaxial growth in Embodiment 1 of the present invention. The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0019] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Example 1 Please see Figure 1 Embodiment 1 of the present invention provides a method for abnormal detection and in-situ repair of self-healing epitaxial growth, comprising: S1, Construction of multidimensional data; After the main program of epitaxial growth starts, in-situ monitoring image sequences, temperature field thermal distribution data and gas flow dynamic curves in the epitaxial reaction chamber are collected in real time to construct a spatiotemporally aligned multidimensional growth state dataset. S2, Defective nucleus germination discrimination; The multidimensional growth state dataset is input into a pre-trained defect kernel identification model. The defect kernel identification model fuses image features and temporal parameter features based on an attention mechanism, and outputs the discrimination result of whether there is a defect kernel on the current growth surface and its defect type label. S3, Adaptive Repair Protocol Triggered; When the judgment result indicates the presence of a defective core, the corresponding intervention strategy is automatically matched according to its defect type label, the current epitaxial growth main program is paused, and the reverse repair pulse sequence is triggered; S4. In-situ reconstruction and verification; The reverse repair pulse sequence is executed to physically remove or reconstruct the lattice in the region where the defect nucleus exists. After the reverse repair pulse sequence is completed, a second data acquisition is immediately performed and the defect nucleus discrimination in step S2 is re-executed to verify the defect elimination status. If the discrimination result is that there is no defect nucleus, the epitaxial growth main program is resumed.

[0022] Understandably, the closed-loop process of S1-S4 enables real-time monitoring, identification, repair, and verification during epitaxial growth, endowing the equipment with a "self-healing" ability similar to that of a living organism. Unlike traditional post-detection methods, this approach eliminates defects at the early stage of nucleus germination (micrometer-level stage), preventing further evolution of macroscopic defects (such as dislocations, polycrystalline structures, and cracks) and significantly improving epitaxial yield. Furthermore, the entire monitoring, repair, and verification process is completed in situ within the reaction chamber, eliminating the need to interrupt the main growth program for chamber cleaning or manual intervention. Growth resumes immediately after the repair pulse sequence is completed, avoiding secondary contamination (particles, oxidation) and interface defects caused by traditional shutdown and chamber opening, thus ensuring crystal continuity and epitaxial layer interface quality. The defect nucleus identification model, based on an attention mechanism that fuses multimodal data, can output defect type labels, providing a basis for precise matching of subsequent repair strategies and avoiding unnecessary damage to the growth surface caused by "one-size-fits-all" repair.

[0023] Furthermore, the reverse repair pulse sequence includes a chemical vapor etching sub-sequence, which specifically includes: While keeping the substrate temperature constant, the precursor source supply is cut off, and halogen-based etching gas is pulsed into the epitaxial reaction chamber for a duration of 0.5 to 3 seconds. The defect core region is converted into gaseous byproducts and discharged from the epitaxial reaction chamber by chemical reaction.

[0024] Specifically, in this embodiment, the chemical vapor etching sub-sequence is used for: repairing GaN epitaxial "island" defects based on chemical vapor etching. During GaN MOCVD growth, the defect core identification model identifies "island" lattice distortions (defect cores) caused by trace impurities in a certain area of ​​the substrate. The defect core identification model determines it as a "lattice-type defect" and automatically triggers the chemical etching repair protocol. While keeping the substrate temperature constant, the TMGa source supply is suspended, while the NH3 supply is maintained to preserve the nitriding environment and prevent surface decomposition. A small amount of Cl2 gas is pulsed in for 1 second. The Cl2 reacts selectively with the GaN protrusions on the surface to generate gaseous GaCl3, which "flattens" the defect nucleus. Purge the epitaxial reaction chamber for 3 seconds to remove byproducts; After the defect core identification model re-acquired images to confirm the surface was flat, the TMGa source supply was restored to continue growth.

[0025] Understandably, for lattice defects (such as dislocations, stacking faults, and impurity precipitation), the defect core region can be selectively removed within 0.5-3 seconds by pulsed injection of halogen-based etching gas, achieving "self-healing" of defects. Unlike passively relying on growth condition optimization, this approach actively intervenes in the defect evolution process. The etching process is carried out while keeping the substrate temperature constant, cutting off the precursor source but maintaining thermal field stability. After repair, the precursor source supply can be restored immediately to continue growth, avoiding thermal stress and interface abrupt changes caused by temperature cycling. This subsequence is specifically designed for chemical / lattice defects.

[0026] Furthermore, the pre-training process of the defect kernel identification model includes: A defect kernel sample library containing different impurity types and different temperature drift degrees was constructed. Micrometer-level feature annotation was performed on the samples, a convolutional neural network was trained to extract image texture features, and a recurrent neural network was combined to analyze the temporal abnormal fluctuations of temperature and flow data.

[0027] Example 2 Embodiment 2 of the present invention also provides a system for performing the above-described method for abnormal detection and in-situ repair of self-healing epitaxial growth, comprising: Multimodal data acquisition module: integrates an in-situ imaging unit, an infrared temperature measurement unit, and a mass flow meter to acquire multidimensional status data in real time; Edge computing analysis module: used to run the defect kernel identification model, identify defect types in real time and generate repair decision instructions; Actuator control module: Used to receive instructions and precisely control the speed of gas valves and motors, and execute reverse repair pulse sequences.

[0028] Example 3 Embodiment 3 of the present invention provides a method for abnormal detection and in-situ repair of self-healing epitaxial growth, which differs from Embodiment 1 in that: Furthermore, the reverse repair pulse sequence also includes a hydrodynamic shearing sub-sequence, which is specifically used for: By controlling the rotational speed and holding time of the carrier disk, the physical impurity particles attached to the surface are peeled off by utilizing the airflow boundary layer shear force and centrifugal force generated by the sudden change in rotational speed. The base rotational speed of the carrier disk is 500 rpm.

[0029] Specifically, in this embodiment, the peak rotational speed is 1.5 to 5 times the base rotational speed, and the peak rotational speed is maintained for 0.5 to 6 seconds.

[0030] Specifically, in this embodiment, the fluid dynamics shearing sub-sequence is as follows: SiC epitaxial particle contamination removal based on fluid dynamics shearing. During SiC epitaxial growth, the defect core identification model, combined with airflow fluctuation data, identifies potential "particle-type" defect cores at the edges. The system determines these as "physical defects" and triggers a variable speed repair protocol. Maintain a constant growth temperature and cut off the source of energy supply; The rotational speed of the carrier disk is instantly increased from the base speed of 500 rpm to the peak speed of 1500 rpm, maintained for 2 seconds, and then the base speed is restored. Secondary data acquisition is started. If the defect core identification model determines that the defect elimination verification is passed, the source supply is restored to continue growth. During this process, high-speed rotation changes the boundary layer thickness, using enhanced shear force to fling impurity particles off the surface.

[0031] Understandably, for physical defects (particles, dust, impurity particles), the airflow boundary layer shear force and centrifugal force generated by sudden changes in rotation speed can peel the particles off the surface within 0.5-6 seconds without introducing additional etching gas, thus avoiding chemical interference with the growth environment. The entire process is achieved solely through disk rotation speed control. Shear pulses are performed while the source supply is paused, without introducing additional etching gas. After the shear pulse ends, secondary data acquisition is initiated. If the defect core identification model verifies that the defect elimination has been successful, the original rotation speed can be immediately resumed to continue growth. The process interruption time is extremely short (millisecond-level rotation speed response). This subsequence is specifically designed for physical defects, complementing chemical etching and avoiding over-etching of the intact lattice region.

[0032] Example 4 Embodiment 4 of the present invention provides a method for abnormal detection and in-situ repair of self-healing epitaxial growth, which differs from Embodiment 3 in that: Furthermore, the disk rotation speed and holding time of the hydrodynamic shear subsequence are dynamically adjusted in stages according to the impurity particle size output by the defect kernel identification model: When the equivalent diameter of impurity particles D ≤ 1 μm, the microparticle removal mode is executed. In the microparticle removal mode, the peak rotation speed of the carrier disk is equal to 2 to 2.5 times the reference rotation speed, and the peak rotation speed is maintained for 1 to 2 seconds. When the equivalent diameter of impurity particles is 1μm < D ≤ 5μm, the medium particle removal mode is executed. In the medium particle removal mode, the peak rotation speed of the carrier disk is equal to 2.5 times to 3.5 times the reference rotation speed, and the peak rotation speed is maintained for 2 to 4 seconds. When the equivalent diameter D of the impurity particles is greater than 5 μm, the large particle removal mode is executed. In the large particle removal mode, the peak rotation speed of the carrier disk is equal to 3.5 times to 5 times the reference rotation speed, and the peak rotation speed is maintained for 4 to 6 seconds. The equivalent diameter D of the impurity particles is calculated by the defect kernel identification model based on particle edge detection and pixel equivalent calibration in the in-situ monitoring image sequence.

[0033] Specifically, in this embodiment, the disk rotation speed and holding time of the hydrodynamic shear subsequence are dynamically adjusted in stages according to the impurity particle size output by the defect kernel identification model. During the GaN-on-Si epitaxial growth process, the multimodal data acquisition module acquires in-situ monitoring image sequences, gas flow dynamic curves, and temperature field thermodynamic distribution in real time. After analyzing the above multimodal data, the defect core identification model performs edge detection and pixel equivalent calibration on the detected particles, and outputs the following size classification results: Microparticle P-01 with an equivalent diameter of 0.6 μm, medium-sized particles P-02 with an equivalent diameter of 2.3 μm, large particles P-03 with an equivalent diameter of 7.2 μm, medium-sized particles P-04 with an equivalent diameter of 1.2 μm, and microparticle P-05 with an equivalent diameter of 0.8 μm; The defect kernel identification model determined it to be a mixed-size physical defect cluster, triggering a size-based hydrodynamic shear repair protocol: Keep the substrate temperature constant at 1050℃, keep the reaction chamber pressure stable, and cut off the supply of trimethylgallium and ammonia precursors. First, the micro-particle removal mode is executed, which controls the disk speed to instantly increase from 500 rpm to the peak speed of 1100 rpm, and maintains the peak speed for 1.5 seconds; Then, in the particle removal mode, the disk speed is instantly increased from 1100 rpm to the peak speed of 1500 rpm, and the peak speed is maintained for 3 seconds; Finally, the large particle removal mode is executed, controlling the disk speed to instantly increase from 1500rpm to the peak speed of 2200rpm, and the peak speed is maintained for 5 seconds; After the large particle removal mode is completed, secondary data acquisition is immediately initiated. The system determines that the defect elimination verification is successful, gradually restores the precursor source supply, and the carrier disk speed drops back to the reference speed of 500 rpm to continue epitaxial growth.

[0034] Example 5 Embodiment 5 of the present invention provides a method for abnormal detection and in-situ repair of self-healing epitaxial growth, which differs from Embodiment 3 in that: Furthermore, the disk rotation speed and holding time of the hydrodynamic shear subsequence are dynamically adjusted in stages according to the impurity particle adhesion level output by the defect kernel identification model: When the critical load Lc between the impurity particles and the substrate surface is less than 10mN, a low adhesion removal mode is executed. In the low adhesion removal mode, the peak rotation speed of the carrier disk is equal to 1.5 to 2 times the reference rotation speed, and the peak rotation speed is maintained for 0.5 to 2 seconds. When the critical load between the impurity particles and the substrate surface is 10mN≤Lc<100mN, the medium adhesion removal mode is executed. In the medium adhesion removal mode, the peak rotation speed of the carrier disk is equal to 2 to 3 times the reference rotation speed, and the peak rotation speed is maintained for 1.5 seconds to 4 seconds. When the critical load Lc between the impurity particles and the substrate surface is greater than or equal to 100 mN, a high adhesion removal mode is executed. In the high adhesion removal mode, the peak rotation speed of the carrier disk is 3 to 4.5 times the reference rotation speed, and the peak rotation speed is maintained for 3 to 6 seconds. The adhesion level is inferred by the defect core identification model based on the morphological characteristics of impurity particles and the airflow fluctuation response characteristics; the critical load Lc is calibrated through offline scratch testing and is used to establish the mapping relationship between the adhesion level and online characteristics.

[0035] Specifically, in this embodiment, the disk rotation speed and maintenance time of the hydrodynamic shear subsequence are dynamically adjusted in stages according to the impurity particle adhesion level output by the defect kernel identification model. During SiC epitaxial growth, the defect core identification model detected multiple carbon particles (D≈2-3μm). Based on particle morphology characteristics (rounded edges, small contact angle) and airflow fluctuation response characteristics (no significant displacement under pulsed carrier gas disturbance), the model comprehensively inferred the particle adhesion level. The classification results are as follows: C-01 is a medium-adhesion particle with an equivalent diameter of 2.1 μm, rounded edges, a contact angle of 25°, and a slight vibration without displacement in response to airflow. C-02 particles with medium adhesion, an equivalent diameter of 2.8 μm, clear edges, a contact angle of 35°, and no obvious airflow response; C-03 is a low-adhesion particle with an equivalent diameter of 1.5 μm, blurred edges, a contact angle of 15°, and a significant airflow response. The defect kernel identification model determined it to be a physical defect with mixed adhesion levels, triggering a hydrodynamic shear repair protocol based on adhesion grading: Maintain the substrate temperature at 1600℃, keep the reaction chamber pressure stable, and cut off the supply of silane and propane precursor sources; First, execute the low adhesion removal mode, control the carrier plate speed to instantly increase from 500 rpm to the peak speed of 900 rpm, and maintain the peak speed for 1 second; In the adhesion removal mode, the carrier plate speed is instantly increased from 900 rpm to a peak speed of 1400 rpm and maintained for 2.5 seconds. After the adhesion removal mode is completed, secondary data acquisition is immediately initiated. The system determines that the defect elimination verification is successful, gradually restores the precursor source supply, and the carrier speed drops back to the reference speed of 500 rpm to continue epitaxial growth.

[0036] Example 6 Embodiment 6 of the present invention provides a method for abnormal detection and in-situ repair of self-healing epitaxial growth, which differs from Embodiment 3 in that: Furthermore, the fluid dynamics shear subsequence also includes a step mode, which specifically includes: The carrier disk speed is controlled to jump from the reference speed to the peak speed, which is 1.5 to 5 times the reference speed. After the peak speed is maintained for 0.5 to 6 seconds, it gradually drops back to the reference speed at a fixed slope of 100 rpm / s to 300 rpm / s.

[0037] Specifically, in this embodiment, the step mode is as follows: During the SiC epitaxial growth process, the defect core identification model detected a single Fe metal particle (D≈8μm), with an adhesion level of high adhesion (critical load Lc≈120 mN). The defect core identification model was determined to be suitable for the step mode. Maintain the substrate temperature at 1600℃, keep the reaction chamber pressure stable, and cut off the precursor source supply; The step mode is executed, controlling the disk speed to climb from 500 rpm to the peak speed of 2200 rpm. After the peak speed is maintained for 4 seconds, it gradually drops back to 500 rpm at a fixed slope of 200 rpm / s. After the step mode is completed, secondary data acquisition is immediately initiated. The system determines that the defect elimination verification has passed, and the precursor source supply is gradually restored to continue epitaxial growth.

[0038] Example 7 Embodiment 7 of the present invention provides a method for abnormal detection and in-situ repair of self-healing epitaxial growth, which differs from Embodiment 3 in that: Furthermore, the fluid dynamics shear subsequence also includes oscillation modes, specifically: The carrier disk rotation speed is controlled to oscillate rapidly between the reference speed and the peak speed. The peak speed is 1.5 to 5 times the reference speed, the frequency is 0.5 Hz to 5 Hz, the amplitude is 50% to 100% of the difference between the peak speed and the reference speed, and the oscillation duration is 2 to 8 seconds.

[0039] Specifically, in this embodiment, the oscillation mode is as follows: During the GaN epitaxial growth process, the defect core identification model detected multiple carbon particles (D≈2-3μm). Morphology analysis showed that the particle edges were rounded and the contact angle was small, indicating that the adhesion was moderate. Considering the fragile nature of carbon particles, the defect core identification model selected an oscillation mode to avoid secondary contamination caused by particle breakage due to high-intensity single impact. Keep the substrate temperature constant at 1050℃, keep the reaction chamber pressure stable, and cut off the precursor source supply; The oscillation mode was executed, controlling the disk speed to climb from 500 rpm to a peak speed of 1400 rpm, with a frequency of 3 Hz, an amplitude of 900 rpm, and an oscillation duration of 4 seconds; After the oscillation mode is completed, secondary data acquisition is immediately initiated. The system determines that the defect elimination verification has passed, and the precursor source supply is gradually restored to continue epitaxial growth.

[0040] Example 8 Embodiment 8 of the present invention provides a method for abnormal detection and in-situ repair of self-healing epitaxial growth, which differs from Embodiment 3 in that: Furthermore, the fluid dynamics shearing subsequence also includes a progressive mode, which specifically includes: The carrier disk speed is controlled to increase from the reference speed to the peak speed at a fixed slope of 100rpm / s-500rpm / s. The peak speed is 1.5 times to 5 times the reference speed. After the peak speed is maintained for 0.5 seconds to 6 seconds, it is restored to the reference speed at the same or different slope of 100rpm / s-500rpm / s.

[0041] Specifically, in this embodiment, the progressive mode is as follows: During the 200mm diameter GaN-on-Si epitaxial growth process, the defect nucleus identification model detected multiple medium-sized particles (D≈3-5μm). Since the large-size substrate is sensitive to mechanical stress, it is necessary to avoid warping or thermal stress caused by sudden changes in rotation speed. Therefore, the defect nucleus identification model selects a progressive mode. Keep the substrate temperature constant, keep the reaction chamber pressure stable, and cut off the precursor source supply; The progressive mode is executed, controlling the disk speed to linearly climb from 500 rpm to the peak speed of 1800 rpm at a fixed slope of 200 rpm / s. After the peak speed is maintained for 3 seconds, it is restored to the reference speed of 500 rpm at the same slope of 200 rpm / s. After the incremental mode is completed, a second data acquisition is immediately initiated. The system determines that the defect elimination verification has passed, and the precursor source supply is gradually restored to continue epitaxial growth.

[0042] In summary, the self-healing epitaxial growth anomaly detection and in-situ repair method and system described in the above embodiments of the present invention, through a closed-loop process of S1-S4, monitors, identifies, repairs, and verifies in real time during the epitaxial growth process, endowing the equipment with a "self-healing" ability similar to that of a living organism. Unlike traditional post-detection methods, this method eliminates defects at the early stage of defect nucleus germination (micrometer-level stage), avoiding further evolution of macroscopic defects (such as dislocations, polycrystalline structures, and cracks), and significantly improving epitaxial yield. Moreover, the entire monitoring, repair, and verification process is completed in-situ within the reaction chamber, without interrupting the main growth program for chamber cleaning or manual intervention. Growth resumes immediately after the repair pulse sequence is completed, avoiding secondary contamination (particles, oxidation) and interface defects caused by traditional shutdown and chamber opening, ensuring crystal continuity and epitaxial layer interface quality. The defect nucleus identification model, based on an attention mechanism and fusion of multimodal data, can output defect type labels, providing a basis for fine matching of subsequent repair strategies and avoiding unnecessary damage to the growth surface caused by "one-size-fits-all" repair.

[0043] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0044] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for detecting and in-situ repairing abnormalities in self-healing epitaxial growth, characterized in that, include: S1, Construction of multidimensional data; After the main program of epitaxial growth starts, in-situ monitoring image sequences, temperature field thermal distribution data and gas flow dynamic curves in the epitaxial reaction chamber are collected in real time to construct a spatiotemporally aligned multidimensional growth state dataset. S2, Defective nucleus germination discrimination; The multidimensional growth state dataset is input into a pre-trained defect kernel identification model. The defect kernel identification model fuses image features and temporal parameter features based on an attention mechanism, and outputs the discrimination result of whether there is a defect kernel on the current growth surface and its defect type label. S3, Adaptive Repair Protocol Triggered; When the judgment result indicates the presence of a defective core, the corresponding intervention strategy is automatically matched according to its defect type label, the current epitaxial growth main program is paused, and the reverse repair pulse sequence is triggered; S4. In-situ reconstruction and verification; The reverse repair pulse sequence is executed to physically remove or reconstruct the lattice of the region with defective nuclei. After the reverse repair pulse sequence is completed, a second data acquisition is immediately performed and the defect nuclei discrimination in step S2 is re-executed to verify the defect elimination status. If the discrimination result is that there are no defective nuclei, the epitaxial growth main program is restored. The reverse repair pulse sequence includes a chemical vapor etching sub-sequence, which specifically includes: While keeping the substrate temperature constant, the precursor source supply is cut off, and halogen-based etching gas is pulsed into the epitaxial reaction chamber for a duration of 0.5 to 3 seconds. The defect core region is converted into gaseous byproducts and discharged from the epitaxial reaction chamber by chemical reaction.

2. The method for abnormal detection and in-situ repair of self-healing epitaxial growth according to claim 1, characterized in that, The reverse repair pulse sequence further includes a hydrodynamic shearing sub-sequence, which is specifically used for: By controlling the rotational speed and holding time of the carrier disk, the physical impurity particles attached to the surface are peeled off by utilizing the airflow boundary layer shear force and centrifugal force generated by the sudden change in rotational speed. The base rotational speed of the carrier disk is 500 rpm.

3. The method for abnormal detection and in-situ repair of self-healing epitaxial growth according to claim 2, characterized in that, The disk rotation speed and holding time of the hydrodynamic shearing subsequence are dynamically adjusted in stages according to the impurity particle size output by the defect kernel identification model: When the equivalent diameter of impurity particles D ≤ 1 μm, the microparticle removal mode is executed. In the microparticle removal mode, the peak rotation speed of the carrier disk is equal to 2 to 2.5 times the reference rotation speed, and the peak rotation speed is maintained for 1 to 2 seconds. When the equivalent diameter of impurity particles is 1μm < D ≤ 5μm, the medium particle removal mode is executed. In the medium particle removal mode, the peak rotation speed of the carrier disk is equal to 2.5 times to 3.5 times the reference rotation speed, and the peak rotation speed is maintained for 2 to 4 seconds. When the equivalent diameter D of the impurity particles is greater than 5 μm, the large particle removal mode is executed. In the large particle removal mode, the peak rotation speed of the carrier disk is equal to 3.5 times to 5 times the reference rotation speed, and the peak rotation speed is maintained for 4 to 6 seconds. The equivalent diameter D of the impurity particles is calculated by the defect kernel identification model based on particle edge detection and pixel equivalent calibration in the in-situ monitoring image sequence.

4. The method for abnormal detection and in-situ repair of self-healing epitaxial growth according to claim 2, characterized in that, The disk rotation speed and holding time of the hydrodynamic shear subsequence are dynamically adjusted in stages according to the impurity particle adhesion level output by the defect kernel identification model: When the critical load Lc between the impurity particles and the substrate surface is less than 10mN, a low adhesion removal mode is executed. In the low adhesion removal mode, the peak rotation speed of the carrier disk is equal to 1.5 to 2 times the reference rotation speed, and the peak rotation speed is maintained for 0.5 to 2 seconds. When the critical load between the impurity particles and the substrate surface is 10mN≤Lc<100mN, the medium adhesion removal mode is executed. In the medium adhesion removal mode, the peak rotation speed of the carrier disk is equal to 2 to 3 times the reference rotation speed, and the peak rotation speed is maintained for 1.5 seconds to 4 seconds. When the critical load Lc between the impurity particles and the substrate surface is greater than or equal to 100 mN, a high adhesion removal mode is executed. In the high adhesion removal mode, the peak rotation speed of the carrier disk is 3 to 4.5 times the reference rotation speed, and the peak rotation speed is maintained for 3 to 6 seconds. The adhesion level is inferred by the defect core identification model based on the morphological characteristics of impurity particles and the airflow fluctuation response characteristics; the critical load Lc is calibrated through offline scratch testing and is used to establish the mapping relationship between the adhesion level and online characteristics.

5. The method for abnormal detection and in-situ repair of self-healing epitaxial growth according to claim 2, characterized in that, The fluid dynamics shear subsequence also includes a step mode, which specifically includes: The carrier disk speed is controlled to jump from the reference speed to the peak speed, which is 1.5 to 5 times the reference speed. After the peak speed is maintained for 0.5 to 6 seconds, it gradually drops back to the reference speed at a fixed slope of 100 rpm / s to 300 rpm / s.

6. The method for abnormal detection and in-situ repair of self-healing epitaxial growth according to claim 2, characterized in that, The fluid dynamics shearing subsequence also includes an oscillation mode, which specifically includes: The carrier disk rotation speed is controlled to oscillate rapidly between the reference speed and the peak speed. The peak speed is 1.5 to 5 times the reference speed, the frequency is 0.5 Hz to 5 Hz, the amplitude is 50% to 100% of the difference between the peak speed and the reference speed, and the oscillation duration is 2 to 8 seconds.

7. The method for abnormal detection and in-situ repair of self-healing epitaxial growth according to claim 2, characterized in that, The fluid dynamics shearing subsequence also includes a progressive mode, which specifically includes: The carrier disk speed is controlled to increase from the reference speed to the peak speed at a fixed slope of 100rpm / s-500rpm / s. The peak speed is 1.5 times to 5 times the reference speed. After the peak speed is maintained for 0.5 seconds to 6 seconds, it is restored to the reference speed at the same or different slope of 100rpm / s-500rpm / s.

8. The method for abnormal detection and in-situ repair of self-healing epitaxial growth according to claim 1, characterized in that, The pre-training process of the defect kernel identification model includes: A defect kernel sample library containing different impurity types and different temperature drift degrees was constructed. Micrometer-level feature annotation was performed on the samples, a convolutional neural network was trained to extract image texture features, and a recurrent neural network was combined to analyze the temporal abnormal fluctuations of temperature and flow data.

9. A system for performing the method for abnormal detection and in-situ repair of self-healing epitaxial growth according to any one of claims 1-8, characterized in that, include: Multimodal data acquisition module: integrates an in-situ imaging unit, an infrared temperature measurement unit, and a mass flow meter to acquire multidimensional status data in real time; Edge computing analysis module: used to run the defect kernel identification model, identify defect types in real time and generate repair decision instructions; Actuator control module: Used to receive instructions and precisely control the speed of gas valves and motors, and execute reverse repair pulse sequences.

Images