Semiconductor integrated circuit chip welding method

Abstract

The invention discloses a semiconductor integrated circuit chip welding method, and relates to the technical field of semiconductor welding processes, and the method comprises the steps: collecting oxide layer thickness data and substrate microstructure data, inputting the data into a pre-trained convolutional neural network, and generating an oxidation state coding matrix; the optimal concentration ratio of the polyethylene glycol and glucan aqueous two-phase solution is calculated through a gradient matching algorithm, zoning path planning is carried out on the optimal concentration ratio, and a zoning spraying path instruction set is generated; the partition spraying path instruction set is transmitted to a micro-nozzle array to execute double-aqueous-phase spraying, meanwhile, liquid film thickness distribution data of a spraying area are collected, and a uniformity feedback data set is generated; and constructing an atomic layer deposition kinetic model, and inputting the uniformity feedback data set into the atomic layer deposition kinetic model. According to the method, the concentration proportion of the two-aqueous-phase solution is dynamically matched by analyzing the oxidation state coding matrix, and the partition spraying path is planned, so that the self-adaptive adjustment of the oxidation activity of the surface of the bonding pad and the solution proportion is realized.

Description

Technical Field

[0001] This invention relates to the field of semiconductor welding technology, and in particular to a method for welding semiconductor integrated circuit chips. Background Technology

[0002] Semiconductor integrated circuit chip bonding technology has evolved from wire bonding, flip chip bonding, and wafer-level packaging to high-density interconnection and micro / nano-scale precision bonding. With increasing chip integration and the widespread adoption of 3D stacking technology, the bonding interface is transitioning from the micrometer to the sub-micrometer scale, placing higher demands on interface thermal stress control, material diffusion behavior, and metallurgical bonding quality. In recent years, processes such as laser-assisted bonding and nano-silver sintering have made progress in the fields of power devices and high-frequency chips.

[0003] Current semiconductor integrated circuit chip welding technology has several shortcomings: insufficient quantitative analysis of the chip oxide layer thickness distribution and substrate microstructure during the pre-welding process; local thickness differences in the oxide layer lead to fluctuations in solder wettability; changes in substrate roughness cause uneven distribution of interfacial contact thermal resistance, significantly reducing the fatigue life of the solder joint; in addition, the aqueous two-phase solution spraying and atomic layer deposition processes lack a real-time thickness feedback mechanism, requiring production line interruption for offline inspection; fluctuations in the spraying liquid film thickness lead to disordered stacking of subsequent deposition layers; decreased atomic layer step coverage results in increased interfacial lattice distortion rate, weakening the shear strength of the weld joint. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for welding semiconductor integrated circuit chips to solve the problems of insufficient microscopic characteristic adaptation and lack of deposition feedback mechanism.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for bonding semiconductor integrated circuit chips, comprising: acquiring oxide layer thickness data and substrate microstructure data, inputting them into a pre-trained convolutional neural network to generate an oxide state encoding matrix; calculating the optimal concentration ratio of polyethylene glycol and dextran aqueous two-phase solution using a gradient ratio algorithm, performing partitioned path planning on the optimal concentration ratio to generate a partitioned spraying path instruction set; transmitting the partitioned spraying path instruction set to a micro-nozzle array to execute aqueous two-phase spraying, while simultaneously acquiring liquid film thickness distribution data in the spraying area to generate a uniformity feedback dataset; constructing an atomic layer deposition dynamics model, and inputting the uniformity feedback dataset into the atomic layer... A deposition kinetics model is used to generate atomic layer control signals, and atomic film thickness parameters are calculated based on these signals. A pre-stored material yield strength database is invoked, and combined with the atomic film thickness parameters, a pulsed laser scanning path and energy density scheme are generated. Theoretical strength parameters are calculated simultaneously, and fluorescence spectral data of the welding interface are collected. Based on the theoretical strength parameters, a defect coordinate set is generated through defect location analysis. A laser repair welding path sequence is generated based on the defect coordinate set, and the repair welding operation is performed. The laser energy parameters and residence time parameters for each repair welding point are corrected until the measured shear strength of the welding interface reaches the preset strength threshold. A welding quality certification report is then output.

[0007] In a preferred embodiment of the semiconductor integrated circuit chip welding method of the present invention, the specific steps for generating the oxidation state encoding matrix are as follows: The collected oxide layer thickness data were normalized, and the collected substrate microstructure data were subjected to Gaussian filtering for noise reduction. The normalized oxide layer thickness data and the noise-reduced substrate microstructure data are combined into a dual-channel input matrix and then input into a pre-trained convolutional neural network. Features are extracted through convolution and max pooling operations of a convolutional neural network, and then mapped through a fully connected layer to generate an oxidized state encoding matrix.

[0008] In a preferred embodiment of the semiconductor integrated circuit chip welding method of the present invention, the specific steps for calculating the optimal concentration ratio of polyethylene glycol and dextran aqueous two-phase solution are as follows: Analyze the spatial dimension features of the oxidation state encoding matrix and extract the oxidation activity level value corresponding to each matrix structure; Based on the oxidation activity level value, the optimal concentration ratio of polyethylene glycol and dextran aqueous two-phase solution was calculated using a gradient ratio algorithm.

[0009] In a preferred embodiment of the semiconductor integrated circuit chip welding method of the present invention, the specific steps for generating the partitioned spraying path instruction set are as follows: The optimal concentration ratio is mapped to the physical space coordinates of the pads corresponding to the oxidation state encoding matrix to generate a heat map of the concentration ratio distribution. Based on the concentration ratio distribution heatmap, the optimal movement path of the micro-nozzle is generated using a path planning algorithm.

[0010] In a preferred embodiment of the semiconductor integrated circuit chip welding method of the present invention, the specific steps for generating the uniformity feedback dataset are as follows: The partitioned spraying path instruction set is parsed into a control signal sequence for the micro-nozzle array, which is then driven to move to the target coordinate position according to the control signal sequence. The two-phase spraying deposition of polyethylene glycol solution and dextran solution is then executed sequentially to generate the scanning spraying area. The spraying area is scanned by a white light interferometer to collect liquid film thickness distribution data, and median filtering noise reduction is performed on the liquid film thickness distribution data. The denoised liquid film thickness distribution data is mapped to a thickness distribution matrix according to the physical space coordinates of the pads, and the standard deviation of the thickness distribution matrix is ​​calculated to generate a uniformity feedback dataset.

[0011] In a preferred embodiment of the semiconductor integrated circuit chip welding method of the present invention, the specific steps for constructing the atomic layer deposition dynamics model are as follows: An input layer is constructed based on a uniformity feedback dataset, a core dynamics layer is constructed based on a surface diffusion equation and a thickness feedback control term, a signal generation layer is constructed based on a PID control equation, and an output layer is constructed based on thickness integral calculation. An atomic layer deposition dynamics model is constructed based on the input layer, core dynamics layer, signal generation layer, and output layer.

[0012] In a preferred embodiment of the semiconductor integrated circuit chip welding method of the present invention, the specific steps for generating the atomic film thickness parameters are as follows: The uniformity feedback dataset is input into the atomic layer deposition dynamics model, and atomic layer control signals are generated through serial hierarchical physical calculations. The atomic film thickness parameter is generated by integral calculation based on the deposition rate parameter and electron signal duration in the atomic layer control signal.

[0013] In a preferred embodiment of the semiconductor integrated circuit chip welding method of the present invention, the specific steps for generating the pulsed laser scanning path and energy density scheme are as follows: The pre-stored material yield strength database is called to extract the yield strength parameters of the chip material and the substrate material. Combined with the atomic film thickness parameter, the stiffness distribution matrix of the welding interface is calculated. Based on the stiffness distribution matrix and yield strength parameters, a pulsed laser scanning path and energy density scheme are generated through path optimization and energy allocation.

[0014] In a preferred embodiment of the semiconductor integrated circuit chip welding method of the present invention, the specific steps for generating the defect coordinate set are as follows: Theoretical intensity parameters were calculated based on the pulsed laser scanning path and energy density scheme, and fluorescence spectral data of the welding interface were collected. By combining theoretical intensity parameters and fluorescence spectral data, defect localization analysis is performed through dual-channel image fusion and morphological segmentation to generate a defect coordinate set.

[0015] In a preferred embodiment of the semiconductor integrated circuit chip welding method of the present invention, the specific steps for outputting the welding quality certification report are as follows: Laser energy parameters and dwell time parameters are initialized based on the defect coordinate set to generate a laser repair welding path sequence; The welding operation is performed according to the laser welding path sequence. The interface shear strength data of the welding area is collected in real time. The strength deviation value is calculated based on the interface shear strength data and the preset strength threshold. If the strength deviation exceeds the limit, the laser energy parameters and dwell time parameters are corrected, and the repair welding operation is repeated until the strength comparison is completed, until the strength deviation does not exceed the limit, and a welding quality certification report is output.

[0016] The beneficial effects of this invention are as follows: By analyzing the oxidation state encoding matrix to dynamically match the concentration ratio of the two-phase aqueous solution and planning the partitioned spraying path, the adaptive adjustment of the oxidation activity of the solder pad surface and the solution ratio is achieved, which accurately eliminates the poor soldering and metal compound segregation caused by uneven wetting. At the same time, the path avoids low viscosity areas, which significantly reduces the risk of splash contamination. By constructing an atomic layer deposition kinetic model to carry out closed-loop control of surface diffusion and real-time thickness feedback, the deposition rate and residence time parameters are dynamically calculated, and the deposition process is continuously controlled online to eliminate liquid film fluctuations and lattice distortion, avoiding efficiency loss due to interrupted detection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of a method for soldering semiconductor integrated circuit chips.

[0019] Figure 2 A flowchart for generating a set of instructions for zoned spraying paths.

[0020] Figure 3 This is a flowchart for generating atomic film thickness parameters.

[0021] Figure 4 A flowchart for generating a welding quality certification report. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for welding semiconductor integrated circuit chips, including the following steps: S1. Collect oxide layer thickness data and substrate microstructure data, input them into a pre-trained convolutional neural network, and generate an oxide state encoding matrix. S1.1 Use an ellipsometry to scan the surface of the semiconductor chip pads to obtain oxide layer thickness data, and at the same time use an atomic force microscope to collect substrate microstructure data in the substrate pad area. It should be noted that the ellipsometry emits a beam of linearly polarized light of a fixed wavelength to illuminate the surface of the semiconductor chip pads. After the linearly polarized light is reflected by the oxide layer interface, its polarization state changes to elliptically polarized light. The oxide layer thickness data is directly collected by the built-in optical sensor, and the oxide layer thickness data of all measurement points are arranged according to spatial coordinates to generate a 500×500 oxide layer thickness data dot matrix. The substrate pad area is scanned by the silicon probe of the atomic force microscope in a contact mode to collect the three-dimensional morphology data of the surface in a tapping mode, and the substrate micro-morphology data dot matrix (500×500) is output. The ellipsometry has a scanning step size of 20 micrometers covering a 10mm×10mm pad area, and the atomic force microscope has a scanning range that matches the pad size and a resolution of 0.1nm. Finally, the oxide layer thickness data and the substrate micro-morphology data are output.

[0026] S1.2 Normalize the oxide layer thickness data to the 0-1 range and perform Gaussian filtering noise reduction on the substrate microstructure data; It should be noted that the global maximum and minimum values ​​of the oxide layer thickness data matrix are extracted, and then scaled point by point according to the scaling formula to generate normalized oxide layer thickness data. The expression is as follows: ; in, This represents the normalized oxide layer thickness data; This indicates the oxide layer thickness data; Represents the global minimum value; Represents the global maximum value; A 3×3 neighborhood window is established with each substrate micro-morphology data point in the substrate micro-morphology data matrix as the center. Based on the weight distribution template (the center has the highest weight and the edge weights decrease), weighted fusion processing is performed on all substrate micro-morphology data within the neighborhood window to generate noise-reduced substrate micro-morphology data. The convolution process is completed by covering the entire 500×500 substrate micro-morphology data matrix point by point through a sliding window.

[0027] S1.3. The normalized oxide layer thickness data and the noise-reduced substrate microstructure data are merged into a dual-channel input matrix and input into a pre-trained convolutional neural network. It should be noted that the normalized oxide layer thickness data and the noise-reduced substrate microstructure data are aligned point by point according to the same spatial coordinate position. The normalized oxide layer thickness data of each coordinate point is filled into the corresponding position of the first layer of the dual-channel structure. At the same time, the noise-reduced substrate microstructure data of the current coordinate point is filled into the corresponding position of the second layer of the dual-channel structure to generate a three-dimensional data block, which is the dual-channel input matrix. A convolutional neural network (CNN) is trained on a large public image dataset (such as ImageNet) to learn general feature recognition capabilities. During training, natural images are input, and the weight arrays of convolutional layers, fully connected layers, and bias values ​​are progressively adjusted based on the image classification results to enable the CNN to correctly identify object categories. After training, the final determined weight arrays of convolutional layers, fully connected layers, and bias values ​​are saved. The CNN is then fine-tuned on a dedicated dataset for the oxidation state of semiconductor pads. This dataset contains paired samples of oxide layer thickness data and substrate microstructure data, along with oxidation activity level labels. The weight arrays of convolutional layers, fully connected layers, and bias values ​​are updated based on the oxidation activity classification results to obtain optimized weight arrays of convolutional layers, fully connected layers, and fixed bias values. The trained CNN is then output. Finally, a dual-channel input matrix is ​​input into the trained CNN.

[0028] S1.4. Features are extracted through convolution and max pooling operations of a convolutional neural network, and then mapped through a fully connected layer to generate an oxidized state encoding matrix.

[0029] It should be noted that the convolution operation calls the convolutional layer weight array and fixed bias value, and slides across the dual-channel input matrix with a stride of 2, covering a 7-row, 7-column region each time. The values ​​of each data point in the covered region are fused with the corresponding position of the weight array to generate new feature values. This process is repeated point by point to cover the entire dual-channel input matrix, forming a 64-layer feature map. The max pooling operation slides a 3×3 window across the 64-layer feature map, and each time the maximum value is directly selected from the 9 new feature values ​​in the sliding window for output. The sliding window is moved with a stride of 2 to cover all regions of the 64-layer feature map, compressing the size to 125×125, and outputting a dimensionality-reduced feature map. The dimensionality-reduced feature map is flattened into a one-dimensional feature vector. This one-dimensional feature vector is then combined with the weight array of the fully connected layer. Each fixed value in the weight array is automatically fused with the corresponding value in the one-dimensional feature vector to generate an intermediate value. A non-linear transformation function is used to zero out each intermediate value (negative values ​​are set to zero, positive values ​​are retained). Positive values ​​are linearly scaled to the 0-1 range, generating 2500 oxidation activity level values. These 2500 oxidation activity level values ​​are then filled into a 50×50 oxidation matrix row by row according to the physical space coordinates of the pads, forming the oxidation state encoding matrix. The row numbers of the oxidation state encoding matrix correspond to the Y-axis coordinates of the pads, and the column numbers correspond to the X-axis coordinates. The value stored at each position in the oxidation state encoding matrix represents the oxidation activity level at the corresponding coordinate point, thus completing the output of the oxidation state encoding matrix.

[0030] S2. Calculate the optimal concentration ratio of polyethylene glycol and dextran aqueous two-phase solution using a gradient ratio algorithm, perform zonal path planning on the optimal concentration ratio, and generate a zonal spraying path instruction set. S2.1 Analyze the spatial dimension features of the oxidation state encoding matrix and extract the oxidation activity level value corresponding to each matrix structure; It should be noted that the oxidation state encoding matrix (50×50 matrix) is scanned row by row and column by column, and the oxidation activity level value (range 0-1) stored at each position is extracted according to the row and column coordinate order of the oxidation state encoding matrix structure, generating an oxidation activity level value list containing 2500 oxidation activity level values.

[0031] S2.2 Based on the oxidation activity level value, the optimal concentration ratio of polyethylene glycol and dextran aqueous two-phase solution is calculated by gradient ratio algorithm; It should be noted that the global average value of the oxidation activity level list is used as the activity baseline. For each oxidation activity level, the optimal polyethylene glycol concentration ratio is dynamically calculated using the polyethylene glycol concentration ratio formula, which is expressed as follows: ; in, This indicates the optimal concentration ratio of polyethylene glycol; Indicates the activity baseline value; This indicates the activity deviation, which is the difference between the current oxidation activity level value and the activity baseline value. This represents the adjustment coefficient, calibrated through gradient welding experiments; The adjustment coefficient was calibrated through gradient soldering experiments. Specifically, 500 sets of solder pad samples with different oxidation activity levels were prepared, and each set of solder pad samples was divided into 100 micro-regions. Different adjustment coefficients were applied to each micro-region. (Step size 0.1) Calculate the polyethylene glycol concentration ratio and perform welding. Detect the weld void rate using X-ray and statistically determine the... The average void ratio of the full pad sample was the lowest. This is the optimal adjustment coefficient; Meanwhile, the optimal concentration ratio of dextran is always constrained to be the difference between the total concentration ratio and the optimal concentration ratio of polyethylene glycol, ensuring that the total concentration ratio of the aqueous two-phase solution is always 100%, that is, the sum of the optimal concentration ratios of polyethylene glycol and dextran remains completely complementary. The optimal concentration ratio sequence is generated by traversing all 2500 oxidation activity level values, and finally a list of optimal concentration ratio distributions of polyethylene glycol and dextran is output.

[0032] S2.3 Map the optimal concentration ratio to the physical space coordinates of the pads corresponding to the oxidation state encoding matrix to generate a heat map of the concentration ratio distribution; It should be noted that the optimal concentration ratio distribution list is mapped point by point to the physical space coordinates of the solder pads according to the original row and column order of the oxidation state coding matrix. The optimal concentration ratio of polyethylene glycol is divided into five continuous intervals according to the concentration value from low to high (based on the physical characteristics of the soldering process and experimental failure analysis). Each interval corresponds to a specific color: the lowest interval (e.g., optimal concentration ratio ≤ 0.3, where the solder cannot spread) is filled with dark blue to indicate the low-activity flux area, and the lower interval (e.g., optimal concentration ratio between 0.3 and 0.5, where the solder viscosity is insufficient) is filled with... The light blue area is used to indicate the low concentration region. The medium range (such as the optimal concentration ratio between 0.5 and 0.7, where humidity and viscosity are in the best balance) is filled with green to indicate the baseline concentration region. The higher range (such as the optimal concentration ratio between 0.7 and 0.8, where the viscosity is high and gas can be trapped) is filled with orange to indicate the medium-high concentration enhancement region. The highest range (such as the optimal concentration ratio ≥ 0.8, where the solder is too viscous and can cause microbubbles) is filled with dark red to indicate the high concentration enhancement region. The five colors form a continuous gradient transition from dark blue to dark red in the color gradation. A continuous color distribution map is generated by covering the corresponding coordinate area of ​​the pad with color levels point by point. After superimposing the pad border scale marks, a heat map of concentration ratio distribution is output.

[0033] S2.4. Based on the concentration ratio distribution heat map, the optimal moving path of the micro-nozzle is generated through a path planning algorithm; It should be noted that, in the concentration ratio distribution heatmap, the spatial distribution boundary between the high-concentration enhancement area marked by dark red and the low-activity flux area marked by dark blue is identified. Taking the high-concentration enhancement area marked by dark red as the starting point and the low-activity flux area marked by dark blue as the ending point, the optimal movement path of the micro-nozzle is generated through the shortest path algorithm: the straight-line distance between the starting point and the ending point is calculated and used as a reference line. 100×100 path nodes (100 micrometers apart) are evenly distributed in the pad area covered by the concentration ratio distribution heatmap. The path nodes are grouped according to the concentration color level (dark red, orange and green nodes are given priority, and light blue and dark blue nodes are given secondary priority). Starting from the starting point, the shortest connection line to the ending point is calculated for each path node. The connection line must pass through the orange and green node areas first (forced passage) and detour around the light blue node areas. The path node connection sequence with the shortest total length and the fewest turning points is selected to generate the optimal movement path of the micro-nozzle. It should also be noted that the light blue node area corresponds to a low-viscosity solder state with an optimal polyethylene glycol concentration ratio between 0.3 and 0.5. In the low-viscosity solder state, the solder viscosity is less than 50 centipoise, resulting in insufficient surface tension. When the micro-nozzle moves through it, the solder is easily subjected to inertial impact, generating splash droplets. The splash droplets will contaminate the adjacent pad area, creating a short circuit risk. At the same time, the low-viscosity solder causes uneven liquid film spreading, resulting in microvoids. To ensure soldering quality and yield, it is necessary to bypass the light blue area and maintain a buffer distance of at least 200 micrometers to isolate the splash risk.

[0034] S2.5. Combine the optimal movement path with the optimal concentration ratio to generate a set of instructions for the zoned spraying path.

[0035] It should be noted that the optimal movement path is matched point-by-point with the optimal concentration ratio distribution list according to the physical space coordinates of the pads. Each movement path node corresponds to an optimal concentration ratio value, generating a triplet instruction containing the coordinate position, the optimal concentration ratio of polyethylene glycol, and the optimal concentration ratio of dextran. Based on the optimal concentration ratio of polyethylene glycol, the dwell time of the micronozzle at the current position is automatically calculated (the dwell time is extended in high-concentration areas and shortened in low-concentration areas). The expression is as follows: ; in, Indicates the duration of stay; Indicates the reference time constant; All triplet instructions are arranged in the order of movement path and integrated into a partition spraying path instruction set; Specifically, the reference time constant is calibrated through a jetting experiment: a piezoelectric drive device is used to generate controllable microdroplet jets, and a high-speed camera is used to record the dynamic evolution of the droplets from formation to stabilization, collecting data on the changes in droplet length, volume, and velocity over time; based on the step response principle, the time corresponding to when the droplet morphology or motion parameters reach a stable state is the reference time constant; during the jetting experiment, it is necessary to ensure that the driving signal is a bipolar trapezoidal wave to stabilize droplet generation, and the droplet breakage delay and volume change are observed by adjusting the coaxial airflow intensity, and the fitting accuracy is optimized by utilizing the damped oscillation characteristics; this residence time only represents the duration of valve opening at the target coordinate point.

[0036] Furthermore, a droplet reaching a stable state means that after the droplet leaves the nozzle, its morphological oscillations significantly weaken (the amplitude of the morphological oscillations decreases to within 5% of the initial diameter, based on the definition of solder microdroplet quality inspection standards) until it is no longer visible to the naked eye, its shape approaches a regular ellipsoid and its surface is smooth; the trajectory of the center of mass transitions from acceleration to uniform linear motion, the tail fracture surface is completely closed and no satellite droplets are generated, at this time the droplet's kinetic energy and surface tension reach dynamic equilibrium, its length, volume and velocity remain constant for a short time, and continuous monitoring shows that the trend of change tends to be stable, which can be judged as entering a stable state.

[0037] It should also be noted that existing technologies use a fixed ratio of polyethylene glycol-dextran solution to uniformly spray the solder pads, relying on operator experience to adjust the concentration. This results in insufficient wetting in low-oxidation-activity areas, leading to cold solder joints, and solution accumulation in high-oxidation-activity areas, causing short circuits. In addition, the spraying path uses a linear reciprocating scanning motion, causing spatter and droplet contamination of adjacent solder pads due to inertial impact in low-viscosity areas. In contrast, S2 dynamically calculates the optimal concentration ratio for each solder joint using an oxidation state coding matrix and plans a path around low-viscosity areas based on the concentration ratio distribution heatmap. This achieves precise, on-demand spraying of the aqueous two-phase solution, solving the welding defects and cost losses caused by static mixing ratios and path collisions in existing technologies.

[0038] S3. Transmit the partitioned spraying path instruction set to the micro-nozzle array to perform dual-phase spraying, and at the same time collect the liquid film thickness distribution data of the spraying area to generate a uniformity feedback dataset. S3.1. Parse the partitioned spraying path instruction set into a control signal sequence for the micro-nozzle array, drive the micro-nozzle array to move to the target coordinate position according to the control signal sequence, and sequentially execute the two-phase spray deposition of polyethylene glycol solution and dextran solution; It should be noted that the partitioned spraying path instruction set is read line by line according to the triplet instruction sequence. The coordinate position, optimal polyethylene glycol concentration ratio, and optimal dextran concentration ratio in each triplet are extracted, along with the dwell time. The physical space coordinates of the solder pads are mapped to the actual position in the micro-nozzle array coordinate system using a coordinate mapping mathematical formula, generating a control signal sequence containing the actual movement coordinates, target concentration ratio, and dwell time. The expression is as follows: ; ; in, Represents a micro-nozzle array coordinate; Represents a micro-nozzle array coordinate; Indicates the physical space of the solder pads coordinate; Indicates the physical space of the solder pads coordinate; surface Indicates the physical space origin of the solder pads coordinate; Represents the physical space origin of the pads The coordinates, with the origin coordinates determined during pad mounting; Indicates the density of the micro-nozzle array; Indicates the number of micro-nozzles; Indicates the pad size; The micro-nozzle array is driven by a sequence of control signals to move to the target coordinate position according to the movement path sequence. Upon arrival, the piezoelectric jet controller is invoked to perform two-phase jetting: the opening of the inlet valve is adjusted according to the current optimal concentration ratio of polyethylene glycol, the opening is maintained according to the residence time and a precise volume of solution is released, the optimal concentration ratio channel of dextran is switched and the opening of the inlet valve and the timed release are repeatedly adjusted, and after the two-phase deposition is completed, the system moves to the next movement coordinate. The cycle is repeated until all control signals are processed.

[0039] S3.2. Scan the spraying area with a white light interferometer to collect liquid film thickness distribution data, and perform median filtering noise reduction on the liquid film thickness distribution data; It should be noted that the white light interferometer probe scans the surface area of ​​the sprayed pads in 20-micrometer steps. The white light interferometer emits a broadband light source to illuminate the liquid film surface, and the reflected light from the upper surface of the liquid film and the substrate interface forms interference fringes. The optical sensor captures the interference fringes in real time, measures the rate of change of the center spacing between adjacent bright or dark interference fringes (widening of the interference fringe spacing corresponds to an increase in thickness, and narrowing corresponds to a decrease in thickness), and simultaneously detects the fringe offset of the overall position of the interference fringes relative to the reference position (where the liquid film thickness is zero). Based on the physical ratio of the light wave center wavelength (e.g., 600 nm) to the fringe offset (a fringe offset of one full cycle corresponds to a thickness change of half a wavelength), the thickness is determined. The center spacing change rate and fringe offset are converted into liquid film thickness values: if the interference fringes shift to the left of the reference reference by a full cycle distance, the liquid film thickness value increases by half a wavelength (e.g., 300 nm); if they shift to the right by a full cycle distance, the liquid film thickness value decreases by half a wavelength. At the same time, the liquid film thickness value is determined by combining the width change of the interference fringe spacing (doubling the interference fringe spacing corresponds to an increase of half a wavelength in liquid film thickness). That is, the liquid film thickness value at the reference reference position (which is 0) is superimposed with the fringe offset and width change to generate the liquid film thickness value. Each scanning point generates a liquid film thickness value, which is filled into a 500×500 liquid film thickness matrix according to the scanning row and column order, and the original thickness data matrix is ​​output. Median filtering is performed point-by-point on the original thickness data matrix: Taking the current matrix structure as the center, nine liquid film thickness values ​​are extracted from the 3×3 neighborhood window. The nine liquid film thickness values ​​are sorted in ascending order of numerical value, and the value of the 5th position after sorting (the median value) is used to replace the original center point liquid film thickness value. Following this rule, starting from the upper left corner (1,1) of the matrix, the window slides to the right with a step size of 1, scanning row by row to the lower right corner (500,500) to achieve noise reduction of the entire original thickness data matrix and output the noise-reduced liquid film thickness distribution data. It should also be noted that the reference position refers to the coordinates of the center position of the interference fringes when the white light interferometer probe is aligned with an ideal plane (such as a polished silicon wafer) before scanning.

[0040] S3.3. Map the noise-reduced liquid film thickness distribution data to a thickness distribution matrix according to the physical space coordinates of the pads, and calculate the standard deviation of the thickness distribution matrix to generate a uniformity feedback dataset.

[0041] It should be noted that the noise-reduced liquid film thickness distribution data is sequentially mapped to the physical space coordinates of the pads according to the original thickness data matrix during white light interferometer scanning. The row number corresponds to the Y-axis coordinate of the pad's physical space, and the column number corresponds to the X-axis coordinate of the pad's physical space. The liquid film thickness value of each scanning point is directly filled into the corresponding row and column positions of the thickness distribution matrix to generate a 500×500 thickness distribution matrix. The arithmetic mean of all liquid film thickness values ​​in the thickness distribution matrix is ​​calculated as the average thickness. The standard deviation is then calculated based on the average thickness. The thickness distribution matrix and the standard deviation are integrated into a uniformity feedback dataset. The expression for calculating the standard deviation is as follows: ; in, This represents the standard deviation value; This represents the total number of matrix structures in the thickness distribution matrix; This represents the row index of the thickness distribution matrix, with values ​​ranging from 1 to 500; The column index represents the thickness distribution matrix, with values ​​ranging from 1 to 500; Indicates the first Line 1 The column shows the liquid film thickness values; This represents the average thickness.

[0042] S4. Construct an atomic layer deposition dynamics model, input the uniformity feedback dataset into the atomic layer deposition dynamics model, generate atomic layer control signals, and calculate the atomic film thickness parameters based on the atomic layer control signals. S4.1. Construct an input layer based on the uniformity feedback dataset, construct a core dynamics layer based on the surface diffusion equation and thickness feedback control term, construct a signal generation layer based on the PID control equation, and construct an output layer based on thickness integral calculation. It should be noted that the thickness distribution matrix in the uniformity feedback dataset is directly used as the thickness spatial distribution structure of the input layer, while the standard deviation value in the uniformity feedback dataset is independently extracted as the uniformity quantification parameter of the input layer. The thickness distribution matrix and the uniformity quantification parameter together constitute the data content structure of the input layer, where the thickness distribution matrix occupies the first 250,000 neurons of the input layer (arranged in row priority order), and the standard deviation value occupies the 250,001st neuron node, thus completing the construction of the input layer. The surface diffusion equation is used as the basic dynamic framework to directly establish the main architecture of the core dynamic layer. The thickness feedback control term is embedded as a dynamic correction structure within the basic dynamic framework to generate a thickness deviation signal. A thickness feedback control interface is reserved in the surface diffusion equation, allowing the thickness distribution matrix and standard deviation values ​​to be injected into the core dynamic layer in real time. This forms a synergistic coupling framework between surface diffusion and thickness feedback adjustment, completing the construction of the core dynamic layer. The expression for the surface diffusion equation coupled with the thickness feedback control term is as follows: ; in, This represents the real-time liquid film thickness, which is a dynamically calculated value from the equation. Indicates a specific moment in time; This represents the surface diffusion coefficient, which is determined by the pad material properties. This represents the feedback gain coefficient, calibrated based on deposition kinetics experiments; This represents the values ​​of the matrix elements in the thickness distribution matrix; Represent the surface diffusion equation; Indicates the thickness feedback control term; This represents the deposition rate parameter; The core framework of the signal generation layer is established using proportional control, integral control, and derivative control structures as its basic structures. A dedicated register array is created within the signal generation layer as a historical deviation accumulation storage area to continuously record the historical sequence of thickness deviation signals. The proportional control structure receives the thickness deviation signal output from the core dynamics layer, the integral control structure connects to the historical deviation accumulation storage area, and the derivative control structure receives the current thickness deviation and the previous thickness deviation to calculate and generate the thickness change rate. Parallel signal paths are established for the proportional, integral, and derivative control structures within the signal generation layer, and a weighted superposition node is established to fuse the outputs of the three structures, outputting the control signal and completing the construction of the signal generation layer. The expression for the PID control equation is as follows: ; in, express Control signals output at all times; express Thickness deviation at any given time; Indicates the proportional gain parameter; Indicates the integral gain parameter; Represents the differential gain parameter; Indicates the rate of change of thickness; Indicates the cumulative historical deviation; This represents a dummy variable during integration, indicating a temporary variable in the time integration process, referring to the time interval from 0 to the current moment. All moments; It should also be noted that the proportional gain parameter, integral gain parameter, and derivative gain parameter are all PID gain parameters, which are calibrated through gradient soldering experiments. The gradient soldering experiments cover 500 sets of solder pad samples, each set is divided into 100 micro-regions, and the gain combination of proportional gain parameter, integral gain parameter, and derivative gain parameter is tested to determine the optimal value with the goal of minimizing void rate and thickness fluctuation. The control signal receiving port is physically connected to the output of the signal generation layer; a signal conversion structure is set inside the output layer, which includes a voltage mapping structure and a time mapping structure; the output port is integrated at the end of the signal conversion structure and connected to the actuator drive interface; at the same time, a voltage conversion coefficient storage area is fixed for the voltage mapping structure, and a duration maintenance path is opened for the time mapping structure, thus completing the construction of the output layer.

[0043] S4.2 Construct an atomic layer deposition dynamics model based on the input layer, core dynamics layer, signal generation layer, and output layer; It should be noted that the input layer, core dynamics layer, signal generation layer, and output layer are connected in series to construct an atomic layer deposition dynamics model; The atomic layer deposition dynamics model does not employ a traditional data-driven training process. Core parameters are pre-calibrated and solidified through prior physical experiments: the surface diffusion coefficient is stored in a material property database; the feedback gain coefficient and PID gain parameters are calibrated to optimal values ​​based on gradient welding experiments; after the atomic layer deposition dynamics model is constructed, deterministic physical calculations are directly performed: the input layer receives the thickness distribution matrix and standard deviation values ​​from the uniformity feedback dataset; the core dynamics layer couples the surface diffusion equation and thickness feedback term to solve the deposition rate parameters and thickness deviation data in real time; the signal generation layer processes the thickness deviation data through proportional, integral, and differential paths to generate control signals; the output layer uses a voltage mapping unit to multiply the deposition rate parameter and voltage conversion coefficient as the voltage amplitude signal; and a time mapping unit uses the product of the control signal value and the time scaling factor as the electronic signal duration. It should also be noted that the voltage conversion coefficient was determined through a voltage-flow calibration experiment of the micro-nozzle. A stepped voltage signal (0 to 10 volts, with a step size of 0.5 volts) was applied to the micro-nozzle. Based on the linear relationship curve between the deposition rate parameter and the voltage signal, the reciprocal of the slope of the linear relationship curve was taken as the voltage conversion coefficient. The time scaling factor was determined through a voltage-response delay calibration experiment of the piezoelectric valve. A stepped voltage signal (0 to 10 volts, with a step size of 0.5 volts) was applied to the piezoelectric valve, and the mechanical delay time from the rising edge of the voltage signal to the full opening of the valve was recorded using a high-speed camera. The delay time value corresponding to each voltage signal was measured, and a linear relationship curve between voltage and delay time was fitted. The slope of the curve was taken as the time scaling factor.

[0044] S4.3 Input the uniformity feedback dataset into the atomic layer deposition dynamics model, and generate atomic layer control signals through serial hierarchical physical calculations; It should be noted that the thickness distribution matrix in the uniformity feedback dataset is input into the first 250,000 nodes of the atomic layer deposition kinetics model input layer in row-major order, and the standard deviation value is input into the 250,001st node; the core kinetics layer calls the surface diffusion equation coupled with the thickness feedback control term, takes the thickness distribution matrix cell value as the target thickness, and calculates the liquid film thickness change in real time according to the expression of the surface diffusion equation coupled with the thickness feedback control term to generate deposition rate parameters. At the same time, the difference between the target thickness and the real-time thickness is calculated to generate thickness deviation data. The signal generation layer receives thickness deviation data, inputs the thickness deviation data into a proportional control structure, and uses the product of the proportional gain parameter and the thickness deviation data as the real-time deviation response component; it also inputs the thickness deviation data into an integral control structure, and uses the product of the integral gain parameter and the historical deviation accumulation as the historical correction component; and it inputs the thickness deviation data into a differential control structure, and uses the product of the differential gain parameter and the thickness change rate as the trend suppression component. The real-time deviation response component, the historical correction component, and the trend suppression component are fused through a weighted overlay node to generate a control signal. The output layer uses the product of the deposition rate parameter and the voltage conversion coefficient as the voltage amplitude signal, and the product of the control signal and the time scaling factor as the electronic signal duration. The voltage amplitude signal and the electronic signal duration are integrated to generate the atomic layer control signal.

[0045] S4.4. Based on the deposition rate parameter and electron signal duration in the atomic layer control signal, the atomic film thickness parameter is generated by integral calculation.

[0046] It should be noted that the deposition rate parameter in the atomic layer control signal is integrated in chronological order, and the duration of the electron signal at the corresponding time point is extracted as the integration time step. Starting from the initial zero time, the first thickness increment is obtained by multiplying the deposition rate parameter at the first time point by the duration of the electron signal at the first time point. Similarly, all thickness increments are accumulated point by point in chronological order to generate the cumulative thickness increment. This is then superimposed on the initial zero-thickness substrate to generate the atomic film thickness value. The atomic film thickness value is filled into the corresponding row and column positions of the thickness distribution matrix according to the physical space coordinate mapping relationship of the pads, and the atomic film thickness parameter is output.

[0047] It should also be noted that existing technologies control the atomic layer deposition process by pre-setting fixed deposition rates and empirical residence time parameters. This relies on manual interruption of the process for offline thickness detection and manual adjustment of parameters, resulting in large fluctuations in deposition thickness and lag in response, leading to local over-deposition. In contrast, this solution constructs an atomic layer deposition kinetic model, which dynamically calculates the deposition rate by collecting liquid film thickness distribution data in real time, combining surface diffusion physics equations and thickness feedback terms, and uses a proportional-integral-differential algorithm to convert thickness deviations into control signals. This enables adaptive and precise adjustment of deposition rate and residence time, significantly reducing the thickness fluctuation range compared to existing technologies, completely eliminating the risk of lattice distortion, and simultaneously enabling online continuous control to avoid production line interruptions, thereby improving deposition efficiency and material utilization.

[0048] S5. Call the pre-stored material yield strength database and combine it with atomic film thickness parameters to generate a pulsed laser scanning path and energy density scheme, and simultaneously calculate theoretical strength parameters, collect fluorescence spectrum data of the welding interface, and generate a defect coordinate set through defect location analysis in combination with theoretical strength parameters; S5.1. Call the pre-stored material yield strength database, extract the yield strength parameters of the chip material and the substrate material, and calculate the stiffness distribution matrix of the welding interface by combining the atomic film thickness parameters. It should be noted that, in preparing standard tensile specimens of chip materials and substrate materials (such as silicon wafers and FR4 boards), axial tensile loads are applied on a universal testing machine, and the deformation of the specimens is recorded in real time using an extensometer. Stress-strain curves are plotted, and the stress value corresponding to the starting point of the yield plateau in the stress-strain curve is extracted as the yield strength parameter. The experiment is repeated thirty times for each material, and the arithmetic mean is taken after removing outliers as the final yield strength parameter. Data such as material type, yield strength value, test temperature, and loading rate are classified and stored in a spreadsheet according to the material number, ultimately forming a material yield strength database. The process involves accessing a material yield strength database and extracting the yield strength parameters of the semiconductor chip material and the substrate material using the material type index. The target point's material region is determined by its physical spatial coordinates. If the target point is located in the chip welding area, the chip material yield strength parameter is selected; otherwise, the substrate material yield strength parameter is selected. The atomic film thickness parameter corresponding to the target point is then associated with the selected material yield strength parameter. Based on the welding interface stiffness calculation formula, the stiffness value at each coordinate position of the pad is calculated point by point. The calculated stiffness values ​​are then filled into a 500×500 matrix according to spatial coordinate order to generate the welding interface stiffness distribution matrix. The expression for the stiffness calculation formula is as follows: ; in, Indicates the stiffness of the weld interface; Indicates the atomic film thickness parameter; This represents the material's yield strength parameter; This represents the geometry factor, which is determined by the shape of the pads.

[0049] S5.2 Based on the stiffness distribution matrix and yield strength parameters, a pulsed laser scanning path and energy density scheme are generated through path optimization and energy allocation; It should be noted that the high-stiffness and low-stiffness regions are determined by stiffness values: the partition range is set by statistical characteristics of the stiffness distribution matrix and welding deformation test data; the 25th percentile (e.g., 100 N / m) and 75th percentile (e.g., 500 N / m) of the stiffness distribution are calculated by performing percentile statistics on the stiffness values ​​of the entire weld pad area; based on the fatigue test data of the welding interface (500 sets of weld point thermal cycle tests), the statistical critical stiffness value for weld cracking is the median value of 300 N / m; combined with the study of the heat accumulation effect of laser welding, when the stiffness is below 100 N / m, the risk of material over-deformation exceeds 90%, and when it is above 500 N / m, the probability of thermal stress cracking is less than 1%; finally, the threshold for the low-stiffness region is set at 100 N / m (i.e., the region with stiffness value < 100 N / m requires high-speed scanning to prevent overheating), and the threshold for the high-stiffness region is set at 500 N / m (i.e., the region with stiffness ≥ 500 N / m requires low-speed strengthening fusion). In high-stiffness areas, a slow pulsed laser scanning path (e.g., 0.5 m / s) is set to enhance energy deposition, while in low-stiffness areas, a fast pulsed laser jumping path (e.g., 2 m / s) is set to avoid overheating damage, thus forming a pulsed laser scanning path. Input the yield strength parameter value of each coordinate point of the pad into the laser energy density formula to generate the energy density of each coordinate point. The expression is as follows: ; in, Indicates energy density; Indicates the reference energy density; Indicates the energy adjustment coefficient; Indicates the yield strength of the reference material; It should also be noted that the reference energy density was determined through gradient energy welding experiments. Specifically, standard pad samples (silicon substrate and tin-silver solder thickness of 50 micrometers) were prepared, and an energy density gradient (e.g., 0.1 to 1.0 J / mm) was set on the laser welding platform. 2 Ten solder joints were welded for each energy density value. The weld penetration depth was measured, and the minimum energy density value at which the penetration depth reached 90% of the solder thickness (based on the process implementation settings of ball grid array and other high-density interconnect technologies) was statistically analyzed. The average value of thirty experiments was taken as the baseline energy density (e.g., 0.5 J / mm²). 2 ); The energy adjustment coefficient was calibrated through yield strength gradient welding experiments. Specifically, ten materials with yield strength parameters ranging from 0.5 GPa to 10 GPa were selected to prepare pad samples. Fifty welding experiments were conducted for each material, with a fixed energy density baseline value of 0.5 J / mm² for each experiment. 2Metallographic testing was performed on the critical energy density of the fully penetrated weld joint. The rate of increase in critical energy density corresponding to each 1 GPa increase in yield strength parameter was statistically analyzed. The slope of the linear regression curve of 500 sets of experimental data was taken as the energy adjustment coefficient (e.g., 0.1). The reference material yield strength is set according to the Semiconductor Equipment and Materials Association standard, which stipulates that 1 GPa is uniformly used as the reference material yield strength for semiconductor welding processes. By integrating the slow scanning path and fast jumping path of the pulsed laser with the energy density, a pulsed laser scanning path and energy density scheme are generated.

[0050] S5.3 Calculate the theoretical intensity parameters based on the pulsed laser scanning path and energy density scheme, and collect the fluorescence spectrum data of the welding interface using a spectrometer; It should be noted that the theoretical intensity parameters are calculated based on the pulsed laser scanning path and energy density scheme, and the expression is as follows: ; ; ; in, Indicates the theoretical strength parameter; Indicates the energy intensity conversion factor; Indicates the thermal accumulation correction factor; Indicates the measured strength parameters; Indicates the thermal accumulation reference coefficient; Indicates the scanning speed; Move the drive spectrometer to the current weld point coordinates, align the spectrometer probe vertically with the weld interface and 0.5 mm away from the surface, and collect fluorescence spectral data at a sampling rate of 100 points per second. It should also be noted that, The shear strength refers to the actual shear strength of the weld joint after welding. It is measured by performing a standard shear test on the weld joint using a universal testing machine. Specifically, a weld joint specimen is prepared, fixed to the shear fixture of the testing machine, and a shear load is applied at a constant rate of 1 mm per minute. Load sensor data is recorded in real time until the weld joint breaks. The ratio of the maximum load value before breakage to the cross-sectional area of ​​the weld joint is taken as the shear strength value. The test is repeated thirty times for each group of weld joints, and after removing outliers, the arithmetic mean of the shear strength values ​​is taken as the final value. ; The thermal accumulation reference coefficient is calibrated through laser welding heat conduction experiments. Specifically, a standard solder pad sample (such as a silicon substrate) is irradiated with a laser beam of fixed energy density. The laser beam is moved uniformly at different scanning speeds (0.1 to 5 meters per second). A high-speed infrared thermal imager is used to record the surface temperature change curve of the solder joint. The cooling time from the peak temperature of the molten pool to the critical temperature of solid-state phase transition (1414 degrees Celsius for silicon material) after the laser beam leaves is extracted. The cooling time data corresponding to different scanning speeds are statistically analyzed, and a linear relationship curve between scanning speed and cooling time is fitted. The slope of the linear relationship curve is taken as the thermal accumulation reference coefficient (e.g., for every 1 meter increase in scanning speed, the cooling time per second decreases by 0.4 seconds, so the coefficient is 0.4 s / m).

[0051] S5.4 Combining theoretical intensity parameters and fluorescence spectral data, defect localization analysis is performed through dual-channel image fusion and morphological segmentation to generate a defect coordinate set.

[0052] It should be noted that the minimum and maximum values ​​of all theoretical strength parameters are determined by arranging the values ​​in ascending order. The difference between the minimum and maximum values ​​of the ultimate strength is calculated, and a unit strength difference is generated based on the difference between each theoretical strength parameter value and the minimum value. The ratio of the unit strength difference to the ultimate strength difference is used as the strength scaling factor. The product of the strength scaling factor and 255 (based on the general definition of computer image processing, standard grayscale images use 8-bit depth encoding, and the maximum value of 8 bits is 255) is used as the strength grayscale value of the corresponding theoretical strength parameter. All strength grayscale values ​​are filled according to the row and column positions of the stiffness distribution matrix to form a 500×500 theoretical strength grayscale image as the first channel image. Calculate the wavelength difference between each peak wavelength value and the minimum peak wavelength value in the fluorescence spectrum data. Use the ratio of the wavelength difference to the minimum peak wavelength value as the wavelength scaling factor. Use the product of the wavelength scaling factor and 255 as the wavelength gray value of the corresponding wavelength value. Fill all the wavelength gray values ​​according to the stiffness distribution matrix to generate a 500×500 spectral feature gray image as the second channel image. An image weighted fusion algorithm is used to merge the first and second channel images into a fused image, fusing the intensity grayscale value and wavelength grayscale value into a pixel grayscale value. The arithmetic mean P of all pixel grayscale values ​​is calculated, and pixels with a grayscale value ≥ 1.5P (based on welding defect detection standards, verified by thousands of weld point experiments as the optimal balance between accuracy and efficiency) are identified as potential defect pixels. Adjacent potential defect pixels are automatically aggregated into connected regions. The centroid pixel coordinates of each connected region are calculated, converted to physical coordinates in micrometers using a formula, and stored in the defect coordinate set for output. The expression for calculating the centroid pixel coordinates is... ; ; in, Indicates the centroid pixel coordinates and row coordinates; This represents the column coordinates of the centroid pixels; This represents the total number of centroid pixels within a connected region; Indicates the centroid pixel index, with a value of 1- ; Indicates the first term in the connected region row coordinates of pixels ; Indicates the first term in the connected region column coordinates of pixels ; The expression for converting the centroid pixel coordinates to pad physical space coordinates is: ; ; in, Indicates pad Axis physical space coordinates; Indicates pad Axis physical space coordinates; This indicates that the pad scanning step size is 20 micrometers.

[0053] S6. Generate a laser repair welding path sequence based on the defect coordinate set and perform the repair welding operation. Correct the laser energy parameters and dwell time parameters of each repair welding point until the measured shear strength of the welding interface reaches the preset strength threshold, and output a welding quality certification report.

[0054] S6.1. Initialize the laser energy parameters and dwell time parameters based on the defect coordinate set, and generate the laser repair welding path sequence; It should be noted that the physical space coordinates of the solder pads and the corresponding defect type (crack or micropore) of each coordinate point in the defect coordinate set are extracted; a repair path sequence is generated according to the physical space coordinates of the solder pads: starting from the lower left corner of the solder pad, all defect points are sorted in ascending order of X coordinate (if X is the same, then in ascending order of Y), the distance between adjacent defect points is calculated, and if the distance exceeds 200 micrometers (based on the physical safety distance of the laser processing equipment and the control requirements of the heat-affected zone of the solder joint), an empty jump instruction is inserted to generate an ordered path sequence; Traverse each defect point in the ordered path sequence and initialize the laser energy parameters according to the defect type: for crack defects, use an energy density of 0.8 J / mm². 2 With a residence time parameter of 50ms (set according to crack healing specifications), micropore defects were treated with an energy density of 1.2J / mm². 2 The dwell time parameter is 20ms (set according to the micropore filling specification); the defect coordinates, laser energy parameters and dwell time parameters of each defect point are integrated into a laser repair welding path sequence; It should also be noted that the empty jump command refers to inserting a movement command into the laser controller when the physical distance between adjacent defect points exceeds 200 micrometers. The command content is to raise the laser head to a safe height (default 5 mm) and move in a straight line to the coordinates of the next defect point at the maximum speed (default 2 m / s). During the movement, the laser energy output is turned off. After reaching the defect point, the working height is restored and the energy output is turned on. No welding operation is performed during this process.

[0055] S6.2. Perform the welding operation according to the laser welding path sequence, collect the interface shear strength data of the welding area in real time, and calculate the strength deviation value based on the interface shear strength data and the preset strength threshold. It should be noted that the laser driver loads the laser welding path sequence and performs welding operations point by point: for ordinary command points, the laser energy parameters and dwell time parameters are used to illuminate the coordinate position of the welding point; for empty jump command points, the laser head is raised to a safe height of 5 mm and moves to the next welding point at a speed of 2 m / s. After each weld point is repaired, the universal testing machine is immediately driven to perform an interface shear test. The shear fixture applies shear force at a constant rate of 1 mm / min until the weld point breaks. The maximum load value is recorded and the ratio of it to the standard cross-sectional area of ​​the solder pad is used as the interface shear strength data. Select the strength threshold according to the material region to which the solder joint belongs (700 MPa for the chip region and 300 MPa for the substrate region, based on the definition of mechanical stress test for semiconductor devices), and use the difference between the interface shear strength data and the strength threshold as the strength deviation value.

[0056] S6.3 If the strength deviation value exceeds the limit, correct the laser energy parameters and dwell time parameters, repeat the repair welding operation until the strength comparison is completed, and output the welding quality certification report.

[0057] It should be noted that if the strength deviation value is less than the strength threshold (set by engineering consensus based on the critical point of fatigue life of welded structure and the risk of material failure, such as the strength deviation value of chip area < -70 MPa or the strength deviation value of substrate area < -30 MPa), it is judged as exceeding the limit. The laser energy correction is the product of the current laser energy parameter and the deviation proportional gain coefficient (defined based on the proportional gain stability criterion in control theory, such as 0.1). The dwell time correction is the product of the current dwell time parameter and the deviation proportional gain coefficient, generating the corrected laser energy parameters and dwell time parameters. The laser energy parameters and dwell time parameters of the corresponding weld points in the laser welding path sequence are updated, and the laser is re-driven to perform the corresponding weld point welding operation. After welding, the interface shear strength data is collected again to calculate the strength deviation value. If it still exceeds the limit, it is cyclically corrected until the deviation value is ≥ -10% of the strength threshold. When the strength deviation value does not exceed the limit, the final interface shear strength data is recorded. After traversing all weld points to complete the correction, the weld point coordinates, final interface shear strength data, correction times, and certification status (qualified and over-limit repair) are summarized to generate a welding quality certification report.

[0058] It should also be noted that when the welding strength is less than 10% of the strength threshold, the metal fatigue crack propagation rate enters the accelerated stage.

[0059] This embodiment also provides a computer device applicable to the method of welding semiconductor integrated circuit chips, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method of welding semiconductor integrated circuit chips as proposed in the above embodiment.

[0060] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0061] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for bonding semiconductor integrated circuit chips as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0062] In summary, this invention achieves adaptive adjustment of the oxidation activity of the solder pad surface and the solution ratio by: dynamically matching the concentration ratio of the two-phase aqueous solution through the analysis of the oxidation state encoding matrix and planning the zoned spraying path; accurately eliminating the poor soldering and metal compound segregation caused by uneven wetting; and significantly reducing the risk of splash contamination by avoiding low viscosity areas through the path. Furthermore, it constructs an atomic layer deposition kinetic model for closed-loop control of surface diffusion and real-time thickness feedback, dynamically calculates the deposition rate and residence time parameters, and continuously controls the deposition process online to eliminate liquid film fluctuations and lattice distortion, thus avoiding efficiency losses due to interrupted detection.

[0063] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for welding semiconductor integrated circuit chips, characterized in that: include, Data on oxide layer thickness and substrate microstructure are collected and input into a pre-trained convolutional neural network to generate an oxide state encoding matrix. The optimal concentration ratio of polyethylene glycol and dextran aqueous two-phase solution is calculated using a gradient ratio algorithm. Then, a zoned path planning is performed on the optimal concentration ratio to generate a zoned spraying path instruction set. The partitioned spraying path instruction set is transmitted to the micro-nozzle array to perform dual-phase spraying, while the liquid film thickness distribution data of the spraying area is collected to generate a uniformity feedback dataset. An atomic layer deposition dynamics model is constructed, and the uniformity feedback dataset is input into the atomic layer deposition dynamics model to generate atomic layer control signals. The atomic film thickness parameters are then calculated based on the atomic layer control signals. The system calls upon a pre-stored material yield strength database and combines it with atomic film thickness parameters to generate a pulsed laser scanning path and energy density scheme. Simultaneously, it calculates theoretical strength parameters, collects fluorescence spectral data of the welding interface, and generates a defect coordinate set through defect location analysis based on the theoretical strength parameters. Based on the defect coordinate set, a laser repair welding path sequence is generated and the repair welding operation is performed. The laser energy parameters and dwell time parameters of each repair welding point are corrected until the measured shear strength of the weld interface reaches the preset strength threshold, and a welding quality certification report is output.

2. The method for welding semiconductor integrated circuit chips as described in claim 1, characterized in that: The specific steps for generating the oxidation state encoding matrix are as follows: The collected oxide layer thickness data were normalized, and the collected substrate microstructure data were subjected to Gaussian filtering for noise reduction. The normalized oxide layer thickness data and the noise-reduced substrate microstructure data are combined into a dual-channel input matrix and then input into a pre-trained convolutional neural network. Features are extracted through convolution and max pooling operations of a convolutional neural network, and then mapped through a fully connected layer to generate an oxidized state encoding matrix.

3. The method for welding semiconductor integrated circuit chips as described in claim 2, characterized in that: The specific steps for calculating the optimal concentration ratio of polyethylene glycol and dextran aqueous two-phase solution are as follows. Analyze the spatial dimension features of the oxidation state encoding matrix and extract the oxidation activity level value corresponding to each matrix structure; Based on the oxidation activity level value, the optimal concentration ratio of polyethylene glycol and dextran aqueous two-phase solution was calculated using a gradient ratio algorithm.

4. The method for welding semiconductor integrated circuit chips as described in claim 3, characterized in that: The specific steps for generating the partitioned spraying path instruction set are as follows. The optimal concentration ratio is mapped to the physical space coordinates of the pads corresponding to the oxidation state encoding matrix to generate a heat map of the concentration ratio distribution. Based on the concentration ratio distribution heatmap, the optimal movement path of the micro-nozzle is generated using a path planning algorithm.

5. The method for welding semiconductor integrated circuit chips as described in claim 4, characterized in that: The specific steps for generating the uniformity feedback dataset are as follows: The partitioned spraying path instruction set is parsed into a control signal sequence for the micro-nozzle array, which is then driven to move to the target coordinate position according to the control signal sequence. The two-phase spraying deposition of polyethylene glycol solution and dextran solution is then executed sequentially to generate the scanning spraying area. The spraying area is scanned by a white light interferometer to collect liquid film thickness distribution data, and median filtering noise reduction is performed on the liquid film thickness distribution data. The denoised liquid film thickness distribution data is mapped to a thickness distribution matrix according to the physical space coordinates of the pads, and the standard deviation of the thickness distribution matrix is ​​calculated to generate a uniformity feedback dataset.

6. The method for welding semiconductor integrated circuit chips as described in claim 5, characterized in that: The specific steps for constructing the atomic layer deposition dynamics model are as follows. An input layer is constructed based on a uniformity feedback dataset, a core dynamics layer is constructed based on a surface diffusion equation and a thickness feedback control term, a signal generation layer is constructed based on a PID control equation, and an output layer is constructed based on thickness integral calculation. An atomic layer deposition dynamics model is constructed based on the input layer, core dynamics layer, signal generation layer, and output layer.

7. The method for welding semiconductor integrated circuit chips as described in claim 6, characterized in that: The specific steps for generating the atomic film thickness parameters are as follows. The uniformity feedback dataset is input into the atomic layer deposition dynamics model, and atomic layer control signals are generated through serial hierarchical physical calculations. The atomic film thickness parameter is generated by integral calculation based on the deposition rate parameter and electron signal duration in the atomic layer control signal.

8. The method for welding semiconductor integrated circuit chips as described in claim 7, characterized in that: The specific steps for generating the pulsed laser scanning path and energy density scheme are as follows: The pre-stored material yield strength database is called to extract the yield strength parameters of the chip material and the substrate material. Combined with the atomic film thickness parameter, the stiffness distribution matrix of the welding interface is calculated. Based on the stiffness distribution matrix and yield strength parameters, a pulsed laser scanning path and energy density scheme are generated through path optimization and energy allocation.

9. The method for welding semiconductor integrated circuit chips as described in claim 8, characterized in that: The specific steps for generating the defect coordinate set are as follows: Theoretical intensity parameters were calculated based on the pulsed laser scanning path and energy density scheme, and fluorescence spectral data of the welding interface were collected. By combining theoretical intensity parameters and fluorescence spectral data, defect localization analysis is performed through dual-channel image fusion and morphological segmentation to generate a defect coordinate set.

10. The method for welding semiconductor integrated circuit chips as described in claim 9, characterized in that: The specific steps for generating the welding quality certification report are as follows. Laser energy parameters and dwell time parameters are initialized based on the defect coordinate set to generate a laser repair welding path sequence; The welding operation is performed according to the laser welding path sequence. The interface shear strength data of the welding area is collected in real time. The strength deviation value is calculated based on the interface shear strength data and the preset strength threshold. If the strength deviation exceeds the limit, the laser energy parameters and dwell time parameters are corrected, and the repair welding operation is repeated until the strength comparison is completed, until the strength deviation does not exceed the limit, and a welding quality certification report is output.

Images