Method and device for planning and executing a process for applying a heat-conducting paste using machine learning methods

EP4469927A1Pending Publication Date: 2024-12-04ROBERT BOSCH GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
EP2023701904
Authority / Receiving Office
EP · EP
Patent Type
Applications
Current Assignee / Owner
Priority Date
2022-01-28
Filing Date
2023-01-23
Publication Date
2024-12-04

AI Technical Summary

Technical Problem

The application of thermally conductive paste on cooling surfaces of heat-producing components is challenging due to increased degrees of freedom, which complicates achieving optimal distribution and compliance with boundary conditions, such as minimizing material usage and avoiding air pockets, and requires expensive expert knowledge and simulations.

Method used

A method using a data-based process parameter model, specifically an artificial neural network, to determine optimal application points, paths, and material quantities for thermally conductive paste application, ensuring efficient heat dissipation by training a quality model to maximize the distribution and quality of the paste after pressing.

Benefits of technology

This approach automates the process of applying thermally conductive paste, optimizing its distribution for improved heat dissipation, reducing material usage, and minimizing cycle time, while ensuring compliance with boundary conditions, thereby enhancing the efficiency and cost-effectiveness of heat management systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 1.1
    Figure 1.1
Patent Text Reader

Abstract

The invention relates to a computer-implemented method for training a data-based process parameter model (11), in particular in the form of an artificial neural network, for providing process parameters (P) of a manufacturing process depending on a distribution map (V) of heat-conducting paste (4) on an application surface after pressing, in particular for use in a method according to any one of claims 1 to 5, wherein the distribution map (V) indicates the distribution of heat-conducting paste (4) on the application surface after pressing, wherein the following steps are carried out for one or more specified distribution maps: - evaluating (S12) the process parameter model (11) by specifying the corresponding specified distribution map (V), in order to obtain process parameters (P), - determining (S14) a quality (G) of the process parameters (P) using a quality model (12, 13) depending on the corresponding specified distribution map (V), wherein the quality model (12, 13) indicates the achievement of the distribution of the heat-conducting paste (4) according to the corresponding specified distribution map (V) by the obtained process parameters (P); - training (S15) the data-based process parameter model (11) using a quality model (12, 13) so that the quality (G) is maximised.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Description

[0002] title

[0003] Method and device for planning and carrying out a process for applying a thermal paste using machine learning methods

[0004] Technical area

[0005] The invention relates to methods for applying a thermal paste to a cooling surface of a heat-generating component, so that optimal heat dissipation is achieved after pressing it onto a heat-dissipating surface of a heat sink. The invention further relates to possibilities for automatically determining process parameters for applying the thermal paste, by which the application process of the thermal paste is carried out in such a way that optimal distribution of the thermal paste is achieved.

[0006] Technical background

[0007] Active components, such as electronic components, power components, and the like, give off heat during operation. To protect them from overheating, they can be equipped with heat sinks designed to dissipate the heat generated during operation as effectively as possible. This usually involves assembling a corresponding system consisting of a component and a heat sink, coating a cooling surface of the component with thermal paste, and then pressing a heat dissipation surface of a heat-dissipating component, such as a heat sink, or a corresponding surface of a housing or the like onto the surface of the electronic component coated with thermal paste. The thermal paste is then distributed over the cooling surface of the component, thus enabling effective heat transfer between the cooling surface of the component and the heat dissipation surface of the heat sink or housing.

[0008] As cooling surfaces become larger, the degrees of freedom when applying the thermal paste before pressing increase, and the distribution of the thermal paste after pressing does not correlate in an obvious way with the application of the thermal paste during the application process. In particular, compliance with boundary conditions, such as that regions adjacent to the cooling surface must not come into contact with the thermal paste, the material used in the thermal paste must be minimized, no air pockets should occur between the cooling surface and the dissipation surface, and that the cycle time for applying the thermal paste to an electronic component should be as short as possible, is crucial. To date, the process parameters that dictate the application of the thermal paste have been determined by expert knowledge. This is supplemented by CFD simulations and experimental investigations, which are, however, complex.

[0009] Disclosure of the invention

[0010] According to the invention, a method for planning and carrying out an application process of a thermal paste on a cooling surface of a component according to claim 1 and a method for training a process parameter model for determining process parameters for the application process as well as corresponding devices according to the independent claims are provided.

[0011] Further embodiments are specified in the dependent claims.

[0012] According to a first aspect, a method for planning a manufacturing process is provided in which, in an application process, thermal paste is applied to an application surface and is then pressed to connect a cooling surface of a component to a heat dissipation surface, comprising the following steps:

[0013] Specifying a distribution map for a cooling surface of the component to be covered with the thermal paste; determining process parameters for the application process of the thermal paste to the application surface using a process parameter model.

[0014] Thermal paste is a pasty material with high viscosity and high thermal conductivity. It is applied between a cooling surface and a heat dissipation surface to assist in the transfer of heat from the cooling surface to the heat dissipation surface. The cooling surface is usually a flat surface on a heat-producing component, while the heat dissipation surface is usually provided on a heat sink or housing. For installation, the thermal paste is applied to the cooling surface or, if applicable, to the heat dissipation surface as one or more dots or drops, or in the form of a straight, curved, or angled strip (application path). The other side of the thermal paste is then pressed onto the cooling surface and the heat dissipation surface so that the applied thermal paste is displaced and widens and flows sideways into the space between the cooling surface and the heat dissipation surface.

[0015] The application process offers a wide range of options, and the spatial distribution of the thermal paste after application on the application surface is particularly crucial for the distribution of the thermal paste in the gap after pressing. Other process parameters, such as the amount of thermal paste at a particular location on the application surface and material properties, can also play a role.

[0016] The above method provides for the determination of the necessary process parameters for an application process for applying thermal paste using a process parameter model. The process parameter model is designed as a data-based model, in particular as an artificial neural network (e.g., a CNN), and makes it possible to determine application points and / or application paths for applying thermal paste to the cooling surface by specifying a distribution area of ​​an electronic component to be coated with thermal paste. The process parameters can thus define an application map that determines map points, each of which determines a position on the application surface where thermal paste is to be applied. Furthermore, the process parameters can determine an application path across the application surface.

[0017] Furthermore, the process parameters can determine the amount of thermal paste material applied to different positions on the application surface and / or the material properties of the thermal paste. For example, the process parameters can define the application path(s), the amount of thermal paste material applied to each application point, and the like.

[0018] The process parameters can thus be used to produce a system comprising the component and a heat dissipator, in particular a heat sink or a housing, by determining the positions of application of the thermal paste to the application surface by the process parameters.

[0019] To produce a component connected to a heat dissipation surface via a thermal paste, thermal paste is applied according to the process parameters and then the application surface, which can be a cooling surface or a heat dissipation surface, is connected to a heat dissipation surface or a cooling surface of a component by pressing (up to a predetermined distance).

[0020] According to a further aspect, a computer-implemented method for training a data-based process parameter model, in particular in the form of an artificial neural network, for providing process parameters of a manufacturing process depending on a distribution map of thermal paste on an application surface after pressing is provided, in particular for use in the above method, wherein the distribution map indicates the distribution of thermal paste on the application surface after pressing, wherein the following steps are carried out for one or more predetermined distribution maps:

[0021] - Evaluating the process parameter model by specifying the corresponding predefined distribution map to obtain process parameters,

[0022] Determining a quality of the process parameters using a quality model depending on the corresponding predefined distribution map, wherein the quality model indicates the achievement of the distribution of the thermal paste according to the corresponding predefined distribution map by the obtained process parameters;

[0023] Training the data-based process parameter model using a quality model so that the quality is maximized.

[0024] The process parameter model should be trained in such a way that, after pressing, the distribution of the thermal paste on the application surface achieves the highest possible quality. This quality is determined by a quality function which, depending on the distribution of the thermal paste after pressing, indicates how good the heat dissipation is and the quality of the system thus connected, consisting of component and heat dissipation. The quality function can, in particular, take into account the proportion of the cooling surface that is wetted by the thermal paste after pressing, the number and size of air pockets, the amount of thermal paste material used, the required application speed using a corresponding dispensing device, the time required for an application process, and / or the amount and size of the area of ​​thermal paste that laterally extends outside the cooling surface.

[0025] It can be provided that the quality model comprises a flow model and a quality function, wherein the flow model maps the process parameters onto a modeled distribution map by means of a particularly heuristic simulation and / or a CFD simulation, wherein the quality function evaluates the modeled distribution map with regard to the predetermined distribution map.

[0026] Alternatively, the quality model can be data-based, in particular in the form of an artificial neural network, and trained to represent a flow model and a subordinate quality function, so that predetermined process parameters are mapped to the quality, wherein the flow model maps the process parameters to a modeled distribution map by means of a particularly heuristic simulation and / or a CFD simulation, wherein the quality function evaluates the modeled distribution map with regard to the predetermined distribution map.

[0027] The process parameter model is trained accordingly according to the above procedure. Training is based on a flow model, which can be heuristic or data-driven, particularly in the form of an artificial neural network. The flow model simulates the pressing process after the thermal paste has been applied to the application surface according to specified process parameters.

[0028] The flow model can simulate, in a rule-based manner, a gradual displacement of material from application points of the thermal paste on the application surface into adjacent areas until a distribution corresponding to a desired distance between the cooling surface and the dissipation surface results as a gap.

[0029] The flow model can also be specified data-based, particularly in the form of an artificial neural network, to accelerate the training of the process parameter model and facilitate the derivation of the quality during the training of the process parameter model used in the training procedure. For this purpose, the heuristic flow model can be trained as a suitable data-based flow model.

[0030] The resulting distribution is evaluated according to the quality function defined above, based on predefined quality criteria. The determined quality is part of the cost function (loss function) for training the data-based process parameter model.

[0031] The process parameter model is trained according to the cost function so that it assigns a planar distribution of the thermal paste after pressing to process parameters that define the application process and that lead to the planar distribution after pressing.

[0032] Furthermore, it can be planned to train the process parameter model using the results of a CFD simulation, whereby the distribution of the thermal paste on a cooling surface is simulated based on specified process parameters. The assignment of the distribution of the thermal paste on the cooling surface after pressing to the process parameters previously used for application represents a training data set for further training of the process parameter model.

[0033] According to a further aspect, an apparatus for carrying out one of the above methods is provided.

[0034] Brief description of the drawings

[0035] Embodiments are explained in more detail below with reference to the attached drawings. They show:

[0036] Figure 1 shows a system comprising a heat-producing component and a heat sink mounted thereon;

[0037] Figure 2 is a flow chart illustrating a method for evaluating a process parameter model to determine process parameters according to which the thermal paste is to be applied to an application surface;

[0038] Figures 3a and 3b show plot maps of the process parameter model;

[0039] Figure 4 is a block diagram of a training system for the

[0040] Process parameter model;

[0041] Figure 5 is a flowchart illustrating a method for training the process parameter model; and

[0042] Figure 6 shows an illustration of the determination of a modeled

[0043] Distribution map using a heuristic flow model.

[0044] Description of Embodiments Figure 1 shows a system 1 comprising a heat-producing component 2 with a cooling surface 21 and a heat sink 3 applied thereto. The heat sink 3 has cooling fins 32 and a heat dissipation surface 31 opposite the cooling surface 21. A thermally conductive paste 4 is introduced between the cooling surface 21 and the heat dissipation surface 31 by means of an application and pressing process.

[0045] Figure 2 shows a flow chart illustrating a method for planning a process for applying thermal paste to an application surface, such as the cooling surface 21 of the component 2 or the heat dissipation surface 31, and a subsequent pressing of a surface, namely the other of the cooling surface 21 and the heat dissipation surface 31. Such a process is usually carried out by applying thermal paste approximately in the center of the application surface and pressing the heat dissipation surface evenly onto the thermal paste so that it is distributed between the cooling surface and the heat dissipation surface and ideally fills this space completely without thermal paste escaping at the edges or the contact surface or heat dissipation becoming too small.

[0046] In an automated process, the thermal paste is applied to the application surface in a point-like manner or along one or more application paths by a so-called dispensing device, whereby the result after pressing depends considerably on the amount of thermal paste applied and the previously selected application path.

[0047] The application process is determined by process parameters that specify the application positions of the application points or the path paths of the application paths. The process parameters can also determine the thickness of the applied thermal paste material.

[0048] The process parameter model can be designed as a data-based neural network and assign the desired result of the distribution of the thermal paste on the application surface to suitable process parameters.

[0049] For this purpose, the application surface is rasterized, for example, in a matrix-like distribution map, and the pixels where thermal paste is to be present after pressing are assigned a corresponding entry, such as the value "1," while the remaining pixels are assigned the value "0." Since the process parameter model is intended to be used universally for a variety of components, the process parameter model assumes a specified size for the distribution map that corresponds to a specified maximum possible cooling surface of the component. For smaller cooling surfaces, for example, only a central area can be designated for use, and the surrounding areas can be filled in such a way that they do not indicate wetting with thermal paste.

[0050] In step S1, the application area is first specified with regard to its dimensions and shape and this is mapped into a distribution map of the same or higher dimension by adding edge areas accordingly.

[0051] Using the trained process parameter model, which can preferably be implemented as a convolutional neural network (CNN), this distribution map is evaluated in step S2. The process parameter model is trained to evaluate the distribution map of the specific dimension.

[0052] Depending on the training or configuration of the process parameter model, the output can be a corresponding application map based on process parameters, on which the elements of the matrix where thermal paste material is to be applied are marked accordingly, for example, depending on the amount of thermal paste material for the respective map pixel with increasing values ​​such as 2, 3, etc. This is shown, for example, in Figure 3a.

[0053] Alternatively, the output of the process parameter model can also be process parameters in the form of path parameters of an application trajectory, as shown, for example, in Figure 3b. The path parameters determine the start and end points of an application path and the material quantity along this application path. If the path parameters are output as process parameters, they can be used directly for the production process, as they already contain an optimization that was performed during the training process for the process parameter model. In step S3, the determined process parameters are used, for example, directly for applying the thermal paste in an application process.Alternatively, when the application card is issued, the path parameters for the application process are first determined according to a known path determination method, which determines an optimal application trajectory with regard to the application time and the properties of the dispensing device.

[0054] The process parameters do not necessarily have to be used directly in the production process. Alternatively, the determined process parameters can also be used as an aid in a development process, in which the determined process parameters are evaluated and adjusted if necessary before the process parameters are used on the machine.

[0055] The training of the process parameter model is enabled by a structure as outlined in Figure 4. Figure 5 shows a flowchart illustrating a method for training the process parameter model.

[0056] The structure shown comprises a process parameter model 11, a flow model 12 and a quality function 13. The flow model 12 and the quality function 13 together form a quality model.

[0057] For training, different distribution maps are successively specified in step S11, which indicate the distribution of the thermal paste in the gap.

[0058] In step S12, the distribution map V is converted into process parameters P using the process parameter model 11, which can be represented as a distribution map A or as path parameters and material properties, as described above. The process parameter model 11 is initially untrained and is continuously trained throughout this process. In step S12, the process parameter model 11 is always used according to the current training status.

[0059] The process parameter model 11 can be implemented as a data-based model, such as a neural network, particularly in the form of a convolutional neural network. In particular, all data-based, trainable models that can convert map information into parameter values ​​are suitable for the method described herein.

[0060] In step S13, the flow model 12 is applied to the process parameters P. The flow model 12 essentially generates a modeled distribution map Vmod from the process parameters P. The flow model 12 can, for example, be a heuristic model in which a distribution map V specified or generated from the process parameter P is evaluated. The distribution map corresponds to a matrix of individual points representing the cooling surface, for which it is determined whether thermal paste should be present after pressing and in what quantity. The quantity of thermal paste material is determined by the final distance between the cooling surface and the heat dissipation surface.

[0061] The flow model 12 can be designed as a heuristic model. Starting from the application map A, in which a material quantity is specified for each map point, the heuristic model can lower the height of a virtual plane that defines a maximum available volume for each map point. As the virtual plane is successively lowered, excess thermal paste material is displaced to neighboring map points. If the amount of thermal paste located there exceeds the maximum available volume, this process is repeated until the displaced material is distributed over the neighboring map points and the amount of thermal paste for all map points does not exceed the maximum available volume. The successive reduction in the volume of each map point occurs until a volume that corresponds to the distance between the opposing surfaces for this map point.The resulting distribution is shown in the distribution map.

[0062] In the example shown in Figure 6, a deposition height of 2 is assumed for a diamond-shaped surface with the matrix. For the thus specified deposition map A, which results from the process parameters, a virtual plane is now provided which is iteratively lowered to height 1. The distribution of the displaced material around the initial area of ​​the applied material can be seen. The distribution is uniform when the virtual plane is lowered parallel to the surface onto the virtual cooling surface (deposition map). If the virtual plane is lowered at an angle, varying heights result for each map point of the deposition map A. The displacement is carried out until each of the map points of the deposition map A has a material quantity that is no greater than the material quantity that is permissible according to the height.

[0063] The heuristic rules result in a distribution of the thermal paste on the cooling surface that corresponds to the modeled distribution map.

[0064] The flow model can also be data-based, e.g., in the form of a neural network, particularly a convolutional neural network (CNN). This is trained according to the results of the heuristic flow model, another rule-based flow model, and / or simulation models, such as CFD simulations. The advantage of transferring the flow model to a data-based model is that it is derivable and can thus be easily used in subsequent gradient formation.

[0065] Using the quality function 13, a quality G can now be determined in step S14 for the determined distribution map Vmod, which results from the desired distribution V of the thermal paste and the modeled distribution of the thermal paste. The quality function 13 can provide that each discretized map point of the modeled distribution map V mOd is assigned a value, whereby the values ​​for each map point are determined depending on whether the desired actual coverage with thermal paste and the modeled coverage with thermal paste match at this map point. Uncovered map points of the distribution map are rated negatively if they should be covered, covered map points of the distribution map are rated positively if they should be covered, covered map points that do not belong to the distribution area to be covered are rated slightly negatively, and taboo zones, which affect areas outside the distribution area, are rated strongly negatively.

[0066] The quality G can further evaluate the distribution of the occupied map points of the application map A before pressing, in particular with regard to the required or estimated time duration of application by a dispensing device, which results from the properties of the dispensing device and the application process, the total amount of applied thermal conductive material, the number of separate application areas and the like.

[0067] Alternatively, the quality can also directly consider the process parameters. If the process parameters specify an application path, this path length and application time can be evaluated directly. Such an evaluation of the process parameters is less complex than evaluating a resulting application map.

[0068] In a further embodiment, the heuristic method described above can also be used as a basis for training a data-based quality model by directly mapping the process parameters to be output by the process parameter model to a quality. For this purpose, the flow model and the quality function are evaluated, and the resulting training data sets consisting of process parameters and qualities (labels) are used to train the quality model. The advantage of transferring the flow model and the quality function into a data-based quality model is that it is derivable and can thus be easily used in subsequent gradient formation.

[0069] In step S15, a training step for the process parameter model takes place, which is based on the quality determined in this way. The quality function then serves as the loss function or cost function for training the process parameter model. Training can be based on backpropagation. The model parameters of the process parameter model are determined by gradient formation of the quality over the model parameters and these are successively adapted in this way. In step S16, a termination criterion is checked. The termination criterion indicates when training of the process parameter model using the above method is complete. If it is determined in step S16 that the termination criterion is met (alternative: yes), the method is terminated and in step S17 the process parameter model is used to determine process parameters for a given distribution map. Otherwise (alternative: no), the process returns to step S11.

[0070] By training the process parameter model using the flow model and the quality function or using the data-based quality model, the process parameters are automatically optimized to maximize quality. This means that during the training process, the process parameter model is repeatedly trained based on predefined distribution maps until an optimized / maximum quality is achieved.

Claims

Claims 1. Method for planning a manufacturing process in which thermal paste (4) is applied to a substrate in a coating process and subsequently pressed to connect a cooling surface (21) of a component (2) with a heat dissipation surface (31), comprising the following steps: Specifying (S1) a distribution map for a cooling surface (21) of the component (2) to be covered with the thermal paste (4); Determine (S2) process parameters (P) for the application process of the thermal paste (4) onto the application surface using a process parameter model (11).

2. Method according to claim 1, wherein the process parameters (P) define an application map which determines map points which each define a position on the application surface where thermal paste (4) is to be applied.

3. Method according to claim 1, wherein the process parameters (P) determine a deposition path over the deposition area.

4. Method according to claim 2 or 3, wherein the process parameters (P) determine a quantity of thermal paste (4) for different positions of the application surface and / or material properties of the thermal paste.

5. Method according to any one of claims 1 to 4, wherein the process parameters (P) are used to produce a system from the component (2) and a heat sink (3), in particular a heat sink or a housing, by determining the positions of the application of the thermal paste to the application surface by the process parameters (P).

6. Computer-implemented method for training a data-based process parameter model (11), in particular in the form of an artificial neural network, for providing process parameters (P) of a manufacturing process depending on a distribution map (V) of Thermal paste (4) on an application surface after compression, in particular for use in a method according to one of claims 1 to 5, wherein the distribution map (V) specifies the distribution of thermal paste (4) on the application surface after compression, wherein the following steps are carried out for one or more predetermined distribution maps: - Evaluating (S12) the process parameter model (11) by specifying the corresponding predefined distribution map (V) to obtain process parameters (P), Determining (S14) a quality (G) of the process parameters (P) using a quality model (12, 13) depending on the corresponding predefined distribution map (V), wherein the quality model (12, 13) indicates the achievement of the distribution of the thermal paste (4) according to the corresponding predefined distribution map (V) by the obtained process parameters (P); Training (S15) the data-based process parameter model (11) using a performance model (12, 13) such that the performance (G) is maximized. Method according to claim 6, wherein the performance model (12, 13) comprises a flow model (12) and a performance function (13), wherein the flow model (12) maps the process parameters (P) to a modeled distribution map (V) by means of a heuristic simulation and / or a CFD simulation. mO d) maps, where the quality function (G) represents the modeled distribution map (V) mOd) evaluated with respect to the predetermined distribution map (V). Method according to claim 6, wherein the quality model (12, 13) is data-based and trained to represent a flow model (12) and a subordinate quality function (G) such that predetermined process parameters (P) are mapped to the quality (G), wherein the flow model (12) maps the process parameters (P) to a modeled distribution map by means of a heuristic simulation and / or a CFD simulation, wherein the quality function (13) represents the modeled distribution map (V). mO d) evaluated with respect to the specified distribution map (V). Device for carrying out one of the methods according to any one of claims 1 to 8. A computer program product comprising instructions that, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 8. A machine-readable storage medium comprising instructions that, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 8.