MOCVD (Metal Organic Chemical Vapor Deposition) film thickness detection method and system based on multi-modal deep learning

Abstract

The invention is applicable to the technical field of semiconductor manufacturing, and provides an MOCVD (Metal Organic Chemical Vapor Deposition) film thickness detection method and system based on multi-mode deep learning, and the method comprises the steps: collecting a film image through a high-resolution industrial camera and a tunable coaxial light source, dividing a grid region, extracting a local RGB (Red, Green, Blue) value, and inhibiting metal reflection through a polarization filter; the RGB-thickness nonlinear mapping relation is dynamically calibrated based on the 3D LUT technology, and light source aging and ambient light interference are eliminated; a multi-task deep learning model is adopted to synchronously execute thickness regression and defect classification: an improved U-Net + CBAM model realizes pixel-level thickness prediction, and a ResNet-50 + LSTM model analyzes thickness time sequence fluctuation and detects process abnormality; calculating a wafer thickness distribution standard deviation if sigma < gt >; if the particle size is 0.3 nm, the flow velocity ratio of the III-group precursor to the V-group precursor is dynamically adjusted through a PID controller, and closed-loop optimization is achieved. The method has sub-nanoscale detection precision, high industrial adaptability and efficient real-time performance, single-image processing is smaller than or equal to 100 ms, the yield of the MOCVD process can be remarkably improved, and the risk of batch scrapping is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of semiconductor manufacturing technology, and in particular relates to a MOCVD film thickness detection method and system based on multimodal deep learning. Background Art

[0002] MOCVD technology, the full name of which is Metal Organic Chemical Vapor Deposition, is a technology for preparing micro-nano-scale semiconductor thin films. The principle of the MOCVD method is to use a carrier gas to separate the Group III source and the Group V source into two paths and introduce them into the cavity respectively, where gas phase chemical reactions and surface chemical reactions occur on the heated epitaxial wafer to grow a homogeneous epitaxial film on the epitaxial wafer. The high quality, high uniformity and high controllability of MOCVD epitaxial growth make it widely used in the production and manufacturing of semiconductor devices. Due to its flexibility, low cost, high efficiency and high productivity, MOCVD technology has become the most widely used technology in the semiconductor industry for producing high-quality semiconductor devices using Group III and Group V source compounds.

[0003] In the MOCVD process, film thickness uniformity is a key indicator affecting device performance. Traditional detection methods have the following drawbacks: contact measurement (such as ellipsometry) is slow and cannot provide real-time feedback; image processing methods based on single-point color mapping have difficulty handling nonlinear color distribution (such as gradient color ring interference); a high-precision, non-contact detection solution is urgently needed. Summary of the Invention

[0004] The purpose of the present invention is to provide a MOCVD film thickness detection method based on multimodal deep learning, aiming to solve the technical problems existing in the prior art identified in the background technology.

[0005] The present invention is implemented as follows: a MOCVD film thickness detection system based on multimodal deep learning includes the following collaborative modules:

[0006] Image acquisition module: High-resolution industrial camera (resolution ≥ 4K, frame rate ≥ 120fps), equipped with a polarization filter to suppress reflections from metal surfaces; tunable coaxial light source (wavelength range 400-700nm) to ensure clarity of interference patterns;

[0007] Dynamic color calibration module: Based on online learning 3D LUT (lookup table) technology, it eliminates device color difference and ambient light interference by fitting the nonlinear RGB-thickness mapping relationship;

[0008] Multi-task deep learning model: Branch 1 (thickness regression network): uses an improved U-Net architecture and embeds a channel-spatial dual attention mechanism (CBAM) to achieve pixel-level thickness prediction; Branch 2 (defect classification network): Based on the ResNet-50 transfer learning model, it combines a temporal LSTM module to identify process anomalies;

[0009] Process feedback module:

[0010] Input the thickness distribution standard deviation (σ) into the MOCVD gas flow PID controller in real time to facilitate dynamic adjustment of the III / V precursor ratio;

[0011] The main steps of core multi-task deep learning model training are as follows:

[0012] Step S1: sampling the prepared thin film material and formulating a thickness conversion table.

[0013] Step S2: Annotate each grid of each wafer with a thickness value (generated according to the color conversion table), and mark unqualified wafers (thickness difference > 0.3nm) and "BL" abnormal areas.

[0014] Step S3: Construct a U-Net+attention mechanism (CBAM) model, train the input data as the film image, train the output data as the thickness value of each pixel, and use the U-Net model to implement the thickness prediction task.

[0015] Step S4: Construct a ResNet-50+LSTM (analyze fluctuations in thickness series) model, with the training input data being the thickness distribution of a single thin film wafer and the output data being pass / fail.

[0016] Step S5: Convert the model to TensorRT format and deploy it on an embedded device (such as NVIDIA Jetson) to achieve a single image processing time of ≤100ms (meeting the production line cycle requirements)

[0017] The beneficial effects of the present invention are:

[0018] 1. Sub-nanometer detection accuracy:

[0019] By fusing multimodal data (polarized light + RGB) with dynamic 3D LUT calibration, the mean square error (MAE) of thickness prediction is ≤ 0.03nm, and the thickness difference detection sensitivity reaches 0.1nm, significantly outperforming traditional optical methods (error > 0.1nm);

[0020] The improved U-Net model's ability to analyze gradient interference patterns is improved by 50%, effectively preventing misjudgments.

[0021] 2. Comprehensiveness of defect detection:

[0022] The multi-task model simultaneously implements thickness regression and defect classification, with a recall rate of ≥99.2% for "BL" bad pixels and edge cracks and a missed detection rate of <0.05%;

[0023] The LSTM module can capture thickness timing fluctuations, provide early warning of process deviations (such as abnormal gas concentration), and reduce the risk of batch scrapping.

[0024] 3. Strong adaptability to industrial environment:

[0025] Dynamic color calibration module eliminates the effects of light source aging (color temperature drift ±500K) and camera noise (SNR≤40dB), ensuring long-term stability;

[0026] Embedded deployment (NVIDIA JetsonAGX Xavier) supports wide operating temperature range of -20°C to 60°C, meeting the harsh environment of semiconductor clean workshops. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flow chart of the main steps provided in an embodiment of the present invention;

[0028] Figure 2 An example of sampling and thickness-color mapping. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0030] like Figure 1 As shown, a MOCVD film thickness detection method based on multimodal deep learning includes:

[0031] S1, using a high-resolution industrial camera and a tunable coaxial light source to capture film images, and then dividing the area into grids to extract local RGB values;

[0032] S2, dynamic color calibration based on 3D LUT technology to eliminate color difference and ambient light interference;

[0033] S3. Using a multi-task deep learning model to simultaneously perform thickness regression and defect classification, the model includes:

[0034] Improved U-Net+CBAM thickness regression network, outputting pixel-level thickness prediction maps;

[0035] ResNet-50+LSTM defect classification network outputs qualification judgment and process abnormality warning;

[0036] S4. Feedback the standard deviation of the film thickness distribution within the wafer to the MOCVD process control system in real time, and dynamically adjust the flow rate ratio of the III-group and V-group precursors through the PID controller.

[0037] Furthermore, the MOCVD film thickness detection system based on multimodal deep learning includes:

[0038] 1. Image acquisition module

[0039] Hardware configuration:

[0040] Industrial camera: Use a high-resolution industrial camera with a resolution of ≥4096×2160 (4K) and a frame rate of ≥120fps, equipped with a polarizing filter with an extinction ratio of ≥100:1 to suppress reflections on metal surfaces (such as attached Figure 2 (reflection noise in the “BL” region shown).

[0041] Light source system: Use tunable coaxial light source with a wavelength range of 400-700nm and a wavelength accuracy of ±2nm to ensure clear imaging of interference patterns (see attached Figure 2 gradient color band as shown in the color bar on the right).

[0042] Data collection process:

[0043] Place the 75x2 inch LED wafer film image (see attached Figure 2 As shown in the figure, the device is placed in the MOCVD chamber, and the industrial camera and light source system are started.

[0044] Real-time images of the film growth process were collected, each wafer was divided into 5×5 grid areas, and the local RGB values ​​of each grid were extracted.

[0045] 2. Dynamic color calibration module

[0046] 3D LUT non-linear mapping:

[0047] Based on the attached Figure 2 The color bar on the right (marked with a thickness range of 438.90-449.50nm) and the actual sampled data (such as the marked "average = 444.6nm") are used to construct a nonlinear mapping relationship between RGB and thickness through online learning.

[0048] Data preparation: Based on the annotated thickness-color mapping table and the actual sampling data (annotated with "average = 444.6 nm"), a nonlinear mapping relationship between RGB values ​​and thickness is generated.

[0049] Online calibration:

[0050] By fitting the nonlinear relationship between the RGB three channels and thickness in real time, the camera sensor chromatic aberration (such as the response deviation of different batches of cameras) is eliminated.

[0051] Dynamically compensate for ambient light interference (such as workshop lighting fluctuations) and color temperature drift (±500K) caused by light source aging.

[0052] Output: Calibrated film image, eliminating noise (SNR≤40dB) and ensuring color data stability.

[0053] 3. Multi-task deep learning model

[0054] Model architecture and training:

[0055] Branch 1: Thickness Regression Network (U-Net+CBAM)

[0056] Input data: Calibrated gridded film image (see attached Figure 2 gridded disc).

[0057] Network improvement: embed the channel-space dual attention mechanism (CBAM) in the jump connection of U-Net to enhance the sensitivity to gradient patterns (such as the attached Figure 2 color transition area) feature extraction capability.

[0058] Output: pixel-level thickness prediction map with mean square error (MAE) ≤ 0.03nm.

[0059] Branch 2: Defect Classification Network (ResNet-50+LSTM)

[0060] Input data: thickness distribution sequence of a single wafer (see attached Figure 2 Grid thickness value marked).

[0061] Network improvement: An LSTM module is introduced based on ResNet-50 to analyze the timing fluctuations of thickness distribution and detect process drift (such as abnormal gas concentration).

[0062] Output: defect classification results (qualified / unqualified / BL abnormal), recall rate ≥ 99.2%, missed detection rate < 0.05%.

[0063] The main steps of multi-task deep learning model training are as follows:

[0064] Step S1: sampling the prepared thin film material and formulating a thickness conversion table.

[0065] Step S2: Annotate each grid of each wafer with a thickness value (generated according to the color conversion table), and mark unqualified wafers (thickness difference > 0.3nm) and "BL" abnormal areas.

[0066] Step S3: Construct a U-Net+attention mechanism (CBAM) model, train the input data as the film image, train the output data as the thickness value of each pixel, and use the U-Net model to implement the thickness prediction task.

[0067] Step S4: Construct a ResNet-50+LSTM (analyze fluctuations in thickness series) model, with the training input data being the thickness distribution of a single thin film wafer and the output data being pass / fail.

[0068] Step S5: Convert the model to TensorRT format and deploy it on an embedded device (such as NVIDIA Jetson) to achieve a single image processing time of ≤100ms (meeting the production line cycle requirements)

[0069] 4. Process feedback module

[0070] Closed-loop control logic:

[0071] Based on the pixel-level thickness prediction map, the standard deviation of the thickness distribution within the wafer is calculated. If it exceeds the threshold (>0.3nm), the process control mechanism is triggered, and the thickness distribution data is input into the PID controller in real time to dynamically adjust the flow rates of group III (such as TMGa) and group V (such as NH3) precursors.

[0072] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0073] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A MOCVD film thickness detection method based on multimodal deep learning, characterized in that: The method comprises: S1, using a high-resolution industrial camera and a tunable coaxial light source to capture film images, and then dividing the area into grids to extract local RGB values; S2, dynamic color calibration based on 3D LUT technology to eliminate color difference and ambient light interference; S3. Using a multi-task deep learning model to simultaneously perform thickness regression and defect classification, the model includes: Improved U-Net+CBAM thickness regression network, outputting pixel-level thickness prediction maps; ResNet-50+LSTM defect classification network outputs qualification judgment and process abnormality warning; S4. Feedback the standard deviation of the film thickness distribution within the wafer to the MOCVD process control system in real time, and dynamically adjust the flow rate ratio of the III-group and V-group precursors through the PID controller.

2. The method according to claim 1, characterized in that The resolution of the high-resolution industrial camera is ≥4096×2160, the frame rate is ≥120fps, and the extinction ratio of the polarizing filter is ≥100:

1.

3. The method according to claim 1, characterized in that The color difference and ambient light interference are eliminated by constructing an RGB-thickness nonlinear mapping relationship through online learning, and the influence of the color temperature drift of ±500K and the camera noise of SNR≤40dB are calibrated.

4. The method according to claim 1, wherein The multi-task deep learning model is deployed on the embedded device NVIDIA JetsonAGX Xavier, and the processing time of a single image is ≤100ms.

5. The method according to claim 1, wherein The calculation of the thickness distribution standard deviation is based on the pixel-level thickness prediction value of the wafer gridded area. When the thickness distribution standard deviation exceeds a threshold value of 0.3 nm, a PID controller is triggered to adjust the precursor ratio.

6. A MOCVD film thickness detection system based on multimodal deep learning, characterized in that: The system comprises: Image acquisition module, used to capture thin film images using a high-resolution industrial camera and a tunable coaxial light source, and to divide the area into grids to extract local RGB values; Dynamic color calibration module, used for dynamic color calibration based on 3D LUT technology to eliminate color difference and ambient light interference; Multi-task deep learning model for simultaneous thickness regression and defect classification using a multi-task deep learning model; The process feedback module is used to feed back the standard deviation of the film thickness distribution within the wafer to the MOCVD process control system in real time, and dynamically adjust the flow rate ratio of the group III and group V precursors through the PID controller.

Images