Construction method and system of metal coating CCD detection model based on GAN

Abstract

The invention discloses a construction method and system of a metal coating CCD detection model based on GAN, and belongs to the technical field of metal coating detection of semiconductor electronic components. An image data processing result which is output by the constructed GAN and meets a threshold range is a good product; and the processing result of the output image data exceeding the threshold range is a missing product, so that the electronic electroplating high-end quality requirement of the semiconductor electronic element is strictly managed according to the standard. Besides, the appearance, shape, color and overall morphology characteristics of the semiconductor device after electronic electroplating are simulated and monitored by adopting a plating piece total quality threshold value judgment module, the accuracy of GAN training is improved by meeting the standard that the minimum threshold value is a plating specification value-5% and the maximum threshold value is a plating specification value + 5%, training distribution of a generator and a discriminator is rapidly achieved, and the accuracy of the GAN training is improved by adopting the standard that the minimum threshold value is the plating specification value-5% and the maximum threshold value is the plating specification value + 5%. Therefore, the optimal selection standard of the CCD detection mechanism model is obtained, and the detection effect of accurately and rapidly identifying good products or missing products is achieved.

Description

Technical Field

[0001] The present invention relates to a method and system for constructing a metal plating CCD detection model based on GAN, and belongs to the technical field of metal plating detection of semiconductor electronic components. Background Art

[0002] The metal plating deposition of semiconductor electronic components is a unique characterization technology that is sensitive to the local electrochemical-physical environmental structure of metal ions during the electronic electroplating process. It has a wide range of applications in the fields of physics, electrochemistry, and electronic electroplating high-end manufacturing. During the operation of electronic electroplating production equipment, the factors that affect the deposition of metal plating are numerous and complex under the electrochemical-physical environmental conditions of metal ions; therefore, the actual deposited metal plating will have various differences due to the interference of various influencing factors. In the mild case, the plating area is smaller than the required range or exceeds the range, or the plating area is misaligned. In the severe case, the plating area is leaked or even the product is deformed.

[0003] At present, the use of CCD detection equipment and technology provided by market suppliers can detect errors in the coating area and product deformation in real time in continuous electronic electroplating production equipment; however, it is difficult to accurately detect the lack of plating in the coating area; therefore, exploring the creation of an online real-time and accurate CCD metal coating detection mechanism method has become a very important topic. Summary of the invention

[0004] In order to improve the CCD detection accuracy of semiconductor metal coatings, the present invention provides a method and system for constructing a CCD detection model of metal coatings based on GAN. The technical solution is as follows: The present invention provides a method for constructing a metal coating CCD detection model, comprising: Step 1: Obtain the size data and coating area of ​​the semiconductor plated part to be processed; Step 2: Obtain the dimension data set of the coating area of ​​the semiconductor plated part to be processed and the metal coating information to be electroplated, and calculate the total mass of the plated part modeling m 建模总质量 , the calculation method is: m 建模总质量 = m 镀件质量 + m 镀层1 + m 镀层2 + ...... + m 镀层n Step 3: Randomly select the coating quality standard conditions, use the electronic electroplating module to perform the actual electroplating operation on the semiconductor plated part to be processed, and after the electroplating is completed, use a precision balance to measure the actual total mass of the plated part m 实测总质量 ; Step 4: Model the total mass of the plated part m 建模总质量 And the total mass of the plated parts m 实测总质量 When the two satisfy the following relationship: m 实测总质量 (100%-5%)≤ m 建模总质量 ≤ m 实测总质量 (100%+5%) Then save this set of data for training the GAN neural network; If the above relationship is not met, the coating quality standard conditions shall be based on the total mass of the plated parts measured. m 实测总质量 Make corrections and calculate the corrected coating quality standard conditions, return to step 3 to re-measure the total mass of the plated parts m 实测总质量 , and re-determine the total mass of the corrected plated model m 建模总质量 Whether the threshold range is met until the total mass of the plated part is modeled m 建模值 When the above relationship is met, the data set can be used for training the GAN neural network; Step 5: Model the total mass of the plated part m 建模总质量 And the total mass of the plated parts m 实测总质量 Import GAN neural network to model the total quality of plated parts m 建模总质量 And the total mass of the plated parts m 实测总质量 The GAN game system is trained and trained. The trained GAN neural network models the total mass of the plated parts according to the input m 建模总质量 And the total mass of the plated parts m 实测总质量 Classify electronic electroplating products into good products and defective products.

[0005] Optionally, the model operation system of the GAN neural network is composed of sensitive condition parameter modeling, and the sensitive condition parameters include: type of electroplating solution, temperature of electroplating solution, specific gravity of electroplating solution, diameter of the pipeline for transporting electroplating solution, type of electroplating power supply, current intensity, electroplating time, cathode current efficiency, surface area of ​​plating area, metal density of electroplating layer, current density and flow rate of electroplating solution.

[0006] Optionally, the semiconductor plated parts to be processed include: precision terminals, lead frames, wafer chips and precision decorations.

[0007] Optionally, the metal plating layer includes: Monomer plating Au, Ag, Ni, Sn, Cu, Pd, Rh, Pt; Alloy coatings Pd-Ni, Ni-P, W-Ni, Ag-Sn, Rh-Ru and Pt-Rh.

[0008] The present invention provides a system for constructing a metal coating CCD detection model, comprising: A memory 510, used for storing data; CCD detection module 320, used to obtain CCD morphological parameters of the semiconductor plated part to be processed; The electronic electroplating module 310 is used to randomly extract the plating quality standard conditions and perform the actual electroplating operation on the semiconductor plated part to be processed; Plated component + coating mass calculation module 330, used to calculate the total mass of plated component modeling m 建模总质量 , the calculation method is: m 建模总质量 = m 镀件质量 + m 镀层1 + m 镀层2 + ...... + m 镀层n The total mass threshold determination module of the plated part is used to model the total mass of the plated part. m 建模总质量 And the total mass of the plated parts m 实测总质量 When the two satisfy the following relationship: m 实测总质量 (100%-5%)≤ m 建模总质量 ≤ m 实测总质量 (100%+5%) Then save this set of data for training the GAN neural network; If the above relationship is not met, the coating quality standard conditions shall be based on the total mass of the plated parts measured. m 实测总质量 Make corrections and calculate the corrected coating quality standard conditions, return to step 3 to re-measure the total mass of the plated parts m 实测总质量 , and re-determine the total mass of the corrected plated model m 建模总质量 Whether the threshold range is met until the total mass of the plated part is modeled m 建模值 When the above relationship is met, the data set can be used for training the GAN neural network; GAN neural network generation module 200, based on the total mass of the plated part model m 建模总质量 And the total mass of the plated parts m 实测总质量 The GAN game system is trained and trained. The trained GAN neural network models the total mass of the plated parts according to the input m 建模总质量 And the total mass of the plated parts m 实测总质量 Classify electronic electroplating products into good products and defective products.

[0009] Optionally, the GAN neural network generation module 200 includes: A training module 210, for training the total mass of plated parts and the total mass actually measured for plated parts; Stability generation module 220, for processing the ordered arrangement and trend of the measured total mass; A coating potential energy calculation module 230, used for coating simulation calculation based on the measured total mass; The model morphology determination module 240 is used to determine the model morphology of the plated part based on the simulation calculation coating result.

[0010] The present invention provides a CCD detection method for semiconductor metal coating based on a GAN neural network, wherein the method uses a CCD detection model constructed by any of the above methods to detect the metal coating of a semiconductor plated part, comprising: Step 1: Use a precision balance to obtain the actual mass of the semiconductor plated part to be tested m 实测值 ; Step 2: Obtain CCD morphological parameters of the semiconductor plated part to be inspected; Step 3: Calculate the total mass simulation value of the semiconductor plated part based on the CCD morphology parameters M CCD , the calculation method is: M CCD = m 镀件质量 + m 镀层1 + m 镀层2 + ...... + m 镀层n m 镀件质量 = ρ v m 镀层1 = p 1 v 1 m 镀层2 = p 2 v 2 ...... m 镀层n = p n v n in, r is the metal density of the semiconductor plated part; v is the metal volume of the semiconductor plated part; r 1 is the metal density of the first coating layer of the semiconductor plated component; v 1 is the volume of the first coating layer; r 2 is the metal density of the second coating layer of the semiconductor plated component; v 2 is the volume of the second coating layer; and so on. r n is the metal density of the nth coating layer of the semiconductor coating; v n is the volume of the nth coating; Step 4: Actual quality of semiconductor coatings m 实测值 and the total mass simulation value M CCD Make a judgment: m 实测值 (100%-5%)≤ M CCD ≤ m 实测值 (100%+5%) If the above relationship is not satisfied, the semiconductor plated part is judged as a defective product; if the above relationship is satisfied, the next step of the GAN neural network judgment is carried out; Step 5: The actual quality of the semiconductor plated partsm 实测值 and the total mass simulation value M CCD Input the trained GAN neural network, and the results of the GAN neural network adversarial processing are divided into two categories: Good product: The image data processing results of the electronic electroplating product meet the morphology threshold range; Missing products: The image data processing results of electronic electroplating products do not meet the morphology threshold range.

[0011] The beneficial effects of the present invention are: The present invention improves and constructs a GAN neural network operation model through a deep learning simulation model to implement training. The image data processing results of the constructed GAN network output that meet the threshold range are "good products", and the image data processing results that exceed the threshold range are "missing products". This standard is used to strictly manage the high-end quality requirements of electronic electroplating of semiconductor electronic components.

[0012] In addition, the total quality threshold discrimination module of the plated parts is used to simulate the monitoring of the appearance, shape, color and overall morphology characteristics of semiconductor devices after electronic electroplating. Furthermore, by meeting the standard that the minimum threshold is -5% of the coating specification value and the maximum threshold is +5% of the coating specification value, the accuracy of GAN network training is improved, and the training distribution of the generator and the discriminator is quickly realized, thereby obtaining the optimization standard of the CCD detection mechanism model and achieving the detection effect of accurately and quickly identifying good or defective products.

[0013] Based on the detection results of the present invention, the optimal CCD detection mechanism for various semiconductor devices with different structures can be quickly established in actual engineering for the electroplating production conditions of electronically plated semiconductor plated parts, and then the optimal local electronic plating control method for semiconductor plated parts can be obtained. This can not only achieve the accuracy of predicting the local coating thickness of semiconductor devices and greatly reduce the error with the actual coating thickness of electroplated products, but also achieve further improvement of the electronic plating equipment, promote the improvement of the uniformity of the local coating thickness of semiconductor devices, and meet the high quality requirements of semiconductor devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0015] Figure 1 This is a simplified structural diagram of the metal coating CCD detection model based on GAN in the present invention.

[0016] Figure 2It is a flow chart of the method and system for constructing a metal coating CCD detection model based on GAN in the present invention.

[0017] Figure 3 It is a schematic diagram of a semiconductor plated test piece to be processed in the third embodiment of the present invention.

[0018] Figure 4 It is a schematic diagram of the copper plating area of ​​each unit in the semiconductor plated component experimental piece to be processed in the third embodiment of the present invention.

[0019] Figure 5 It is a schematic diagram of the silver-plated area of ​​each unit in the semiconductor plated component experimental piece to be processed in the third embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0021] Embodiment 1: This embodiment provides a method for constructing a metal coating CCD detection model. The detection model is based on a generative adversarial network (GAN). The construction process mainly includes two parts, namely, the construction of a data set and the training of a GAN network.

[0022] First, the data set is constructed based on the semiconductor plated parts to be processed. By collecting data such as the size and coating area of ​​these semiconductor plated parts, the following steps are specifically included: Step 1: Obtain the size data and coating area of ​​the semiconductor plated part to be processed; Step 2: Obtain the dimension data set of the coating area of ​​the semiconductor plated part to be processed and the metal coating information to be electroplated, and calculate the total mass of the plated part modeling m 建模总质量 , the calculation method is: m 建模总质量 = m 镀件质量 + m 镀层1 + m 镀层2 + ...... + m 镀层n (1) m 镀件质量 = ρ v m 镀层1 = p 1 v 1 m 镀层2 = p 2 v 2 ...... m 镀层n = p n v n in, r is the metal density of the semiconductor plated part; v is the metal volume of the semiconductor plated part; r 1 is the metal density of the first coating layer of the semiconductor plated component; v 1 is the volume of the first coating layer; r 2 is the metal density of the second coating layer of the semiconductor plated component; v 2 is the volume of the second coating layer; and so on. r n is the metal density of the nth coating layer of the semiconductor coating; v n is the volume of the nth coating layer.

[0023] Step 3: Randomly select the coating quality standard conditions, use the electronic plating module to perform the actual plating operation on the semiconductor plated parts to be processed, and after the plating is completed, use the ultra-precision analytical balance to measure the total mass of the plated parts. m 实测总质量 ; Step 4: Model the total mass of the plated part m 建模总质量 And the total mass of the plated parts m 实测总质量 When the two satisfy the following relationship: m 实测总质量 (100%-5%)≤ m 建模总质量 ≤ m 实测总质量 (100%+5%) (2) This set of data is saved for training the GAN neural network.

[0024] If the above relationship is not satisfied, the data set cannot be used for GAN neural network training, and the coating quality standard conditions are based on the total mass of the measured plated parts. m 实测总质量 Make corrections and calculate the corrected coating quality standard conditions, return to step 3 to re-measure the total mass of the plated parts m 实测总质量 , and re-determine the total mass of the corrected plated model m 建模总质量Whether the threshold range is met until the total mass of the plated part is modeled m 建模值 Only when the above relationship is met can this set of data be used for training the GAN neural network.

[0025] After the dataset is constructed, it is used to train the GAN neural network.

[0026] Step 5: Model the total mass of the plated part m 建模总质量 And the total mass of the plated parts m 实测总质量 Import GAN neural network to model the total quality of plated parts m 建模总质量 And the total mass of the plated parts m 实测总质量 The GAN game system is trained and trained. The trained GAN neural network will model the total mass of the input plated parts. m 建模总质量 And the total mass of the plated parts m 实测总质量 Classify electronic electroplating products into good products and defective products.

[0027] The GAN neural network model operation system is composed of sensitive condition parameter modeling, which includes the type of electroplating solution; the temperature of the electroplating solution; the specific gravity of the electroplating solution; the diameter of the pipeline for transporting the electroplating solution; the type of electroplating power supply; the current intensity; the electroplating time; the cathode current efficiency; the surface area of ​​the plating area; the metal density of the electroplating layer; the current density; and the flow rate of the electroplating solution.

[0028] This embodiment applies the Nash equilibrium principle to model the total mass data set of semiconductors and the CCD detection visual images and data using a differentiable function. D and G They represent the discriminator and generator respectively; the generator deeply learns the real data and distribution, while the discriminator correctly judges the results of deep learning operations based on the threshold standard, whether it can be used in the GAN adversarial network model to construct a precise detection method for metal coating CCD.

[0029] The basic principle of the metal coating CCD detection mechanism based on the deep learning simulation model of this embodiment is that any differentiable function can be used to represent the generator and discriminator of the GAN neural network. D and G The input is real data x and random variables z . G ( z ) G The generated data satisfy the minimum and maximum thresholds and are close to the real data distribution. Pdata Sample. If the input of the discriminator comes from the real data detected by CCD, it is marked as 1; if the input is a random modeling variable z of G ( z ), it is marked as 0. Here D The goal is to achieve a two-category discrimination of the data source, that is, the real data that can be used for modeling x distribution or non-modeled data from the generator that does not meet the minimum and maximum threshold ranges G ( z ),and G The goal is to make the non-modeled data generated by yourself fail to meet the minimum and maximum threshold ranges G ( z ),exist D Performance on D ( G ( z )) and real data that can be used for modeling x exist D Performance on D ( x ), these two processes of competing and iterative optimization make D and G The performance of D When the ability to distinguish data is improved to a certain level and the source of data cannot be correctly distinguished, the skill considers that the generator G The distribution of the modeled data has been learned.

[0030] In the CCD detection system of metal coating, given the generator G In the case of D . As with the general Sigmoid-based binary classification model training, training the discriminator D It is also the process of minimizing the cross entropy. The loss function equation (3) is: (3) In the formula, x Sampling on CCD to accurately detect data distribution P data ( x ), z is sampled from the modeled random prior distribution P z (z), E(·) represents the expected value. The actual training is different from the classification model. The training data set of the discriminator comes from the distribution of the CCD image data set. P data ( x ), marked as 1, and the modeled data distribution of the generator P g (x ), marked as 0. Given a generator G , we need to obtain the optimal solution of minimization equation (4) in continuous space.

[0031] (4) In this embodiment, for any non-zero real number m and n , and the real value y ∈【0,1】,expression - m log( y )- n log(1- y ) gets the minimum value at m / (m+n). Therefore, under the condition of the given generator, the objective function gets the minimum value at equation (5), which is the optimal solution of the discriminator. From equation (4), it can be seen that the metal plating model GAN ​​constructed in this embodiment evaluates the ratio of two probability distribution densities.

[0032] (5) Further, D ( x ) represents x The probability of image data satisfying the threshold range or generated data exceeding the threshold range from the CCD detection device. x hour, D The goal is to make the output probability value D ( x ) approaches 1, and when the input comes from the generator with generated data outside the threshold range G ( z )hour, D The goal is to correctly determine the source of data so that D (G(z)) approaches 0, and G The goal is to make it approach 1.

[0033] The CAN of the metal coating CCD detection mechanism based on the deep learning simulation model in this embodiment is a kind of G and D Game system, generator G The loss function is Obj G ( i G )=-Obj D ( i D， i G ), so the optimized GAN is a minimax problem, which leads to the objective function as shown in equation (6).

[0034] (6) In summary, the learning process of the CAN game system of the metal coating CCD detection mechanism based on the deep learning simulation model in this embodiment requires training the model D To maximize the identification of image data that meets the threshold range or generates data distribution that exceeds the threshold range from the CCD detection device G ( z ) accuracy, and at the same time, it is necessary to train the modeling data model G To minimize log(1- D ( G (z))). The game system can adopt an alternating optimization method: first fix the generator G , optimize the discriminator D , so that D The accuracy of discrimination is maximized; then the generator is fixed D , optimize the discriminator G , so that D The accuracy of the discrimination is the most accurate. P data = P g When , the global optimum is achieved, which becomes the preferred standard for CCD detection institutions. When training GAN, in the unified round parameter update, generally D Parameter update K Second, G The parameters of are updated once.

[0035] Embodiment 2: This embodiment provides a CCD detection method for semiconductor metal coatings based on a GAN neural network, which uses the detection model constructed by the method described in the first embodiment to classify semiconductor plated parts as good or defective, and specifically includes the following steps: Step 1: Use a precision balance to obtain the actual mass of the semiconductor plated part to be tested m 实测值 ; Step 2: Obtain CCD morphological parameters of the semiconductor plated part to be inspected; Step 3: Calculate the total mass simulation value of the semiconductor plated part based on the CCD morphology parameters M CCD , the calculation method is: M CCD = m 镀件质量 + m 镀层1 + m 镀层2 +...... + m 镀层n m 镀件质量 = ρ v m 镀层1 = p 1 v 1 m 镀层2 = p 2 v 2 ...... m 镀层n = p n v n in, r is the metal density of the semiconductor plated part; v is the metal volume of the semiconductor plated part; r 1 is the metal density of the first coating layer of the semiconductor plated component; v 1 is the volume of the first coating layer; r 2 is the metal density of the second coating layer of the semiconductor plated component; v 2 is the volume of the second coating layer; and so on. r n is the metal density of the nth coating layer of the semiconductor coating; v n is the volume of the nth coating layer.

[0036] Step 4: Actual quality of semiconductor coatings m 实测值 and the total mass simulation value M CCD Make a judgment: m 实测值 (100%-5%)≤ M CCD ≤ m 实测值 (100%+5%) If the above relationship is not satisfied, the semiconductor plated part is judged as a defective product; if the above relationship is satisfied, the next step of the GAN neural network judgment is carried out; Step 5: The actual quality of the semiconductor plated parts m 实测值 and the total mass simulation value M CCD The GAN neural network trained in Example 1 is input, and the results of the GAN neural network adversarial processing are divided into two categories: Good product; the image data processing results of electronic electroplating products meet the morphology threshold range; Missing products: The image data processing results of electronic electroplating products do not meet the morphology threshold range.

[0037] This embodiment achieves accurate detection of good and defective electronic electroplating products through the constructed GAN adversarial network.

[0038] Embodiment three: This embodiment describes in detail the process of constructing the detection model data set of the present invention in conjunction with an actual case of a semiconductor plated part to be processed, which specifically includes the following contents.

[0039] The semiconductor plated test piece to be processed in this embodiment is as follows Figure 3 The copper alloy material is 60×60mm and the thickness is 0.127mm. There are 16 basic units in total. The size of the copper-plated area 10 of each unit is 11.5mm×11.5mm, and its area is 132.25mm 2 ; The total area of ​​a single side of a semiconductor copper plated piece is 16×132.25=2166mm 2 In addition, the size of the silver-plated area 20 of each unit is 7.3mm×7.3mm, and its area is 53.29mm 2 ; The total area of ​​a single silver-plated semiconductor is 16×53.29=852.64mm 2 .

[0040] The electroplating specifications of the plating area are: the thickness of the Cu plating layer is 1700nm, and the thickness of the Ag plating layer is 5000nm.

[0041] The CCD detection center position of the semiconductor plated part coating area in this embodiment is as follows Figure 5 The coordinates of the black point 21 in the center of the electroplating area of ​​each semiconductor plating unit are: 4 (14.6, 7.3), 3 (14.6, 21.9), 2 (14.6, 36.5), 1 (14.6, 51.1) in column I; 4 (21.9, 7.3), 3 (21.9, 21.9), 2 (21.9, 36.5), 1 (21.9, 51.1) in column II; 4 (36.5, 7.3), 4 (36.5, 21.9), 2 (36.5, 36.5), 1 (36.5, 51.1) in column III; 4 (51.1, 7.3), 3 (51.1, 21.9), 2 (51.1, 36.5), 1 (51.1, 51.1) in column IV.

[0042] like Figure 3 The quality of a semiconductor plated part before electroplating is shown m 镀件质量 It is further calculated and derived from Formula 2.

[0043] m 实测总质量 = m 镀件质量 + m 镀层1 + m 镀层2 + ...... + m 镀层n (Formula 2) According to Figure 3 The 3D design drawing of the semiconductor plated part shown, the volume of a semiconductor plated part v mm 3 The specific gravity of the copper alloy used is r mg / mm 3 , so its total mass is: m 镀件质量 = ρ v According to the copper-plated surface area of ​​a single side of a semiconductor plated piece s 2166mm 2 , the thickness of Cu plating h 铜1 The surface area of ​​a single-sided silver-plated semiconductor plated part is 1700nm. s 852.64mm 2 , Ag coating h 银2 is 5000nm, so its total mass is: m 镀层总质量 = m 铜镀层1 + m 银镀层2 = p 铜1 v 铜1 + p 银2 v 银2 = p 铜1 s 铜1 h 铜1 + p 银2 s 银2 h 银2 Step 1: The quality of a semiconductor plated part before electroplating m镀件质量 After alkaline degreasing and acid activation and drying in the electronic plating module 310, the quality of the plated parts before semiconductor plating after pretreatment is tested using an ultra-precision analytical balance m 镀件质量 , the test results are shown in Table 1.

[0044] Table 1 Quality of plated parts before semiconductor electroplating after pretreatment

[0045] As can be seen from Table 1, the mass range of samples 1 to 8 before electronic electroplating is between 2.6757 and 2.6793 grams. Taking the average value of 2.6773 grams as the standard, the minimum error is -0.06% and the maximum error is 0.08%. From the perspective of quality control before electroplating of semiconductor plated parts, compared with the overall quality control threshold of -5.0% to 5.0% of semiconductor plated parts, the modeling conditions are met.

[0046] Step 2: Use a dedicated electronic electroplating device to perform electronic copper plating treatment one by one.

[0047] It is well known in the field of electroplating technology that under conventional electronic plating conditions, a constant current is set to carry out the electronic plating process. However, due to the real-time changes of the components in the electronic plating solution, the surface environment of the plated parts will change differently during the actual electronic plating process, resulting in the applied constant current not being 100% effective for the electrolytic deposition of copper. Therefore, the current efficiency is usually used to represent the effective current percentage of electronic plating. The characteristic of this is that during the electronic plating process, the current efficiency fluctuates within a certain range as the electronic plating implementation environment changes. When the electronic plating process is completed, the semiconductor plated parts are cleaned and their quality is tested using an ultra-precision analytical balance. That is, after the pre-treatment, the quality of the copper-plated semiconductor plated parts is tested. m 镀件质量 + m 铜镀层质量 Indicated by Mix 1, the quality of copper plating m 铜镀层质量 Use mixed 1 quality m 镀件质量 + m 铜镀层质量 minus m 镀件质量 , the results are shown in Table 2.

[0048] Table 2 Copper plating quality

[0049] It can be seen from Table 2 that after electronic copper plating of samples 1 to 8, the mass range of the copper plating layer is between 0.0348 and 0.0377 grams. Taking the average value of 0.0360 grams as the standard, the minimum error is -3.4% and the maximum error is 4.6%. From the perspective of quality control of semiconductor plated parts after electronic copper plating, compared with the overall quality control threshold of semiconductor plated parts -5.0% to 5.0%, the modeling conditions are met.

[0050] Step 3: For the copper-plated samples 1 to 8 whose mass has been measured in Table 2, a dedicated electronic electroplating device is used to perform silver plating treatment one by one, and after cleaning, an ultra-precision analytical balance is used to test their mass, that is, the mass of the semiconductor plated parts after the final silver plating. m 镀件质量 + m （铜+银）镀层质量 Indicated by Mix 2, the quality of the silver coating m 银镀层质量 Use mixed 2 quality m 镀件质量 + m （铜+银）镀层质量 Subtract Mix 1 mass m 镀件质量 + m 铜镀层质量 , the results are shown in Table 3.

[0051] Table 3 Silver coating quality

[0052] It can be seen from Table 3 that after electronic silver plating of samples 1 to 8, the silver plating mass range is between 0.0477 and 0.0514 grams. Taking the average value of 0.0494 grams as the standard, the minimum error is -3.3% and the maximum error is 4.2%. From the perspective of quality control of semiconductor plated parts after electronic silver plating, compared with the overall quality control threshold of -5.0% to 5.0% of semiconductor plated parts, the modeling conditions are met.

[0053] Step 4: After step 1 is implemented and before step 2 is implemented, use a CCD detection mechanism to perform morphology tests on samples 1 to 8 before electroplating.

[0054] The plated sample test uses a dedicated CCD device to detect its overall morphology. After the acquired data is processed by the plated part + coating quality calculation module 330 and judged as Yes by the plated total quality threshold judgment module 340, its test data can be imported into the GAN neural network generation module 200 for deep learning. Otherwise, in the case of No, it will be judged as a defective product and cannot enter the next step of implementation.

[0055] The deep learning model includes: a training module 210, a stability generation module 220, a coating potential energy calculation module 230 and a model morphology discrimination module 240.

[0056] The training module 210 is used for training the adversarial training between the total mass of the plated parts and the total mass actually measured of the plated parts, so as to obtain a model that can be used for classification of good products and defective products.

[0057] Stability generation module 220, for processing the ordered arrangement and trend of the measured total mass; A coating potential energy calculation module 230, used for coating simulation calculation based on the measured total mass; The model morphology determination module 240 is used to determine the model morphology of the plated part based on the simulation calculation coating result.

[0058] If the model judgment result is Yes, it is a good product and its corresponding data is backed up in the memory 510. If the judgment result is No, it is a defective product and its measured image data cannot be backed up and stored.

[0059] Quality of semiconductor plated parts before electroplating based on deep learning model morphology discrimination of CCD device detection image data M 镀件质量 , the results are shown in Table 4.

[0060] Table 4 Quality of semiconductor plated parts detected by CCD

[0061] It can be seen from Table 4 that the mass range of samples 1 to 8 before electronic electroplating based on CCD device detection is between 2.6759 and 2.6785 grams. Taking the average value of 2.6770 grams as the standard, the minimum error is -0.04% and the maximum error is 0.06%. From the perspective of quality control before electroplating of semiconductor plated parts, compared with the overall quality control threshold of -5.0% to 5.0% of semiconductor plated parts, the modeling conditions are met.

[0062] Step 5: After step 2 and before step 3, a CCD detection mechanism is used to perform morphology tests on electronic electroplated copper samples 1 to 8; the test uses a dedicated CCD device to detect its overall morphology, and after the data is obtained and judged as qualified by the total quality threshold judgment module 340 of the plated part, the test data is imported into the GAN neural network generation module 200 for deep learning, and the content of the deep learning is similar to step 4.

[0063] Quality of electroplated copper in semiconductor parts based on deep learning model morphology discrimination of CCD device detection image data M 镀件质量 + M 铜镀层质量 Indicated by Mix 3, the copper plating quality M铜镀层质量 Use mixed 3 quality M 镀件质量 + M 铜镀层质量 minus M 镀件质量 , the results are shown in Table 5.

[0064] Table 5 Quality of copper coating on semiconductor plated parts detected by CCD

[0065] It can be seen from Table 5 that the mass range of the electronic electroplated copper coating of Samples 1 to 8 detected by the CCD device is between 0.0343 and 0.0363 grams. Taking the average value of 0.0351 grams as the standard, the minimum error is -2.2% and the maximum error is 3.5%. From the perspective of quality control before electroplating of semiconductor plated parts, compared with the overall quality control threshold of -5.0% to 5.0% of semiconductor plated parts, these measured data meet the modeling conditions.

[0066] It can be further observed that for the same test samples 1 to 8, the copper plating quality of the semiconductor plated parts obtained by the metal plating CCD detection method based on the deep learning simulation model of this embodiment is M (g) The error range is -2.2% to 3.5%, which is better than the measured semiconductor copper coating quality obtained in step 1. m (g) The error range is -3.4% to 4.6%. This result confirms that the GAN neural network adversarial computing model of the present invention can further improve and optimize the quality of the measured semiconductor copper coating. m (g) Precision of the results.

[0067] Step 6: For the silver-plated samples 1 to 8 after step 3, use a CCD detection mechanism to perform morphology tests on the electronically electroplated silver samples 1 to 8.

[0068] The overall morphology is detected by a dedicated CCD device. After the acquired data is judged as qualified by the plated part total quality threshold judgment module 340, the test data is imported into the GAN neural network generation module 200 for deep learning. The deep learning includes contents similar to step 4 and will not be repeated here.

[0069] Quality of silver electroplating on semiconductor parts based on deep learning model morphology discrimination of CCD device detection image data M 镀件质量 + M （铜+银）镀层质量 Indicated by Mix 4, the quality of the silver coating M 银镀层质量 Use Mix 4 Quality M 镀件质量 + M（铜+银）镀层质量 Subtract Mix 3 Mass M 镀件质量 + M 铜镀层质量 , the results are shown in Table 6.

[0070] Table 6 Silver coating quality of semiconductor plated parts detected by CCD

[0071] It can be seen from Table 6 that after electronic silver plating of samples 1 to 8, their mass range is between 0.0465 and 0.0490 grams. Taking the average value of 0.0475 grams as the standard, the minimum error is -2.1% and the maximum error is 3.2%. From the perspective of quality control of semiconductor plated parts after electronic copper plating, compared with the overall quality control threshold of -5.0% to 5.0% of semiconductor plated parts, the modeling conditions are met.

[0072] It can be further observed that for the same test samples 1 to 8, the quality of the silver coating of the semiconductor plated parts obtained by the metal coating CCD detection method based on the deep learning simulation model of the present invention is M (g) The error range is -2.1% to 3.2%, which is better than the measured semiconductor silver plating quality obtained in step 3. m (g) The error range is -3.3% to 4.2%. This result further confirms that the GAN neural network adversarial computing model of the present invention can effectively improve the quality of the measured semiconductor silver coating. m (g) Precision of the results.

[0073] Comparative Example Based on the use of CCD detection equipment and technology provided by market suppliers, only the errors in the coating area and the lack of product deformation can be detected in real time; however, the lack of plating leakage in the coating area is difficult to detect accurately.

[0074] The existing CCD detection technology can control the plating area by setting the electroplating area data model, but it can only be used when the color of the electroplated coating is significantly different from the raw material or the bottom metal coating. In addition, the deformation of the plated part can be compared with the standard image model to distinguish various deformations of the plated part. First, the shape and size test of the semiconductor single plated part is shown in Table 7.

[0075] Table 7 Shape and size of a single semiconductor plated component

[0076] From Table 7, it can be seen that the shape and size test results of samples 1 to 8 before electroplating that were cleaned in the pre-treatment ranged from 59.95 to 60.06 mm, with an average value of 60.005 mm. Compared with the product design standard value of 60 mm, the average error is 0.01%, the minimum error is -0.08%, and the maximum error is 0.10%. Therefore, the existing CCD can use the precise features of the plated part size test to control product size abnormalities caused by deformation.

[0077] Secondly, the existing CCD test on the size of the copper-plated local area is shown in Table 8.

[0078] Table 8 Existing CCD test on the size of copper-plated local area

[0079] As can be seen from Table 8, the copper plating area size test results for electronic copper plating samples 1 to 10 are in the range of 11.47 to 11.56 mm, and the average value of 11.506 mm is compared with the standard value of 11.5 mm for the copper plating area of ​​the product design. The average error is 0.05%, the minimum error is -0.26%, and the maximum error is 0.52%. Therefore, the existing CCD can use the precise characteristics of the plated part size test to control the product size abnormalities caused by deformation; however, since the color boundary of the copper plating area cannot be accurately identified, it is impossible to accurately determine whether the plating is leaking or seeping.

[0080] Furthermore, the test results of the silver-plated local area size of electronic copper electroplating samples 1 to 8 are shown in Table 9.

[0081] Table 9 CCD detection of silver-plated local area size test results

[0082] As can be seen from Table 9, the copper plating area size test results for electronic copper plating samples 1 to 8 are in the range of 7.27 to 7.35 mm, and the average value is 7.301 mm. Compared with the standard value of the copper plating area in product design of 7.3 mm, the average error is 0.01%, the minimum error is -0.41%, and the maximum error is 0.68%. Therefore, the existing CCD can use the precise characteristics of the plated part size test to control the product size abnormalities caused by deformation; however, since the color boundary of the hard-plated area cannot be accurately identified, it is impossible to accurately determine whether the silver plating is leaking or seeping.

[0083] Since the silver-plated layers of samples 1 to 8 are silvery white and do not form a clear contrast with the color of the underlying nickel plating layer, the current CCD detection technology is unable to detect and identify metal plating layers that do not form a clear contrast in color, that is, it cannot meet the requirements for accurate detection of whether the electroplating area is leaking or plated outside the area.

[0084] The current CCD detection technology can be used to accurately detect the lack of plating or penetration in the plating area, such as gold-plated semiconductor samples, due to its distinct color difference. Figure 3 The semiconductor plated parts shown are the starting raw materials. Figure 4 The electroplating area is replaced by nickel plating, and the nickel plating film thickness is 2500nm; further, the Figure 5 The electroplating area is replaced by gold plating, and the gold plating film thickness is 300nm. Samples 9 to 16 are prepared based on the nickel plating and gold plating film thickness. The preparation method is the same as that of sample 1. The CCD inspection results of the current stage of color difference of samples 1 to 8 and samples 9 to 16 are shown in Table 10 using the CCD detection equipment of the comparative example. The tests of samples 1 to 8 and samples 9 to 16 are based on the completion of silver plating and gold plating.

[0085] Table 10 CCD inspection results with color difference

[0086] It can be seen from Table 10 that the current CCD detection technology, for the detection results of the aforementioned silver-white samples 1 to 8, although it can monitor the deformation of each area of ​​the product, it cannot confirm the silver-white electroplating. Therefore, it is not suitable for the detection and control of such electroplated products; on the other hand, the detection results of the gold-plated samples 9 to 16 show that the plating area can be identified and confirmed, that is, the leakage plating within the plating area or the infiltration plating outside the plating area can be qualitatively detected and controlled.

[0087] Table 11 Comparison results between the present invention and the existing CCD detection method

[0088] From the comparison results in Table 11, it can be seen that the existing CCD detection device not only does not have the function of quantitative detection of all electroplating products in terms of electroplating color confirmation, but also does not have the ability to realize quantitative analysis and control of various metal coatings by detecting and controlling the quality of the plated parts, that is, the quality of the raw materials of the plated parts before electroplating and the quality of each coating.

[0089] Therefore, it is not difficult to see from the comparison results in Table 11 that the existing CCD detection devices are largely not fully applicable to the detection and quality management functions of electronic electroplating products of semiconductor plated parts; these technical problems that need to be solved urgently in the field of electronic electroplating, the present invention solves the technical problems existing in the existing CCD detection devices through the construction and application of a metal plating CCD detection method based on a deep learning simulation model, realizes the quality detection of electronically plated semiconductor plated parts, the simulation model of the CCD detection mechanism of GAN operation and its discrimination management, and provides an innovative detection and control method for the semiconductor electronic electroplating industry.

[0090] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0091] 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, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a metal coating CCD detection model, characterized in that: The method comprises: Step 1: Obtain the size data and coating area of ​​the semiconductor plated part to be processed; Step 2: Obtain the dimension data set of the coating area of ​​the semiconductor plated part to be processed and the metal coating information to be electroplated, and calculate the total mass of the plated part modeling m 建模总质量 , the calculation method is: m 建模总质量 = m 镀件质量 + m 镀层1 + m 镀层2 + ...... + m 镀层n Step 3: Randomly select the coating quality standard conditions, use the electronic electroplating module to perform the actual electroplating operation on the semiconductor plated part to be processed, and after the electroplating is completed, use a precision balance to measure the actual total mass of the plated part m 实测总质量 ; Step 4: Model the total mass of the plated part m 建模总质量 And the total mass of the plated parts m 实测总质量 When the two satisfy the following relationship: m 实测总质量 （100%-5%）≤ m 建模总质量 ≤ m 实测总质量 （100%+5%） Then save this set of data for training the GAN neural network; If the above relationship is not met, the coating quality standard conditions shall be based on the total mass of the plated parts measured. m 实测总质量 Make corrections and calculate the corrected coating quality standard conditions, return to step 3 to re-measure the total mass of the plated parts m 实测总质量 , and re-determine the total mass of the corrected plated model m 建模总质量 Whether the threshold range is met until the total mass of the plated part is modeled m 建模值 When the above relationship is met, the data set can be used for training the GAN neural network; Step 5: Model the total mass of the plated part m 建模总质量 And the total mass of the plated parts m 实测总质量 Import GAN neural network to model the total quality of plated parts m 建模总质量 And the total mass of the plated parts m 实测总质量 The GAN game system is trained and trained. The trained GAN neural network models the total mass of the plated parts according to the input m 建模总质量 And the total mass of the plated parts m 实测总质量 Classify electronic electroplating products into good products and defective products.

2. The method for constructing a metal coating CCD detection model according to claim 1, characterized in that: The model operation system of the GAN neural network is composed of sensitive condition parameter modeling, and the sensitive condition parameters include: type of electroplating solution, temperature of electroplating solution, specific gravity of electroplating solution, diameter of pipeline for conveying electroplating solution, type of electroplating power supply, current intensity, electroplating time, cathode current efficiency, surface area of ​​plating area, metal density of electroplating layer, current density and flow rate of electroplating solution.

3. The method for constructing a metal coating CCD detection model according to claim 1, characterized in that: The semiconductor plated parts to be processed include: precision terminals, lead frames, wafer chips and precision decorations.

4. The method for constructing a metal coating CCD detection model according to claim 1, characterized in that: The metal coating comprises: Monomer plating Au, Ag, Ni, Sn, Cu, Pd, Rh, Pt; Alloy coatings Pd-Ni, Ni-P, W-Ni, Ag-Sn, Rh-Ru and Pt-Rh.

5. A system for constructing a metal coating CCD detection model, characterized in that: The system comprises: A memory (510) for storing data; A CCD detection module (320) is used to obtain CCD morphological parameters of the semiconductor plated part to be processed; An electronic electroplating module (310) is used to randomly extract a plating quality standard condition and perform an actual electroplating operation on the semiconductor plated part to be processed; Plated component + coating mass calculation module (330), used to calculate the total mass of plated component modeling m 建模总质量 , the calculation method is: m 建模总质量 = m 镀件质量 + m 镀层1 + m 镀层2 + ...... + m 镀层n The total mass threshold determination module of the plated part is used to model the total mass of the plated part. m 建模总质量 And the total mass of the plated parts m 实测总质量 When the two satisfy the following relationship: m 实测总质量 （100%-5%）≤ m 建模总质量 ≤ m 实测总质量 （100%+5%） Then save this set of data for training the GAN neural network; If the above relationship is not met, the coating quality standard conditions shall be based on the total mass of the plated parts measured. m 实测总质量 Make corrections and calculate the corrected coating quality standard conditions, return to step 3 to re-measure the total mass of the plated parts m 实测总质量 , and re-determine the total mass of the corrected plated model m 建模总质量 Whether the threshold range is met until the total mass of the plated part is modeled m 建模值 When the above relationship is met, the data set can be used for training the GAN neural network; GAN neural network generation module (200), based on the total mass of the plated part modeling m 建模总质量 And the total mass of the plated parts m 实测总质量 The GAN game system is trained and trained. The trained GAN neural network models the total mass of the plated parts according to the input m 建模总质量 And the total mass of the plated parts m 实测总质量 Classify electronic electroplating products into good products and defective products.

6. The system for constructing a metal coating CCD detection model according to claim 5, characterized in that: The GAN neural network generation module (200) comprises: A training module (210) is used for training the total mass of the plated parts and the total mass actually measured for the plated parts through adversarial training; A stability generation module (220) for processing the ordered arrangement and trend of the measured total mass; A coating potential energy calculation module (230), used for coating simulation calculation based on the measured total mass; The model morphology discrimination module (240) is used to discriminate the model morphology of the plated part based on the simulation calculation plating result.

7. A CCD detection method for semiconductor metal coating based on GAN neural network, characterized in that: The method uses a CCD detection model constructed according to any one of claims 1 to 4 to detect a metal coating of a semiconductor plated part, comprising: Step 1: Use a precision balance to obtain the actual mass of the semiconductor plated part to be tested m 实测值 ; Step 2: Obtain CCD morphological parameters of the semiconductor plated part to be inspected; Step 3: Calculate the total mass simulation value of the semiconductor plated part based on the CCD morphology parameters M CCD , the calculation method is: M CCD = m 镀件质量 + m 镀层1 + m 镀层2 + ...... + m 镀层n m 镀件质量 = ρ v m 镀层1 = ρ 1 v 1 m 镀层2 = ρ 2 v 2 ...... m 镀层n = ρ n v n in, ρ is the metal density of the semiconductor plated part; v is the metal volume of the semiconductor plated part; ρ 1 is the metal density of the first coating layer of the semiconductor plated component; v 1 is the volume of the first coating layer; ρ 2 is the metal density of the second coating layer of the semiconductor plated component; v 2 is the volume of the second coating layer; and so on. ρ n is the metal density of the nth coating layer of the semiconductor coating; v n is the volume of the nth coating; Step 4: Actual quality of semiconductor coatings m 实测值 and the total mass simulation value M CCD Make a judgment: m 实测值 （100%-5%）≤ M CCD ≤ m 实测值 （100%+5%） If the above relationship is not satisfied, the semiconductor plated part is judged as a defective product; if the above relationship is satisfied, the next step of the GAN neural network judgment is carried out; Step 5: The actual quality of the semiconductor plated parts m 实测值 and the total mass simulation value M CCD Input the trained GAN neural network, and the results of the GAN neural network adversarial processing are divided into two categories: Good product: The image data processing results of the electronic electroplating product meet the morphology threshold range; Missing products: The image data processing results of electronic electroplating products do not meet the morphology threshold range.

Images