Flip chip underfill optimization using screen printing technology

Abstract

The present application relates to the technical field of semiconductor chip packaging, and particularly relates to a flip chip underfill optimization method using steel mesh printing technology. The method uses a prediction method to simulate the entire underfill process according to various parameters in the early stage of the process, and then judges whether the current process will have a void defect according to the prediction result. In the dynamic prediction model of the underfill process, two types of loss functions are introduced, which respectively represent the result loss between the continuous prediction process of the model and the training data, and the physical consistency loss formed by the authenticity and reliability of the continuous prediction process. The dynamic prediction model of the underfill process constructed by the present application can simulate the dynamic process of the chip surface pressure field, and the engineers can determine the steel mesh area with process problems according to the continuous prediction process of the model in the process optimization stage. Furthermore, the model can be used to simulate the underfill result after optimization after the process optimization, thereby forming a reliable and effective process optimization process.

Description

Technical Field

[0001] This invention relates to the field of semiconductor chip packaging technology, and more specifically to a flip chip underfill optimization method utilizing stencil printing technology. Background Technology

[0002] In flip-chip packaging, epoxy resin-based underfill material is used to fill the gap between the chip and the substrate to mitigate thermal expansion. This filling process relies on the material's capillary action, flowing and curing under heating conditions. As chip size increases, the underfill material must flow through longer and more complex micropaths composed of thousands of solder balls, leading to increased flow resistance. Simultaneously, the heating environment of reflow soldering triggers a chemical curing reaction in the underfill material, causing its viscosity to irreversibly increase over time, further increasing flow resistance. This process constitutes a direct competition between two dynamic processes: capillary-driven flow and thermo-chemical curing. When the flow resistance caused by increased viscosity eventually exceeds the capillary driving force, flow prematurely stops in certain areas of the chip. If these areas are not fully filled at this point, voids will form. Voids become stress concentration points, significantly reducing the long-term reliability of the package.

[0003] Current technologies for detecting and preventing voids typically employ online real-time monitoring and analysis. This involves using real-time monitoring data to analyze factors such as temperature or filler morphology to determine the occurrence of voids. However, while this method achieves online real-time monitoring, the filling process has already begun when the void occurs, resulting in material waste. Furthermore, to optimize the process, engineers continuously modify process parameters and conduct new experimental fillings, leading to long development cycles and high material consumption. Summary of the Invention

[0004] To address the existing technical problems of material waste caused by real-time monitoring of void defects and the high material consumption required for process optimization, the present invention aims to provide a flip-chip underfill optimization method utilizing stencil printing technology. The specific technical solution adopted is as follows:

[0005] This invention proposes a flip chip underfill optimization method using stencil printing technology, the method comprising:

[0006] The solder ball flowability map on the chip surface is obtained based on the chip structure; the initial pressure change map of the colloid is obtained based on the rate of change of the height of the colloid in space after printing and before curing; the solder ball flowability map, the initial pressure change map of the colloid, and the curing reaction rate of the colloid are used as the initial conditions of the process.

[0007] A dynamic prediction model for the bottom filling process is trained based on the initial process conditions in the historical production flow and the final cavity distribution map. The input of the dynamic prediction model for the bottom filling process is the initial process conditions, and the output is the predicted effective pressure field. The generation process of the predicted effective pressure field is a continuous prediction based on the generation result at the previous moment. The loss function of the dynamic prediction model for the bottom filling process includes result consistency loss and physical consistency loss. The predicted cavity distribution map is obtained by using data from the predicted effective pressure field, and the result consistency loss is obtained based on the difference between the predicted cavity distribution map and the final cavity distribution map. The difference between the change law of the predicted effective pressure field in the continuous prediction process and the preset physical law constitutes the physical consistency loss.

[0008] The initial process conditions during the real-time bottom filling process are input into the dynamic prediction model of the bottom filling process, and the predicted effective pressure field at the final moment is output, and the predicted cavity distribution map is obtained; the early warning signal is fed back based on the predicted cavity distribution map, and the process is optimized based on the continuous prediction process.

[0009] Furthermore, the method for obtaining the solder ball flowability diagram includes:

[0010] A two-dimensional mesh is constructed for the bottom region of the chip based on the chip design file. On the two-dimensional mesh, the element values ​​of the mesh cells occupied by the solder ball entities are set to 0, and the element values ​​of the mesh cells initially applied to the bottom filler are set to 1. The steady-state Laplace equation is solved in the two-dimensional mesh to obtain the element values ​​of each mesh cell, thus obtaining the solder ball flow capability map.

[0011] Furthermore, the method for obtaining the initial pressure change map of the colloid includes:

[0012] A height distribution map of the chip surface is obtained, and the gradient at each location point on the height distribution map is obtained. The gradient at each location point constitutes the initial pressure change map of the colloid.

[0013] Furthermore, the dynamic prediction model for the bottom filling process is a U-Net structure.

[0014] Furthermore, the method for obtaining the predicted cavity distribution map includes:

[0015] The predicted effective pressure field corresponding to the last time of the continuous prediction is taken as the pressure field to be analyzed; the positions in the pressure field to be analyzed where the element value is less than the preset pressure threshold are taken as the void positions, and the predicted void distribution map is obtained.

[0016] Furthermore, the consistency loss is the Dice loss between the predicted cavity distribution map and the final cavity distribution map.

[0017] Furthermore, the difference between the predicted effective pressure field's variation pattern during the continuous prediction process and the preset physical laws constitutes the physical consistency loss, including:

[0018] For any element in the predicted effective pressure field, the preset physical law is expressed by the formula: ;in This represents the element at position (i,j) in the predicted effective pressure field at time t during the continuous prediction process. The rate of change over time; This represents the element value at position (i,j) in the solder ball flowability diagram; The curing reaction rate C of the colloid is a function of time during the underfilling process; For elements Gradient in the predicted effective pressure field;

[0019] The physical consistency loss is expressed by the formula:

[0020] ;in The physical consistency loss is represented by T; T represents the total number of time points in the persistent prediction process. N This represents the total number of element points included in the predicted effective pressure field. This is the L2 norm operator.

[0021] Furthermore, the dynamic prediction model for the bottom filling process uses the weighted sum of the result consistency loss and the physical consistency loss as the final loss.

[0022] Furthermore, the step of feeding back the early warning signal based on the predicted cavity distribution map includes:

[0023] Obtain the area ratio of void regions in the predicted void distribution map; if the area ratio is greater than a preset warning threshold, a warning signal is fed back.

[0024] Furthermore, the process optimization based on the continuous prediction process includes:

[0025] The flow bottleneck region is determined based on the continuous prediction process. The structure of the flow bottleneck region is modified and new initial process conditions are determined. A new predicted void distribution map is obtained based on the new initial process conditions. The predicted void distribution maps before and after modification are compared to determine the optimization effect.

[0026] The present invention has the following beneficial effects:

[0027] To avoid material waste during real-time monitoring of void defects, this invention employs a predictive approach to simulate the entire underfill process based on initial process parameters. The prediction results then determine whether void defects will exist in the current process. Furthermore, the dynamic prediction model for the underfill process constructed in this invention can simulate and predict the dynamic process of the pressure field on the chip surface. During the process optimization phase, engineers can identify stencil areas with process problems based on the model's continuous prediction process. After process optimization, the model can be further used to simulate the optimized underfill results, thus forming a reliable and effective process optimization process. This invention introduces two types of loss functions into the dynamic prediction model for the underfill process. These two loss functions represent the outcome loss between the model's continuous prediction process and the training data, and the physical consistency loss resulting from the realism and reliability of the continuous prediction process. By setting the loss functions, the dynamic prediction model for the underfill process can effectively learn the characteristics of the underfill process while adhering to physical laws, thereby achieving realistic and effective process simulation. Attached Figure Description

[0028] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0029] Figure 1 This is a flowchart of a flip chip underfill optimization method using stencil printing technology, provided as an embodiment of the present invention. Detailed Implementation

[0030] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a flip chip underfill optimization method utilizing stencil printing technology proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0032] The following description, in conjunction with the accompanying drawings, details a specific scheme for an optimized underfill method for flip-chips using stencil printing technology provided by the present invention.

[0033] Please see Figure 1The diagram illustrates a flowchart of a flip chip underfill optimization method using stencil printing technology, according to an embodiment of the present invention. The method includes:

[0034] Step S1: Obtain the solder ball flowability map on the chip surface based on the chip structure; obtain the initial pressure change map of the colloid based on the rate of change of the height of the colloid in space after printing and before curing; the solder ball flowability map, the initial pressure change map of the colloid, and the curing reaction rate of the colloid are used as the initial conditions of the process.

[0035] The embodiments of the present invention aim to simulate the entire underfill process based on the initial process conditions. Therefore, it is first necessary to quantify the various state characteristics of the chip before the underfill process.

[0036] First, it's necessary to quantify the complexity of the microscopic flow channels formed by the fixed array of solder balls on the bottom of the chip. Because the flow path of the bottom filler is strictly limited by these solder ball obstacles, the inherent flow permeability of each region of the chip must be calculated beforehand. The fixed positions and physical characteristics of these solder balls are recorded in the chip design documents, so the solder ball flow capability map of the chip surface can be directly obtained from the chip structure in the design documents. The size of the solder ball flow capability map corresponds to the bottom region of the chip. Larger element values ​​in the map indicate a more open distribution of solder balls at the corresponding location, representing a region with smooth flow; conversely, smaller element values ​​indicate a denser distribution of solder balls at the corresponding location, representing a region with difficult flow. It should be noted that for a pre-designed chip, the solder ball flow capability map is a fixed, static feature, a parameter characteristic fixed only in the initial stages of the process, and only needs to be obtained once.

[0037] Furthermore, this invention also requires determining the initial driving force for the filler flow. In this embodiment, an online 3D solder paste inspection device can be used to scan the bottom filler. After printing and before curing, the actual colloid on the substrate is optically scanned to obtain a three-dimensional colloid height map. That is, the three-dimensional colloid height map can reflect the height of the colloid at various locations on the substrate surface. Since surface tension is the main driving force for capillary flow, its effect is directly related to the rate of change of the colloid surface height. Therefore, this embodiment can obtain the initial pressure change map of the colloid based on the rate of change of the colloid height in space. Similar to the solder ball flow capability map, the initial pressure change map of the colloid also corresponds to the bottom region of the chip.

[0038] Finally, in this embodiment of the invention, the curing reaction rate information of the bottom filler is determined by consulting parameter data, etc., and this characteristic is related to the material properties of the bottom filler. The solder ball flowability diagram, the initial pressure change diagram of the colloid, and the curing reaction rate of the colloid are used as the initial process conditions.

[0039] Preferably, in this embodiment of the invention, the method for obtaining the solder ball flowability diagram includes:

[0040] A two-dimensional mesh is constructed for the bottom region of the chip based on the chip design documents. On this mesh, the element values ​​of the mesh cells occupied by the solder balls are set to 0, and the element values ​​of the mesh cells initially used for the bottom filler are set to 1. The steady-state Laplace equation is solved in the two-dimensional mesh to obtain the element values ​​of each mesh cell, thus yielding the solder ball flowability map. The steady-state Laplace equation is in the form of… , The equation representing the solder ball flow capability describes the potential distribution in a passive field and effectively characterizes the potential flowability of fluid through porous media. In this embodiment of the invention, the equation can be solved using the finite difference method to obtain the potential value at each grid point, i.e., each location point, within the chip region.

[0041] The solder ball flowability map can also be viewed as a two-dimensional matrix. Each element in the matrix corresponds to a grid point as described above, which is also a location point in the bottom region of the chip. The position of the element value, i.e. the potential value, in the solder ball flowability map is 0, which is the absolute physical boundary that the flow cannot enter; the position of the element value is 1, which is the source of the potential energy of the flow.

[0042] Other numerical values ​​are between 0 and 1. They do not directly represent flow velocity or pressure. Instead, they quantify the topological accessibility of flow at a corresponding location, which is determined by a fixed set structure. The closer the element value is to 0, the more tightly the location is surrounded by solder balls in the topology, indicating a difficult flow region. The closer the element value is to 1, the more open the corresponding region is, indicating a smooth flow region.

[0043] Preferably, in this embodiment of the invention, gradient information is selected as the rate of change information, i.e., the method for obtaining the colloid initial pressure change map includes:

[0044] A height distribution map of the chip surface is obtained, and the gradient at each location point on the height distribution map is obtained. The gradient at each location point constitutes the initial pressure change map of the colloid. It should be noted that because the gradient has both direction and value, the resulting initial pressure change map of the colloid is a two-dimensional vector field. In this vector field, the vector direction at each location point represents the direction of the strongest initial flow tendency of the colloid at that location due to the surface tension gradient, and the magnitude of the vector quantifies the magnitude of this initial driving force. Therefore, the initial pressure change map of the colloid directly reflects the actual quality of the current stencil printing process and its direct impact on the initial flow behavior. That is, the initial pressure change map of the colloid changes dynamically with each printing, and the printing quality determines the result of the initial pressure change map of the colloid.

[0045] Step S2: Train the dynamic prediction model for the bottom filling process based on the initial process conditions in the historical production flow and the final void distribution map; the input of the dynamic prediction model for the bottom filling process is the initial process conditions, and the output is the predicted effective pressure field; the generation process of the predicted effective pressure field is a continuous prediction based on the generation results of the previous moment in time; the loss function of the dynamic prediction model for the bottom filling process includes result consistency loss and physical consistency loss; obtain the predicted void distribution map through the data in the predicted effective pressure field, and obtain the result consistency loss based on the difference between the predicted void distribution map and the final void distribution map; the difference between the change law of the predicted effective pressure field in the continuous prediction process and the preset physical law constitutes the physical consistency loss.

[0046] This invention utilizes historical data from all historical production processes stored in a statistical database. This historical data includes the initial process conditions for each bottom-filling process and the final void distribution map after bottom-filling. This data is then used to obtain training data for a dynamic prediction model of the bottom-filling process. It should be noted that the historical data should be selected from chips of the same type used on the production line; that is, each type of chip can correspond to one dynamic prediction model for the bottom-filling process. In the final void distribution map, this invention sets the void area to 1 and other areas to 0, resulting in a binary distribution map corresponding to the coordinates of the chip's bottom area.

[0047] Preferably, the dynamic prediction model for the underfilling process in this embodiment of the invention adopts a U-Net structure. The reason for choosing U-Net is that its symmetrical structure and skip connection design enable it to capture contextual information simultaneously through downsampling encoding and capture precise local details through upsampling decoding and skip connections when processing image-to-image conversion tasks. This allows the model to meet the requirement of predicting dynamic spatiotemporal fields based on static initial condition maps.

[0048] In this embodiment of the invention, the input of the dynamic prediction model for the underfill process is the initial process conditions, and the output is the predicted effective pressure field. The generation process of the predicted effective pressure field is a continuous prediction based on the generation result of the previous time step. The input of the model consists of three independent channels for the initial process conditions, and the output is a three-dimensional spatiotemporal data field. That is, the predicted effective pressure field is a three-dimensional spatiotemporal data field that is related to the two-dimensional coordinates of the chip area and is time-dependent. At each time point t, the model receives the predicted effective pressure field output at migration time t-1, as well as the constant initial conditions solder ball flow capability map and the colloid initial pressure change map, and then predicts the predicted effective pressure field at the current time point t. This continuous prediction process can naturally simulate the time evolution characteristics of the physical process.

[0049] Since the predicted effective pressure field proposed in this embodiment of the invention is simulated, it is not an intermediate process that can be directly monitored by monitoring equipment in historical production processes. Therefore, traditional supervised learning methods cannot be used for model training. Thus, this embodiment of the invention employs two loss functions for model training, allowing the model to learn the correct physical simulation process by modifying its own parameters and altering the prediction results, thus constrained by the loss functions. The two loss functions are result consistency loss and physical consistency loss.

[0050] The purpose of the consistency loss is to ensure that the model's predictions match the actual detection results. Since the final result obtained from the training data is the final void distribution map, the calculation of the consistency loss involves determining the predicted effective pressure field corresponding to the last time point. Because the values ​​at each location in the predicted effective pressure field represent the flow pressure of the colloid, a smaller value indicates that the colloid has stopped flowing at that location, forming voids. Therefore, the corresponding predicted void distribution map can be obtained from the data in the predicted effective pressure field corresponding to the last time point. It should be noted that, similar to the final void distribution map, the predicted void distribution map is also a binary image. Therefore, based on the same type of predicted void distribution map and the final void distribution map, the consistency loss can be obtained through the difference in void distribution between the two.

[0051] Preferably, in this embodiment of the invention, the method for obtaining the predicted cavity distribution map includes:

[0052] The predicted effective pressure field at the last time point of the continuous prediction is taken as the pressure field to be analyzed; the positions in the pressure field whose element values ​​are less than a preset pressure threshold are taken as void locations, thus obtaining the predicted void distribution map. The preset pressure threshold can be set based on the detection accuracy of voids in the actual scenario; a lower pressure threshold can be set for high accuracy requirements, and a higher pressure threshold can be set for low accuracy requirements. In this embodiment of the invention, the pressure threshold is set to the data corresponding to the first 10% of positions after the obtained pressure data is sorted in ascending order.

[0053] Preferably, in this embodiment of the invention, the result consistency loss is the Dice loss between the predicted hole distribution map and the final hole distribution map. That is, the hole distribution map is treated as a set, and the result consistency loss is obtained by comparing the intersection and union of the two sets. The Dice loss is chosen instead of pixel-level cross-entropy loss because hole defects are typically sparse across the entire chip area, and the Dice loss performs better for this type of class imbalance problem, more effectively guiding the model to learn the shape and location of holes.

[0054] Furthermore, the purpose of physical consistency loss is to limit the variation in the model's continuous prediction process and prevent the continuous prediction process from deviating from physical laws. Therefore, the difference between the variation of the predicted effective pressure field in the continuous prediction process and the preset physical laws constitutes the physical consistency loss.

[0055] Preferably, the physical laws in this embodiment of the invention are based on Darcy's law and the law of conservation of mass. For any element in the predicted effective pressure field, the preset physical laws are expressed by the following formula: ;in This represents the element at position (i,j) in the predicted effective pressure field at time t during the continuous prediction process. The rate of change over time; This represents the element value at position (i,j) in the solder ball flowability diagram; The curing reaction rate C of the colloid is a function of time during the underfilling process; For elements The gradient in the predicted effective pressure field.

[0056] It should be noted that the above formula is a simplified partial differential equation, in which... The curing reaction rate information obtained in step S1 can be considered as prior knowledge. This embodiment of the invention expects the model's persistent prediction to satisfy the above equation, i.e., the equation holds. Therefore, for the generated predicted effective pressure field, its physical consistency loss is expressed by the formula:

[0057] ;in The physical consistency loss is represented by T; T represents the total number of time points in the persistent prediction process. N This represents the total number of element points included in the predicted effective pressure field. This is the L2 norm operator.

[0058] The physical consistency loss is the sum of the L2 norms of the residuals of the partial differential equations at various locations on the predicted effective pressure field. The residuals at each location characterize the difference between the generated results and the expected physical laws. This physical consistency loss then constrains the model's generated results, forcing the model to change its parameters and ultimately learn the true and effective characteristics of pressure field changes.

[0059] Preferably, in this embodiment of the invention, the dynamic prediction model for the underfilling process uses the weighted sum of the result consistency loss and the physical consistency loss as the final loss. The weights of the two loss functions aim to balance the accuracy of the final result with the physical plausibility; in this embodiment, both weights can be set to 0.5. Based on the final loss, the Adam optimizer is used to minimize it. The model parameters are continuously adjusted until the final loss reaches its minimum value, indicating that the model can generate a dynamic pressure field that both conforms to physical laws and accurately predicts the final void. It should be noted that other specific training processes of the model are well-known to those skilled in the art and will not be described in detail here.

[0060] After the dynamic prediction model for the bottom filling process is trained and validated offline, it can be deployed to the production line for rapid inference. For new substrates delivered from the production line, the system can extract the initial process conditions from step S1 and input them into the model. The model will then iteratively, continuously, and rapidly deduce the predicted effective pressure field change process from the initial moment to the final moment. The neural network computation is highly parallel, and the network inference process in this embodiment does not require backpropagation; the entire simulation process can be completed in milliseconds.

[0061] Step S3: Input the initial process conditions of the real-time bottom filling process into the dynamic prediction model of the bottom filling process, output the predicted effective pressure field at the final moment, and obtain the predicted cavity distribution map; feed back the early warning signal based on the predicted cavity distribution map, and optimize the process based on the continuous prediction process.

[0062] As described in step S2, for the real-time underfilling process, a dynamic prediction model for the underfilling process can be used to simulate the change in the pressure field during underfilling. Furthermore, based on the same method, a predicted void distribution map corresponding to the predicted effective pressure field at the final simulated moment can be obtained. In the predicted void distribution map, the more void areas there are, the lower the yield under the process conditions, thus requiring a warning signal. Workers can use the warning signal to intercept the chip substrates in the production line, preventing direct underfilling and avoiding material waste in subsequent processes.

[0063] Preferably, in this embodiment of the invention, the feedback of the early warning signal based on the predicted cavity distribution map includes:

[0064] The area ratio of void regions in the predicted void distribution map is obtained; if the area ratio is greater than a preset warning threshold, a warning signal is fed back. In this embodiment of the invention, the warning threshold is set to 0.1, that is, a warning signal is fed back when the area ratio of void regions is greater than 0.1.

[0065] Because the dynamic prediction model for the underfill process provided in this embodiment of the invention can simulate the dynamic process of underfilling, if an early warning signal is fed back during the above process, the staff can optimize the process based on the continuous prediction process of the model. By retrieving the predicted effective pressure field generated at each moment during the continuous prediction process, the staff can observe the dynamic changes of the pressure gradient in various regions inside the chip during the simulated filling process, thereby identifying the problem area and performing targeted process optimization.

[0066] Preferably, in this embodiment of the invention, the flow bottleneck region can be determined based on a continuous prediction process. The flow bottleneck region is the area where the pressure value decays too early and too quickly during the evolution of the predicted effective pressure field. Due to its abnormal decay, the surrounding pressure gradient approaches 0 before the final moment of the underfilling process. Engineers can make targeted modifications based on the located flow bottleneck region. For example, if the diagnosis finds that the pressure in the central region of the chip has dissipated in the middle of the filling process, a direct optimization strategy is to add a new adhesive dot printing opening upstream of the bottleneck region by modifying the stencil CAD file.

[0067] It should be noted that the specific location of the flow bottleneck region in this embodiment of the invention needs to be determined by the engineer. In other embodiments of the invention, the flow bottleneck region can also be determined by setting various threshold conditions. Further details and limitations will not be elaborated upon.

[0068] By modifying the structure of the flow bottleneck region and determining new initial process conditions, engineers can eliminate the need for physical pattern making or experimentation. They simply use the modified initial process conditions as input to obtain a new predicted void distribution map. By comparing the predicted void distribution maps before and after the modification, the optimization effect can be determined. Specifically, if the void area in the predicted void distribution map is significantly reduced after the modification, it indicates effective optimization, and the greater the reduction, the better the optimization effect. Through this simulation and comparison method, engineers can repeat this closed-loop process of diagnosis-modification-virtual verification, using minimizing the void area distribution as the single, clear optimization objective. This allows for efficient exploration of various design possibilities on a computer until a steel mesh design scheme with the theoretically lowest void risk is found.

[0069] In summary, this invention employs a predictive approach to simulate the entire underfill process based on various parameters at the initial stage of the process, and then determines whether void defects will exist in the current process based on the prediction results. Two types of loss functions are introduced into the dynamic prediction model for the underfill process. These two loss functions represent the outcome loss between the model's continuous prediction process and the training data, and the physical consistency loss resulting from the reliability of the continuous prediction process. The dynamic prediction model for the underfill process constructed in this invention can simulate and predict the dynamic process of the pressure field on the chip surface. During the process optimization stage, engineers can identify stencil areas with process problems based on the model's continuous prediction process. Furthermore, after process optimization, the model can be further used to simulate the optimized underfill results, thus forming a reliable and effective process optimization process.

[0070] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0071] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A flip chip underfill optimization method utilizing stencil printing technology, characterized in that, The method includes: The solder ball flowability map on the chip surface is obtained based on the chip structure; the initial pressure change map of the colloid is obtained based on the rate of change of the height of the colloid in space after printing and before curing; the solder ball flowability map, the initial pressure change map of the colloid, and the curing reaction rate of the colloid are used as the initial conditions of the process. A dynamic prediction model for the bottom filling process is trained based on the initial process conditions in the historical production flow and the final cavity distribution map. The input of the dynamic prediction model for the bottom filling process is the initial process conditions, and the output is the predicted effective pressure field. The generation process of the predicted effective pressure field is a continuous prediction based on the generation result at the previous moment. The loss function of the dynamic prediction model for the bottom filling process includes result consistency loss and physical consistency loss. The predicted cavity distribution map is obtained by using data from the predicted effective pressure field, and the result consistency loss is obtained based on the difference between the predicted cavity distribution map and the final cavity distribution map. The difference between the change law of the predicted effective pressure field in the continuous prediction process and the preset physical law constitutes the physical consistency loss. The initial process conditions during the real-time bottom filling process are input into the dynamic prediction model of the bottom filling process, and the predicted effective pressure field at the final moment is output, and the predicted cavity distribution map is obtained; the early warning signal is fed back based on the predicted cavity distribution map, and the process is optimized based on the continuous prediction process. The difference between the predicted effective pressure field's variation during the continuous prediction process and the preset physical laws constitutes the physical consistency loss, including: For any element in the predicted effective pressure field, the preset physical law is expressed by the formula: ;in This represents the element at position (i,j) in the predicted effective pressure field at time t during the continuous prediction process. The rate of change over time; This represents the element value at position (i,j) in the solder ball flowability diagram; The curing reaction rate C of the colloid is a function of time during the underfilling process; For elements Gradient in the predicted effective pressure field; The physical consistency loss is expressed by the formula: ;in The physical consistency loss is represented by T; T represents the total number of time points in the persistent prediction process. N This represents the total number of element points included in the predicted effective pressure field. This is the L2 norm operator.

2. The flip chip underfill optimization method using stencil printing technology according to claim 1, characterized in that, The method for obtaining the solder ball flowability diagram includes: A two-dimensional mesh is constructed for the bottom region of the chip based on the chip design file. On the two-dimensional mesh, the element values ​​of the mesh cells occupied by the solder ball entities are set to 0, and the element values ​​of the mesh cells initially applied to the bottom filler are set to 1. The steady-state Laplace equation is solved in the two-dimensional mesh to obtain the element values ​​of each mesh cell, thus obtaining the solder ball flow capability map.

3. The flip chip underfill optimization method using stencil printing technology according to claim 1, characterized in that, The method for obtaining the initial pressure change diagram of the colloid includes: A height distribution map of the chip surface is obtained, and the gradient at each location point on the height distribution map is obtained. The gradient at each location point constitutes the initial pressure change map of the colloid.

4. The flip chip underfill optimization method using stencil printing technology according to claim 1, characterized in that, The dynamic prediction model for the bottom filling process is a U-Net structure.

5. The flip chip underfill optimization method using stencil printing technology according to claim 1, characterized in that, The method for obtaining the predicted cavity distribution map includes: The predicted effective pressure field corresponding to the last time of the continuous prediction is taken as the pressure field to be analyzed; the positions in the pressure field to be analyzed where the element value is less than the preset pressure threshold are taken as the void positions, and the predicted void distribution map is obtained.

6. The flip chip underfill optimization method using stencil printing technology according to claim 1, characterized in that, The consistency loss is the Dice loss between the predicted cavity distribution map and the final cavity distribution map.

7. The flip chip underfill optimization method using stencil printing technology according to claim 1, characterized in that, The dynamic prediction model for the bottom filling process uses the weighted sum of the result consistency loss and the physical consistency loss as the final loss.

8. The flip chip underfill optimization method using stencil printing technology according to claim 1, characterized in that, The method of feeding back early warning signals based on the predicted cavity distribution map includes: Obtain the area ratio of void regions in the predicted void distribution map; if the area ratio is greater than a preset warning threshold, a warning signal is fed back.

9. The flip chip underfill optimization method using stencil printing technology according to claim 1, characterized in that, The process optimization based on the continuous prediction process includes: The flow bottleneck region is determined based on the continuous prediction process. The structure of the flow bottleneck region is modified and new initial process conditions are determined. A new predicted void distribution map is obtained based on the new initial process conditions. The predicted void distribution maps before and after modification are compared to determine the optimization effect.

Images