Method for thermally simulating semiconductor device and associated apparatus
By constructing a segmented simulation model and a flux conservation autoencoder to predict thermal characteristics, the problem of excessive computing power consumption in the core particle system was solved, and efficient and accurate thermal simulation effects were achieved.
Patent Information
- Application Number
- CN202510887346.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-27
- Publication Date
- 2025-09-26
AI Technical Summary
Traditional thermal simulation methods consume too much computing power in core-particle systems and are unable to meet high computing power requirements. In addition, existing equivalent thermal models and machine learning methods have limitations in accuracy and applicability.
A simulation model is constructed and divided into multiple subdomain units. The thermal characteristics are encoded and predicted through subdomain encoders and flux conservation autoencoders to meet the heat flux conservation constraints between subdomain units. A deep neural network architecture is used for training and iterative processing.
It achieves accurate prediction of the thermal characteristics of any core particle system shape and size, significantly reduces computing power consumption, improves simulation analysis speed and accuracy, and has better stability and robustness.
Smart Images
Figure CN120706273A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of semiconductor device simulation, and in particular to a method for performing thermal simulation on a semiconductor device, a computer program product, a computer-readable storage medium, and a computer device. Background Art
[0002] To meet the increasing integration requirements of transistors in high-computing-power chips, chip-scale packaging technologies featuring heterogeneous chiplet integration and three-dimensional vertical stacking are becoming increasingly prominent within the industry. Consequently, issues surrounding chiplet thermal management and structural reliability are becoming increasingly prominent. Due to the complexity of chiplet systems, traditional, rigorous thermal simulations based on physical models struggle to meet the computing power demands of large-scale computations, posing a significant challenge to design simulations.
[0003] Because the core-particle system uses chip-level packaging technology with heterogeneous core integration and three-dimensional vertical stacking, traditional thermal simulation methods (such as the finite element method and the finite difference method) face huge computing overhead in thermal simulation, which significantly increases the difficulty of design simulation. However, the equivalent thermal model method and end-to-end machine learning methods also have significant limitations in the thermal simulation of core-particle systems. For example, the prediction accuracy of the equivalent thermal model method is highly dependent on the construction quality of the equivalent structure. Due to the multi-level and complex structure of the core-particle system, the equivalent structure is difficult to accurately construct, which in turn restricts the simulation accuracy and efficiency. The end-to-end machine learning method is limited by the model's lack of interpretability and generalization ability. When the core-particle system layout changes, its simulation applicability is greatly reduced. Summary of the Invention
[0004] In view of the above problems, the present invention is proposed to provide a method and related equipment for performing thermal simulation on semiconductor devices that overcome the above problems or at least partially solve the above problems, solve the problem of excessive computing power consumption in the thermal simulation of the core particle system in the prior art, and achieve the effect of saving computing power.
[0005] Specifically, according to one aspect of the present invention, the present invention provides a method for performing thermal simulation on a semiconductor device, comprising:
[0006] Constructing a simulation model for describing characteristic information of the semiconductor device, and dividing a plurality of subdomain units from the simulation model; wherein one of every two adjacent subdomain units is a target subdomain unit and the other is a neighboring unit;
[0007] Encoding the thermal characteristics of each sub-domain unit through a sub-domain encoder to obtain a sampling latent vector group corresponding to each sub-domain unit;
[0008] Combining the sampled latent vector groups of each target sub-domain unit and the corresponding neighboring units to obtain an input latent vector set corresponding to each target sub-domain unit;
[0009] Processing the input latent vector set by a flux conservation autoencoder to obtain an output latent vector set; wherein the flux conservation autoencoder at least satisfies a constraint condition of heat flux conservation between the target subdomain unit and the neighboring unit;
[0010] The output latent vector set is decoded by a subdomain decoder corresponding to the subdomain encoder to obtain a target thermal feature corresponding to the target subdomain unit.
[0011] Optionally, the method for performing thermal simulation on a semiconductor device further includes:
[0012] Constructing a flux conservation autoencoder based on a deep neural network architecture; wherein the autoencoder has at least four fully connected layers;
[0013] The flux conservation autoencoder is trained using the input latent vector set, and the fully connected layer is guided to satisfy the constraint condition of heat flux conservation between the target subdomain unit and the neighboring units during the training process.
[0014] Optionally, processing the input latent vector set by a flux conservation autoencoder to obtain an output latent vector set includes:
[0015] The input latent vector set is input into the flux conservation autoencoder for fixed-point iterative processing until the output result of the flux conservation autoencoder converges or the number of iterations of the flux conservation autoencoder reaches a set number, thereby obtaining the output latent vector set; wherein,
[0016] The sampling latent vector group includes at least a first thermal feature latent vector and a second thermal feature latent vector;
[0017] The iteration variables in the fixed-point iterative process are the first thermal feature latent vectors of the input latent vector set, while the second thermal feature latent vectors remain unchanged; the first thermal feature latent vectors in the input latent vector set are assigned initial values;
[0018] The obtained target thermodynamic characteristic corresponding to the target subdomain unit is the first thermodynamic characteristic corresponding to the target subdomain unit.
[0019] Optionally, after obtaining the target thermal characteristics corresponding to the target subdomain unit, the method for performing thermal simulation on a semiconductor device further includes:
[0020] Taking each of the subdomain units as the target subdomain unit, obtaining target thermal characteristics of each of the subdomain units of the simulation model;
[0021] Each of the target thermal characteristics is mapped onto the simulation model to obtain information reflecting the distribution of the target thermal characteristics on the semiconductor device.
[0022] Optionally, constructing a simulation model for describing characteristic information of the semiconductor device includes:
[0023] Constructing a three-dimensional model according to the layout of the semiconductor device, and using the three-dimensional model as the simulation model; and
[0024] A plurality of subdomain units are divided from the simulation model, including:
[0025] The simulation model is divided on each coordinate axis according to a Cartesian coordinate system to obtain a plurality of subdomain units; and
[0026] The neighborhood unit includes at least one in-plane neighborhood unit and / or at least one out-of-plane neighborhood unit, the in-plane neighborhood unit and the target sub-domain unit are adjacent on the X-axis or Y-axis of the Cartesian coordinate system, and the out-of-plane neighborhood unit and the target sub-domain unit are adjacent on the Z-axis of the Cartesian coordinate system.
[0027] Optionally, the sub-domain units are equal in size.
[0028] Optionally, encoding the thermal characteristics of each subdomain unit by a subdomain encoder includes:
[0029] respectively acquiring the thermal characteristics from different sampling points within each of the subdomain units;
[0030] Each of the thermal features is encoded separately by the sub-domain encoder to encode the thermal features into a one-dimensional feature latent vector, and a combination of the one-dimensional feature latent vectors corresponding to the sub-domain encoder is used as the sampling latent vector group corresponding to the sub-domain encoder.
[0031] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the computer program implements the steps of any of the above methods for performing thermal simulation on a semiconductor device.
[0032] According to another aspect of the present invention, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the steps of any one of the above methods for performing thermal simulation on a semiconductor device.
[0033] According to another aspect of the present invention, a computer device is provided, characterized in that it includes a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to implement any one of the steps of the above-mentioned method for performing thermal simulation on a semiconductor device.
[0034] In the method for thermal simulation of semiconductor devices of the present invention, the target thermal characteristics of the target subdomain unit and the neighboring units are used as input, and the target thermal characteristics of the target subdomain unit are predicted by the flux conservation autoencoder, thereby realizing the prediction of the thermal characteristics of any core particle system shape and size. During the prediction process, not only the thermal information of the target subdomain unit itself is taken into account, but also the thermal information of the adjacent subdomain units, as well as the constraints of the heat flux conservation between the two, which can make the prediction results more accurate. Moreover, by dividing the entire simulation model into multiple small units, when performing prediction calculations, the problem of excessive computing power consumption in the thermal simulation of the core particle system in the prior art can be solved, computing power consumption can be significantly reduced, and the effect of saving computing power can be achieved.
[0035] Compared to traditional finite element and finite difference methods, this method can significantly accelerate simulation analysis, reduce computing power requirements, and meet the iteration speed requirements for thermal information perception in core-grain system layout design. Furthermore, compared to other purely data-driven "black box" machine learning methods, the predictions of this method based on the flux conservation autoencoder have better stability, robustness, and accuracy, and can handle more complex geometries and a wider range of model parameters.
[0036] Based on the following detailed description of specific embodiments of the present invention in conjunction with the accompanying drawings, those skilled in the art will become more aware of the above and other objects, advantages and features of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Hereinafter, some specific embodiments of the present invention will be described in detail in an exemplary and non-limiting manner with reference to the accompanying drawings. The same reference numerals in the accompanying drawings indicate the same or similar components or parts. It should be understood by those skilled in the art that these drawings are not necessarily drawn to scale. In the accompanying drawings:
[0038] Figure 1 is a schematic flow chart of a method for performing thermal simulation on a semiconductor device according to one embodiment of the present invention;
[0039] Figure 2 is a schematic partial flow chart of a method for performing thermal simulation on a semiconductor device according to one embodiment of the present invention;
[0040] Figure 3 is an architectural diagram of a flux conservation autoencoder in a method for performing thermal simulation on a semiconductor device according to an embodiment of the present invention;
[0041] Figure 4 is a schematic diagram of a fixed-point iterative processing process of a flux conservation autoencoder in a method for performing thermal simulation on a semiconductor device according to an embodiment of the present invention;
[0042] Figure 5 is a schematic partial flow chart of a method for performing thermal simulation on a semiconductor device according to one embodiment of the present invention;
[0043] Figure 6 Schematic diagram of a subdomain encoder and a subdomain decoder in a method for thermal simulation of a semiconductor device according to an embodiment of the present invention
[0044] Figure 7 is a schematic diagram of the working process of a sub-domain encoder and a sub-domain decoder in a method for performing thermal simulation on a semiconductor device according to an embodiment of the present invention;
[0045] Figure 8 is a schematic flow chart of a method for performing thermal simulation on a semiconductor device according to one embodiment of the present invention;
[0046] Figure 9 is a schematic diagram of a computer program product according to one embodiment of the present invention;
[0047] Figure 10 is a schematic diagram of a computer-readable storage medium according to one embodiment of the present invention; and
[0048] Figure 11 is a schematic diagram of a computer device according to one embodiment of the present invention. DETAILED DESCRIPTION
[0049] Refer to the following Figures 1 to 11 The method and related apparatus for performing thermal simulation on a semiconductor device according to an embodiment of the present invention are described.
[0050] Figure 1 FIG. 1 is a schematic flow chart of a method for performing thermal simulation on a semiconductor device according to an embodiment of the present invention. Figure 1 As shown, an embodiment of the present invention provides a method for performing thermal simulation on a semiconductor device, which includes:
[0051] Step S100: Construct a simulation model for describing characteristic information of a semiconductor device and segment the simulation model into multiple subdomain units. Of every two adjacent subdomain units, one is a target subdomain unit and the other is a neighboring unit. The simulation model is segmented into multiple units, and when performing prediction, prediction is first performed on each unit.
[0052] In step S200, a subdomain encoder is used to encode the thermal characteristics of each subdomain unit to obtain a set of sampled latent vectors corresponding to each subdomain unit. The thermal characteristics are used to reflect the heat-related information of the subdomain unit. The thermal characteristics include at least temperature, power, and / or thermal conductivity. In some optional embodiments of the present invention, the thermal characteristics include at least temperature, power, thermal conductivity, and / or the boundaries of the subdomain unit.
[0053] Step S300: Combine the sampled latent vectors of each target sub-domain unit and the corresponding neighboring units to obtain an input latent vector set corresponding to each target sub-domain unit. In this step, the target sub-domain unit to be predicted is associated with its neighboring sub-domain units.
[0054] Step S400: Process the input latent vector set using a flux conservation autoencoder to obtain an output latent vector set. The flux conservation autoencoder satisfies at least the constraint of heat flux conservation between the target subdomain unit and its neighboring units. When processing the input latent vector set, the flux conservation autoencoder considers the constraint of heat flux conservation between the target subdomain unit and its neighboring units, ensuring accurate prediction.
[0055] In step S500, the output latent vector set is decoded by a subdomain decoder corresponding to the subdomain encoder to obtain a target thermal feature corresponding to the target subdomain unit. The target thermal feature is the thermal feature to be predicted corresponding to the target subdomain unit, preferably temperature.
[0056] In the method for thermal simulation of semiconductor devices according to an embodiment of the present invention, since the thermal characteristics of the target subdomain unit and the neighboring units are used as input, the target thermal characteristics of the target subdomain unit are predicted by the flux conservation autoencoder, thereby realizing the prediction of the thermal characteristics of any core particle system shape and size. In the prediction process, not only the thermal information of the target subdomain unit itself is taken into account, but also the thermal information of the adjacent subdomain units, as well as the constraints of the heat flux conservation between the two, which can make the prediction results more accurate. Moreover, by dividing the entire simulation model into multiple small units, when performing prediction calculations, the problem of excessive computing power consumption in the thermal simulation of the core particle system in the prior art can be solved, computing power consumption can be significantly reduced, and the effect of saving computing power can be achieved.
[0057] Compared to traditional finite element and finite difference methods, this method can significantly accelerate simulation analysis, reduce computing power requirements, and meet the iteration speed requirements for thermal information perception in core-grain system layout design. Furthermore, compared to other purely data-driven "black box" machine learning methods, the predictions of this method based on the flux conservation autoencoder have better stability, robustness, and accuracy, and can handle more complex geometries and a wider range of model parameters.
[0058] In some embodiments of the present invention, Figure 1 As shown, after obtaining the target thermal characteristics corresponding to the target subdomain unit, the method for performing thermal simulation on the semiconductor device further includes:
[0059] Step S600 : Taking each subdomain unit as a target subdomain unit, a target thermal characteristic of each subdomain unit of the simulation model is obtained.
[0060] Step S700 : mapping each target thermal characteristic onto a simulation model to obtain information reflecting the distribution of the target thermal characteristics on the semiconductor device.
[0061] In order to more comprehensively reflect the thermal information of the simulation model, after obtaining the target thermal characteristics of each subdomain unit (such as the predicted temperature), the subdomain units are spliced together, so that the predicted target thermal characteristics can be reflected on the simulation model, which can intuitively understand the thermal information of the semiconductor device. It is also convenient to perform post-processing analysis on the predicted target thermal characteristics of the semiconductor device, extract the hot spot information of key parts, and output the prediction report.
[0062] In some embodiments of the present invention, the subdomain encoder, flux conservation autoencoder, and subdomain decoder are integrated into a single overall encoder. That is, the overall encoder takes as input the subdomain units and outputs the target thermal signature corresponding to the target subdomain units. Of course, the subdomain encoder, flux conservation autoencoder, and subdomain decoder can also be independently configured. Alternatively, they can be partially integrated, such as by integrating the subdomain encoder and subdomain decoder together.
[0063] In some embodiments of the present invention, Figure 2 and Figure 3 As shown, the method for performing thermal simulation on a semiconductor device further includes:
[0064] Step S800: construct a flux conservation autoencoder based on a deep neural network architecture, wherein the flux conservation autoencoder has at least four fully connected layers.
[0065] Step S900 trains a flux conservation autoencoder using the input latent vector set, guiding the fully connected layer to satisfy the heat flux conservation constraint between the target subdomain unit and its neighboring units during training. In other words, during the training of the flux conservation autoencoder, a mapping of the physical laws governing heat flux conservation between subdomain units can be obtained, facilitating the prediction of thermal information for semiconductor devices.
[0066] In some embodiments of the present invention, Figure 4As shown, the above step S400 processes the input latent vector set through the flux conservation autoencoder to obtain the output latent vector set, specifically: the input latent vector set is input into the flux conservation autoencoder for fixed-point iterative processing until the output result of the flux conservation autoencoder converges or the number of iterations of the flux conservation autoencoder reaches the set number, then the output latent vector set is obtained. Among them, the sampling latent vector group includes at least the first thermal feature latent vector and the second thermal feature latent vector. The iteration variables in the fixed-point iterative processing are the first thermal feature latent vectors of the input latent vector set, while the second thermal feature latent vectors remain unchanged, that is, in Figure 4 As shown in , a Fixed condition is set in the deep neural network to ensure that each second thermal feature latent vector remains unchanged. The first thermal feature latent vector in the input latent vector set is assigned an initial value. The target thermal feature is the first thermal feature corresponding to the target subdomain unit. For example, the first thermal feature latent vector is the temperature latent vector, and the second thermal feature latent vector is the power latent vector, the thermal conductivity latent vector and / or the boundary latent vector of the subdomain unit. In an embodiment of the present invention, the prediction result can be made accurate by judging whether the output result converges or whether the maximum number of iterations is met.
[0067] In some embodiments of the present invention, constructing a simulation model for describing characteristic information of a semiconductor device includes: constructing a three-dimensional model according to a layout of the semiconductor device, and using the three-dimensional model as the simulation model.
[0068] In some embodiments of the present invention, dividing a plurality of subdomain units from a simulation model includes: dividing the simulation model on each coordinate axis according to a Cartesian coordinate system to obtain a plurality of subdomain units. In addition, the neighborhood unit includes at least one in-plane neighborhood unit and / or at least one out-of-plane neighborhood unit, the in-plane neighborhood unit and the target subdomain unit are adjacent on the X-axis or Y-axis of the Cartesian coordinate system, and the out-of-plane neighborhood unit and the target subdomain unit are adjacent on the Z-axis of the Cartesian coordinate system. Preferably, the sizes of the subdomain units are equal, which can significantly improve computing efficiency and save computing power, and will not cause uneven distribution of computing power or waste of computing power because some subdomain units are relatively large and some subdomain units are relatively small.
[0069] In some embodiments of the present invention, Figures 5 to 7 As shown, the above step S200, encoding the thermal characteristics of each sub-domain unit by the sub-domain encoder, includes:
[0070] Step S210 , obtaining thermal characteristics from different sampling points in each subdomain unit.
[0071] In step S220, each thermal feature is encoded separately by a sub-domain encoder to encode the thermal feature into a one-dimensional feature latent vector, and the combination of the one-dimensional feature latent vectors corresponding to the sub-domain encoder is used as the sampling latent vector group corresponding to the sub-domain encoder.
[0072] like Figure 7 As shown, the thermal features that can be obtained at the sampling points are temperature, power, thermal conductivity and / or the boundaries of the subdomain units, and the one-dimensional feature latent vectors obtained are respectively the temperature latent vector, the power latent vector, the thermal conductivity latent vector, and the boundary latent vector of the subdomain unit. Then the sampling latent vector group of the subdomain unit includes four one-dimensional feature latent vectors: the temperature latent vector V_temp, the power latent vector V_power, the thermal conductivity latent vector V_kappa, and the boundary latent vector V_bc of the subdomain unit. In some alternative embodiments of the present invention, the thermal features that can be obtained at the sampling points are temperature, power, and thermal conductivity, and the one-dimensional feature latent vectors obtained are respectively the temperature latent vector, the power latent vector, and the thermal conductivity latent vector. Then the sampling latent vector group of the subdomain unit includes three one-dimensional feature latent vectors: the temperature latent vector V_temp, the power latent vector V_power, and the thermal conductivity latent vector V_kappa.
[0073] In some embodiments of the present invention, the present invention provides a machine learning architecture based on a combined encoder, which can quickly obtain the target thermal feature distribution of complex three-dimensional models, such as temperature distribution, by dividing the core particle system model into subdomains, extracting and encoding the thermal information such as temperature, material and / or boundary of the subdomain, and using the flux conservation autoencoder between subdomain units for fixed-point iteration. Figure 8 The input for this method of thermal simulation of semiconductor devices includes the following: a core system design layout file, which is used to construct a 3D model of the semiconductor device for temperature prediction; the assembly relationship of the cores, specifically the arrangement of the cores and the interconnection between the cores and the interposer through microbumps; and material and boundary conditions, which require the thermal conductivity and power of each part of the core system, as well as the heat exchange coefficient of the heat transfer boundary.
[0074] The method for performing thermal simulation on a semiconductor device according to an embodiment of the present invention specifically includes:
[0075] Step S101 constructs a 3D model based on the original design layout of the core system. This 3D model serves as the aforementioned simulation model. Specifically, this step involves constructing a 3D model based on the semiconductor device layout and using this 3D model as the simulation model. Furthermore, equally spaced sampling points are set on the 3D model to prepare for the next step of sub-domain unit segmentation.
[0076] Step S102: Divide the entire 3D model into sub-domain units of equal size. This allows each sub-domain unit to have a fixed number of sampling points n in the x, y, and z directions. x , n y , n z It should be noted that when cutting the boundary of the simulation model, the subdomain unit needs to exceed the boundary of the simulation model, so that the corresponding subdomain unit can include the boundary of the simulation model to ensure the accuracy of the prediction results.
[0077] In step S103, the thermal feature information of the sub-domain unit is sampled and encoded into a one-dimensional feature latent vector. Multiple one-dimensional feature latent vectors are aggregated to obtain a sampled latent vector group corresponding to the sub-domain unit. Specifically, this can be performed using a sub-domain encoder. The sub-domain encoder has a corresponding sub-domain decoder. In other words, the sub-domain encoder encodes the thermal features of each sub-domain unit to obtain a sampled latent vector group corresponding to each sub-domain unit.
[0078] like Figure 6 and Figure 7 As shown in the figure, after the three-dimensional model is divided into sub-domain units of equal size, each sub-domain unit has a fixed number of sampling points. For the thermal feature information and sub-domain unit information in different sub-domain units, such as the thermal feature information is temperature, power, thermal conductivity, and the sub-domain unit information is the boundary condition, the convolutional neural network is used to construct the downsampling and upsampling process of the sub-domain encoder and the symmetrical sub-domain decoder, and the one-dimensional vector is obtained by flattening the input channel between the sub-domain encoder and the symmetrical sub-domain decoder, and then a symmetrical fully connected layer is set, with the one-dimensional feature latent vector v as the connection between the encoder and the decoder. The architecture of the sub-domain encoder and the symmetrical sub-domain decoder is shown in the figure. Figure 6 As shown in Figure 2, the subdomain encoder and the symmetrical subdomain decoder have a symmetrical structure, where the convolution layer is 3 layers, the fully connected layer is 2 layers, and the length of the one-dimensional feature latent vector depends on the size of the subdomain unit. x =n y =n z =16 as an example, the latent vector length can be 64.
[0079] Step S104: constructing one-dimensional latent vectors of adjacent sub-domain units. That is, this step is to combine the sampled latent vector groups of each target sub-domain unit and the corresponding neighboring units to obtain an input latent vector set corresponding to each target sub-domain unit.
[0080] Step S105: Use the flux conservation autoencoder between subdomain units to perform fixed-point iteration. In other words, this step is to process the input latent vector set through the flux conservation autoencoder to obtain the output latent vector set.
[0081] The key to this step is the flux conservation autoencoder, which is used to describe the conduction of heat flux between subdomains. In other words, the flux conservation autoencoder at least satisfies the constraint condition of heat flux conservation between the target subdomain unit and the neighboring unit. The flux conservation autoencoder between adjacent subdomain units is specifically a flux conservation autoencoder based on a deep neural network (DNN) architecture with fully connected layers. The architecture of the flux conservation autoencoder is as follows: Figure 3 As shown, the fully connected layer has 4-6 layers, the length of the latent vector ξ can be 1 / 10 to 1 / 5 of the input parameter length, the output has the same length as the input, and its total length can be calculated as: l = (l temp +l power +l kappa +l bc )×n s , where l temp is temperature, l power is power, l kappa is the thermal conductivity, l bc is the length of the latent vector ξ, n s is the number of adjacent subdomain units, that is, the total number of a target subdomain unit and the corresponding neighboring units. Under three-dimensional conditions, n s =7, that is, 5 in the plane and 1 in each out-of-plane direction. In some alternative embodiments of the present invention, the number of adjacent sub-domain units can be n s =27, that is, in a 3x3x3 rectangular array.
[0082] When training the flux conservation autoencoder, the subdomain encoder can be used to obtain a one-dimensional sampling latent vector group of temperature, thermal conductivity, power, and boundary corresponding to the subdomain unit; according to the adjacent relationship of the subdomain units, a group of latent vectors are spliced together; that is, the sampling latent vector groups of the target subdomain unit and all the neighboring units corresponding to the target subdomain unit are spliced together. Figure 4 As shown, the one-dimensional feature latent vectors of the target subdomain unit and the target subdomain unit 1, target subdomain unit 2, target subdomain unit 3, and target subdomain unit adjacent to the target subdomain unit 0 are spliced together as the input of the flux conservation autoencoder, that is, as the input latent vector set of the flux conservation autoencoder.
[0083] The temperature prediction of the core particle system is achieved through fixed-point iteration of the flux conservation autoencoder. In the prediction mode, the temperature of the target subdomain unit can be given an initial value, such as 0, and the subdomain encoder is used to encode the temperature, thermal conductivity, power, boundary, etc. into a one-dimensional vector. At the same time, the latent vectors of adjacent subdomain units are spliced to obtain the input latent vector set corresponding to the target subdomain unit, that is, as a whole {v1,...,v n}, and then input the flux conservation autoencoder; the output of the flux conservation autoencoder is a set of one-dimensional vectors {v1′, .., v n ′}. The temperature latent vectors of the target subdomain unit and / or neighboring units are updated, while the remaining one-dimensional feature latent vectors (thermal conductivity, power, and boundary) remain unchanged. The temperature of the target subdomain unit is input again to complete one iteration. After multiple iterations, the output one-dimensional vector is judged to have converged. If convergence has occurred or the maximum number of iterations has been reached, the iteration is terminated.
[0084] Step S106: decode the output latent vector set by the sub-domain decoder corresponding to the sub-domain encoder, such as Figure 6 and Figure 7 As shown in Figure 2, the temperature latent vector of the target subdomain unit obtained after iteration is decoded using the trained subdomain decoder to obtain the predicted temperature of each subdomain unit at the sampling point.
[0085] Step S107: Map the predicted temperature of each subdomain unit at the sampling point to the entire 3D model. The subdomain units are spliced together, and the temperatures of the sampling points are mapped to the global 3D model, i.e., the entire 3D model, to obtain the overall temperature distribution of the core particle system.
[0086] Step S108, post-processing output: Perform post-processing analysis on the obtained core particle system temperature prediction, extract key hot spot information, and output a prediction report.
[0087] The method for thermal simulation of semiconductor devices according to an embodiment of the present invention can take into account factors such as the presence of a large number of repetitive characteristic structures in the core particle system and the fact that the material parameters and power distribution of the device meet a certain distribution range. Starting from the structure and materials of the basic devices, a data-driven machine learning method is constructed to solve the problem of large-scale, real-time thermal prediction of the core particle system. The core particle system temperature prediction method provided by the embodiment of the present invention, based on a subdomain encoder, a flux conservation autoencoder, and a subdomain decoder, specifically captures the thermal characteristics of the core particle and predicts the global temperature field by learning local thermal characteristic information, regional configuration, and the law of heat flux conservation. Compared with traditional finite element and finite difference methods, it can greatly accelerate simulation analysis, reduce computing power requirements, and meet the iteration speed requirements of thermal information perception in core particle system layout design. At the same time, compared with other purely data-driven "black box" machine learning methods, the temperature prediction based on the autoencoder of this method has better stability, robustness, and accuracy, and can handle more complex geometric structures and a wider range of model parameters.
[0088] The flowchart provided in this embodiment is not intended to indicate that the operations of the method will be performed in any particular order, or that all operations of the method are included in all every case. In addition, the method may include additional operations. Within the scope of the technical ideas provided by the method of this embodiment, additional changes can be made to the above method.
[0089] It should be understood that in some embodiments, each part can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system.
[0090] This embodiment also provides a computer program product 10 , a computer-readable storage medium 20 , and a computer device 30 . Figure 9 is a schematic diagram of a computer program product 10 according to one embodiment of the present invention, Figure 10 is a schematic diagram of a computer-readable storage medium 20 according to one embodiment of the present invention, Figure 11 is a schematic diagram of a computer device 30 according to one embodiment of the present invention. Computer program product 10 includes a computer program 11. When executed by a processor 32, computer program 11 implements the steps of any of the aforementioned methods for thermal simulation of a semiconductor device. Computer-readable storage medium 20 stores computer program 11. When executed by processor 32, computer program 11 implements the steps of any of the aforementioned methods for thermal simulation of a semiconductor device. Computer device 30 may include memory 31, processor 32, and computer program 11 stored in memory 31 and executed by processor 32.
[0091] The computer program 11 for performing the operations of the present invention can be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, configuration data for an integrated circuit, or source code or object code written in any combination of one or more programming languages and procedural programming languages. The computer program 11 can be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any type of network (including a local area network (LAN) or a wide area network (WAN)), or can be connected to an external computer. In some embodiments, to perform various aspects of the present invention, an electronic circuit including, for example, a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA) can execute computer-readable program instructions by utilizing state information of the computer-readable program instructions to personalize the electronic circuit.
[0092] In the description of this embodiment, the computer program product 10 is a related product including the computer program 11 .
[0093] For the purposes of the description of this embodiment, the computer-readable storage medium 20 is a tangible device capable of retaining and storing the computer program 11, and can be any device that can contain, store, communicate, propagate, or transmit the program 11 for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable storage medium 20 include the following: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, and any suitable combination of the foregoing.
[0094] The computer device 30 can be, for example, a server, a desktop computer, a laptop computer, a tablet computer, or a smartphone. In some examples, the computer device 30 can be a cloud computing node. The computer device 30 can be described in the general context of computer system executable instructions (such as program modules) executed by a computer system. Generally, program modules can include routines, programs, object programs, components, logic, data structures, etc. that perform specific tasks or implement specific abstract data types. The computer device 30 can be implemented in a distributed cloud computing environment where remote processing devices linked via a communication network perform tasks. In a distributed cloud computing environment, program modules can be located on local or remote computing system storage media, including storage devices.
[0095] The computer device 30 may include a processor 32 adapted to execute stored instructions, and a memory 31 that provides temporary storage for the instructions during operation. The processor 32 may be a single-core processor, a multi-core processor, a computing cluster, or any number of other configurations. The memory 31 may include random access memory (RAM), read-only memory, flash memory, or any other suitable storage system.
[0096] The computer device 30 may also include a network adapter / interface and an input / output (I / O) interface. The I / O interface allows data to be input and output with external devices that can be connected to the computer device. The network adapter / interface can provide communication between the computer device and a network, which is generally shown as a communication network.
[0097] At this point, those skilled in the art will recognize that, although a number of exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications consistent with the principles of the present invention may be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the present invention. Therefore, the scope of the present invention should be understood and deemed to cover all such other variations or modifications.
Claims
1. A method for thermal simulation of a semiconductor device, characterized in that: include: Constructing a simulation model for describing characteristic information of the semiconductor device, and dividing a plurality of subdomain units from the simulation model; wherein one of every two adjacent subdomain units is a target subdomain unit and the other is a neighboring unit; Encoding the thermal characteristics of each sub-domain unit through a sub-domain encoder to obtain a sampling latent vector group corresponding to each sub-domain unit; Combining the sampled latent vector groups of each target sub-domain unit and the corresponding neighboring units to obtain an input latent vector set corresponding to each target sub-domain unit; Processing the input latent vector set by a flux conservation autoencoder to obtain an output latent vector set; wherein the flux conservation autoencoder at least satisfies a constraint condition of heat flux conservation between the target subdomain unit and the neighboring unit; The output latent vector set is decoded by a subdomain decoder corresponding to the subdomain encoder to obtain a target thermal feature corresponding to the target subdomain unit.
2. The method according to claim 1, characterized in that Also includes: Constructing the flux conservation autoencoder based on a deep neural network architecture; wherein the fully connected layers in the flux conservation autoencoder have at least 4 layers; The flux conservation autoencoder is trained using the input latent vector set, and the fully connected layer is guided to satisfy the constraint condition of heat flux conservation between the target subdomain unit and the neighboring units during the training process.
3. The method according to claim 1, characterized in that The input latent vector set is processed by a flux conservation autoencoder to obtain an output latent vector set, including: The input latent vector set is input into the flux conservation autoencoder for fixed-point iterative processing until the output result of the flux conservation autoencoder converges or the number of iterations of the flux conservation autoencoder reaches a set number, thereby obtaining the output latent vector set; wherein, The sampling latent vector group includes at least a first thermal feature latent vector and a second thermal feature latent vector; The iteration variables in the fixed-point iterative process are the first thermal feature latent vectors of the input latent vector set, while the second thermal feature latent vectors remain unchanged; the first thermal feature latent vectors in the input latent vector set are assigned initial values; The obtained target thermodynamic characteristic corresponding to the target subdomain unit is the first thermodynamic characteristic corresponding to the target subdomain unit.
4. The method according to claim 1, wherein After obtaining the target thermal characteristics corresponding to the target subdomain unit, the method further includes: Taking each of the subdomain units as the target subdomain unit, obtaining target thermal characteristics of each of the subdomain units of the simulation model; Each of the target thermal characteristics is mapped onto the simulation model to obtain information reflecting the distribution of the target thermal characteristics on the semiconductor device.
5. The method according to claim 1, wherein Constructing a simulation model for describing characteristic information of the semiconductor device, including: Constructing a three-dimensional model according to the layout of the semiconductor device, and using the three-dimensional model as the simulation model; and A plurality of subdomain units are divided from the simulation model, including: The simulation model is divided on each coordinate axis according to a Cartesian coordinate system to obtain a plurality of subdomain units; and The neighborhood unit includes at least one in-plane neighborhood unit and / or at least one out-of-plane neighborhood unit, the in-plane neighborhood unit and the target sub-domain unit are adjacent on the X-axis or Y-axis of the Cartesian coordinate system, and the out-of-plane neighborhood unit and the target sub-domain unit are adjacent on the Z-axis of the Cartesian coordinate system.
6. The method according to claim 1, wherein The subdomain units have equal sizes; The target thermal characteristic is temperature; The thermal characteristics include at least temperature, power and thermal conductivity.
7. The method according to claim 1, characterized in that Encoding the thermal characteristics of each subdomain unit by a subdomain encoder includes: respectively acquiring the thermal characteristics from different sampling points within each of the subdomain units; Each of the thermal features is encoded separately by the sub-domain encoder to encode the thermal features into a one-dimensional feature latent vector, and a combination of the one-dimensional feature latent vectors corresponding to the sub-domain encoder is used as the sampling latent vector group corresponding to the sub-domain encoder.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for performing thermal simulation on a semiconductor device according to any one of claims 1 to 7 are implemented.
9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method for performing thermal simulation on a semiconductor device according to any one of claims 1 to 7 are implemented.
10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method for thermal simulation of a semiconductor device according to any one of claims 1 to 7.