A topology optimization method and device for embedded chip cooling

Abstract

The application provides a topology optimization method and device for embedded chip cooling. The application comprises the following steps: obtaining multi-physical field distribution data and material distribution factors, and normalizing to construct a condition characteristic vector; inputting the characteristic vector into a pre-trained generative model, and obtaining a latent variable characteristic increment through low-dimensional latent space single-step denoising reasoning; obtaining a material distribution factor differential evolution amount through a variational autoencoder decoder reconstruction; superimposing the updated material distribution factor and topology configuration, and performing cyclic iteration calculation until convergence. The scheme improves the topology evolution efficiency and multi-physical field adaptability, and optimizes the iteration convergence speed.

Description

Technical Field

[0001] This invention relates to the field of cooling high heat flux density electronic devices, and more particularly to a topology optimization method and apparatus for cooling embedded chips. Background Technology

[0002] Embedded chip cooling, as a core technology for thermal management of high heat flux density electronic devices, is widely used in high-performance computing and artificial intelligence. As chip power consumption approaches the kilowatt level, related technologies have constructed an intelligent design system for microchannel structures through the collaborative operation of pseudo-3D modeling, fluid-thermal coupling simulation, and topology optimization algorithms. Specifically, this system covers the entire process from numerical solution of physical fields to iterative evolution of material distribution, including key aspects such as temperature field monitoring, velocity field analysis, and geometric configuration generation.

[0003] However, existing topology optimization methods directly employ high-frequency sensitivity analysis based on the adjoint matrix and pure numerical iteration paths, without utilizing historical physical field characteristics for intelligent inference acceleration. The lack of adaptive balancing mechanisms for multi-objective weights and prior knowledge mining results in low convergence efficiency and huge resource consumption during the computation process. This makes it difficult to quickly obtain the optimal configuration that balances low thermal resistance and low voltage drop under limited computing power, severely restricting the development cycle and engineering implementation efficiency of chip heat dissipation solutions. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose a topology optimization method for embedded chip cooling.

[0006] Another objective of this invention is to provide a topology optimization device for embedded chip cooling.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention proposes a topology optimization method for embedded chip cooling, comprising:

[0010] S1, Obtain multiphysics distribution data and material allocation factors under the current topology, and normalize the multiphysics distribution data to construct a conditional feature vector; S2, the conditional feature vector is injected into the pre-trained latent diffusion model, and single-step denoising inference is performed in the low-dimensional latent space to predict the latent variable feature increments that characterize the evolution trend of the topological structure. S3, the decoder part of the variational autoencoder is called to reconstruct the latent variable feature increment and restore the material allocation factor differential evolution in physical space; S4. The material allocation factor differential evolution quantity is superimposed with the material allocation factor at the current moment to update the topology configuration, and the updated topology configuration is fed back to the physics solver for the next round of multiphysics calculation until the preset convergence condition is met.

[0011] In one embodiment of the present invention, S1 includes: A pseudo-3D simplified model is used to perform fluid-thermal coupling simulation, and the temperature field, fluid velocity field and material distribution factor of the current iteration step are collected as the original multiphysics field distribution data. The temperature field and the fluid velocity field are respectively subjected to normalization based on global extrema to generate dimensionless temperature field distribution data and velocity field distribution data, wherein the material allocation factor remains unchanged from its original value. The dimensionless temperature field distribution data, the dimensionless velocity field distribution data, and the original material allocation factor are concatenated into a multi-channel tensor, and then mapped into a high-dimensional conditional feature vector through a conditional encoder.

[0012] In one embodiment of the present invention, the normalization processing based on global extrema is performed on the temperature field and the fluid velocity field, respectively, including: Traverse the three-dimensional data matrix with height h, width w, and time t under the current topology. Determine the maximum value in the three-dimensional data matrix. and minimum value ; Using formula For each pixel in the temperature field and fluid velocity field, a point-by-point calculation is performed to obtain the normalized physical quantity field. ,in Represents the field of physical quantities in the SI unit system. .

[0013] In one embodiment of the present invention, S2 includes: The conditional feature vectors are used as strong constraints and injected into the denoising network of the latent diffusion model through a cross-attention mechanism. Inside the denoising network, the latent space features are dynamically adjusted using an adaptive layer normalization mechanism, wherein the scaling factor and offset factor are dynamically generated by the conditional feature vector and the current noise time step through a small multilayer perceptron. A single-step denoising operation is performed based on the adjusted latent space features to predict the noise estimate at the current noise time step, and the noise estimate is transformed into a latent variable feature increment that characterizes the evolution trend of the topological structure.

[0014] In one embodiment of the present invention, S3 includes: During the inference phase, the encoder part of the variational autoencoder is skipped, and the latent variable feature increments output by the latent diffusion model are received directly. The latent variable feature increments are input into a pre-trained multi-layer convolutional decoder to perform upsampling and feature restoration operations, mapping the abstract feature increments in the low-dimensional latent space back to the high-dimensional pixel space. The output is the differential evolution of the material allocation factor in the restored physical space, wherein the edge smoothness of the differential evolution is guaranteed by the reparameterization technique used in the training phase of the variational autoencoder.

[0015] In one embodiment of the present invention, the step of residually superimposing the differential evolution of the material allocation factor with the material allocation factor at the current moment to update the topology configuration includes: Read the material allocation factor distribution of the current iteration step and the differential evolution of the material allocation factor output by the variational autoencoder decoder; Perform residual addition operation to superimpose the material allocation factor differential evolution onto the material allocation factor distribution of the current iteration step to generate the updated topology for the next iteration step; The updated topology is fed back as a new geometric boundary condition to the pseudo-3D physics solver, triggering the next round of multiphysics calculations to obtain new temperature and velocity field distributions, thus forming a closed-loop self-iterative process of physics calculations and AI incremental inference.

[0016] In one embodiment of the present invention, the preset convergence condition includes at least one of the following: The updated topology corresponds to a chip with a maximum temperature lower than a preset threshold, a material allocation factor change rate between two adjacent iterations less than a preset tolerance value, or a cumulative iteration count reaching a preset maximum iteration round. When any of the preset convergence conditions is met, the iteration process is terminated and the final microchannel topology configuration is output. The final microchannel topology configuration is used to guide the fabrication of the physical entity of the microfluidic channel in the silicon-based bonding and packaging process.

[0017] To achieve the above objectives, a second aspect of the present invention provides an embedded chip cooling microchannel topology optimization device, comprising: The data acquisition and normalization module is used to acquire multiphysics distribution data and material allocation factors under the current topology, and to normalize the multiphysics distribution data to construct a conditional feature vector. The latent space denoising inference module is used to inject the conditional feature vector into the pre-trained latent diffusion model, perform single-step denoising inference in the low-dimensional latent space, and predict the latent variable feature increments that characterize the evolution trend of the topological structure. The incremental reconstruction mapping module is used to call the decoder part of the variational autoencoder to reconstruct the feature increment of the latent variable and restore the differential evolution of the material allocation factor in physical space. The topology iteration update module is used to superimpose the differential evolution of the material allocation factor with the material allocation factor at the current moment to update the topology configuration, and feed the updated topology configuration back to the physics solver for the next round of multiphysics calculation until the preset convergence condition is met.

[0018] The present invention provides a topology optimization method and apparatus for embedded chip cooling, which significantly shortens the topology optimization iteration cycle, compresses the design time from hours to seconds, and greatly improves the R&D efficiency of embedded chip cooling structures.

[0019] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing a topology optimization method for embedded chip cooling as described in the first aspect embodiment.

[0020] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a topology optimization method for embedded chip cooling as described in the first aspect embodiment.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] Figure 1 This is a flowchart of a topology optimization method for embedded chip cooling according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a typical pseudo-3D chip liquid cooling problem modeling according to an embodiment of the present invention; Figure 3 This is a diagram showing the final material allocation factor, temperature field, and velocity field distribution of the topology optimization according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the training and inference process of a topology optimization scheme according to an embodiment of the present invention. Figure 5 This is a detailed structural diagram of the diffusion model according to an embodiment of the present invention; Figure 6 This is a complex topological feature structure diagram according to an embodiment of the present invention; Figure 7 This is a structural diagram of a topology optimization device for embedded chip cooling according to an embodiment of the present invention; Figure 8 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] The following description, with reference to the accompanying drawings, describes a topology optimization method and apparatus for embedded chip cooling according to an embodiment of the present invention.

[0026] Figure 1 This is a flowchart of a topology optimization method for embedded chip cooling according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S1, Obtain multiphysics distribution data and material allocation factors under the current topology, and normalize the multiphysics distribution data to construct a conditional feature vector; S2, the conditional feature vector is injected into the pre-trained latent diffusion model, and single-step denoising inference is performed in the low-dimensional latent space to predict the latent variable feature increments that characterize the evolution trend of the topological structure. S3, the decoder part of the variational autoencoder is called to reconstruct the latent variable feature increment and restore the material allocation factor differential evolution in physical space; S4. The material allocation factor differential evolution quantity is superimposed with the material allocation factor at the current moment to update the topology configuration, and the updated topology configuration is fed back to the physics solver for the next round of multiphysics calculation until the preset convergence condition is met.

[0027] This invention proposes another method for optimizing the topology of embedded chip cooling microchannels, including: a multiphysics data-driven topology evolution framework: a topology optimization dataset is constructed using multiphysics simulation tools such as COMSOL. Intermediate step data, such as temperature field, fluid velocity field, and material distribution factor, are collected during the optimization iteration process through normalization processing to establish a mapping relationship from physical field characteristics to structural evolution. The normalization processing step can be expressed as: ,,

[0028] Here, U represents the physical quantity field (temperature, velocity) in SI units. The data source is a topology optimization dataset with dimensions h×w×t, where h and w are the number of pixels occupied by the height and width of the physical field, and t is the time step. Since the material allocation factor itself is a value in the range [0,1], no normalization is required.

[0029] Furthermore, a generative design based on the Latent Diffusion Model (LDM) is employed: A conditionally injected diffusion model is trained to address the nonlinear evolution characteristics of the microchannel topology. This model, conditioned on the current physical field distribution and material allocation factor, predicts the single-step differential evolution of the material allocation factor to achieve intelligent reasoning regarding the direction of topology updates. The specific structure of the diffusion model is as follows... Figure 5 As shown, the external physical field is encoded into high-dimensional conditional features by a conditional encoder, fused with the noise channel, and then passed through temporal injection and a multi-layer fully convolutional network. Finally, a VAE decoder is used to restore it to the expected material allocation factor single-step differential evolution.

[0030] Furthermore, high-dimensional feature extraction and spatial mapping: A variational autoencoder (VAE) is used to reduce the dimensionality of the material allocation factor in topology optimization, extracting its high-dimensional latent variable features in spatial distribution. Diffusion inference is performed in the low-dimensional latent space, effectively capturing the complex topological features of the flow channel configuration while reducing computational overhead. Figure 6 This demonstrates the original pixel space results corresponding to linear interpolation in a high-dimensional feature space for two completely different structures. It shows that the network can learn the topological features of the original image.

[0031] Furthermore, physical randomness and multimodal preservation: By leveraging the probabilistic generation characteristics of the diffusion model, the randomness of physical evolution is preserved during the inference process. Compared to traditional regression prediction models, this technical solution can avoid the "fuzzy averaging" effect among multiple potential optimal modes, thereby generating more innovative microchannel configurations that conform to physical laws. Furthermore, high-efficiency convergence and inference: The method utilizes single-step inference of a pre-trained model to replace the tedious iterative process of solving the incomplete Jacobian matrix based on the adjoint method in traditional topology optimization, significantly shortening the design cycle and improving convergence efficiency.

[0032] Furthermore, the dataset construction process employs a pseudo-3D simplified model for thermal-fluid coupling simulation, which significantly reduces the cost of generating training samples while ensuring the capture of thermal transport characteristics perpendicular to the chip plane.

[0033] Furthermore, the conditional injection mechanism inputs the temperature field T, velocity field V, and material distribution ρ of the current step as multi-channel tensors into the denoising network of the diffusion model, and guides the denoising evolution of the latent space features through adaptive layer normalization.

[0034] Adaptive layer normalization is described as follows:

[0035] Where γ and β are scaling and offset factors, respectively. They are dynamically generated by a small MLP from the condition C and time step t. The denoising network is trained by minimizing the following objective function with respect to the network parameters θ:

[0036] That is, for noise ε distributed according to a standard normal distribution, at a given noise time step t and conditional feature τ(C), its relationship with the predicted ε is... θ The MSE error should be minimized.

[0037] Furthermore, the VAE network structure includes a multi-layer convolutional encoder and decoder. During the training phase, a reparameterization trick is used to ensure the continuity of the latent space distribution, resulting in smooth edges of the generated microchannels and good manufacturing feasibility.

[0038] Furthermore, the single-step prediction task refers to the model directly predicting the material distribution increment in the (n+1)th step or across steps based on the state in the nth step. When used with a multiphysics solver, such as the COMSOL general-purpose multiphysics solver, the optimal topological solution is rapidly approximated through iterative reasoning.

[0039] Furthermore, the microchannel topology generated by the optimized design method is ultimately physically fabricated and integrated using the aforementioned silicon-based bonding and packaging processes (such as silicon-silicon and silicon-glass bonding).

[0040] like Figure 2 The diagram illustrates a typical pseudo-3D chip liquid cooling problem modeled by this software. This model constructs the physical boundary of an embedded microfluidic cooling system, simulating the flow and heat transfer process of the cooling medium at a microscale by integrating a microchannel array on the back of the chip or within the packaging layer. The pseudo-3D modeling method effectively simplifies heat transport calculations perpendicular to the chip plane while maintaining computational accuracy, providing an efficient physical simulation foundation for subsequent high-scale topology optimization.

[0041] like Figure 3 The figure shows the final material distribution factor, temperature field, and velocity field distribution after topology optimization in a case with a central local heat source. The figure visually illustrates the non-intuitive flow channel configuration automatically evolved by the software: in the core region of the heat source, the material distribution factor exhibits a highly integrated branched structure to guide the cooling medium for precise scouring; the corresponding temperature field shows that the hotspot temperature is significantly suppressed, and the velocity field confirms that the optimized flow channel effectively reduces local flow resistance while maintaining high heat transfer efficiency, achieving a dual balance between thermodynamic and hydrodynamic performance.

[0042] like Figure 4 The diagram illustrates the training and inference framework of the topology optimization scheme described in this work. The core of this framework lies in utilizing variational autoencoders (VAEs) and latent diffusion models (LDMs) to achieve intelligent acceleration of the physical evolution laws. Training phase: By collecting intermediate iteration data of multiphysics, VAE is trained to extract high-dimensional latent features of topology and to construct a diffusion model with temperature field (T) and velocity field (V) as conditions.

[0043] Inference Phase: The software employs an asymmetric architecture, where the physical solver outputs real-time field information to guide the LDM in single-step evolution prediction. It only uses the VAE decoder to reconstruct the differential increment (δρ) of the material distribution factor, thus bypassing the extremely time-consuming adjoint matrix solution process in traditional methods. The working principle of this invention is as follows: During software operation, the pseudo-3D physics solver, which is at the core of the process, first assumes the current topological state and calculates the real-time temperature and velocity distributions. This physical field information serves as a strong constraint and is injected into the pre-trained latent diffusion model (LDM) through a cross-attention mechanism.

[0044] LDM performs single-step inference within the latent space, predicting the feature increments required for the current structure to evolve to the optimal state. Subsequently, the system only calls the decoder part of VAE to restore the abstract features of the latent space to the material distribution factor difference component (δρ) in the physical space. This increment is added to the distribution residual of the previous time step to generate the updated topology and feed it back to the physical solver, forming a closed-loop iteration.

[0045] This principle, through the collaborative mechanism of "AI inference-physical verification", not only preserves the randomness of physical evolution and avoids the configuration ambiguity that traditional models are prone to, but also, thanks to the extremely high efficiency of model inference, shortens the iteration process that originally required several hours to the second level, greatly improving the response speed of embedded chip cooling design.

[0046] The embodiments of this invention also have the following technical effects: Significantly improves the efficiency of topology optimization design and reduces the R&D cycle: The method of this invention uses a pre-trained latent diffusion model (LDM) to replace the most time-consuming adjoint method solution process in traditional topology optimization. Accurate prediction of the material allocation factor difference component can be achieved through single-step inference, shortening the complex fluid-thermal coupling iteration process that originally required hours or even days to the second level, greatly improving the design efficiency of embedded chip cooling microchannels and facilitating rapid solution iteration in the early stages of chip design. Accurately captures physical evolution characteristics, ensuring the scientific nature of the heat dissipation configuration: The method of this invention extracts high-dimensional latent features of the structure through a variational autoencoder (VAE) and combines it with a conditional injection mechanism of the temperature and velocity fields, so that the generated topology not only has geometric continuity but also deeply conforms to the physical laws of fluid dynamics and heat transfer. Compared with traditional regression models, this scheme can retain the modal diversity of physical evolution, and the generated microchannels have better targeting and robustness in suppressing local hotspots. Reducing reliance on high-performance computing resources and improving algorithm versatility: The method of this invention adopts an asymmetric VAE-LDM architecture, enabling only the decoder part in the inference stage, and combining it with a pseudo-3D simplified model for physics field verification. This scheme significantly reduces the stringent requirements of large-scale topology optimization on video memory and computing power, making it possible to design high-resolution (e.g., 169x169 mesh) embedded cooling systems on ordinary workstations and even mobile devices, effectively lowering the application threshold of advanced thermal management technologies.

[0047] To achieve the above embodiments, such as Figure 7 As shown, this embodiment also provides a topology optimization device 10 for embedded chip cooling, including: The data acquisition and normalization module 100 is used to acquire multiphysics distribution data and material allocation factors under the current topology, and to normalize the multiphysics distribution data to construct a conditional feature vector. The latent space denoising inference module 200 is used to inject the conditional feature vector into the pre-trained latent diffusion model, perform single-step denoising inference in the low-dimensional latent space, and predict the latent variable feature increments that characterize the evolution trend of the topological structure. The incremental reconstruction mapping module 300 is used to call the decoder part of the variational autoencoder to reconstruct the feature increment of the latent variable and restore the differential evolution of the material allocation factor in physical space. The topology iteration update module 400 is used to superimpose the differential evolution of the material allocation factor with the material allocation factor at the current moment to update the topology configuration, and feed the updated topology configuration back to the physics solver for the next round of multiphysics calculation until the preset convergence condition is met.

[0048] This invention provides a topology optimization device for embedded chip cooling, which significantly shortens the topology optimization iteration cycle, compresses the design time from hours to seconds, and greatly improves the R&D efficiency of embedded chip cooling structures.

[0049] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 8 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the topology optimization method for embedded chip cooling described above.

[0050] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a topology optimization method for embedded chip cooling as described in the foregoing embodiments.

[0051] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0052] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A topology optimization method for embedded chip cooling, characterized in that, include: S1, Obtain multiphysics distribution data and material allocation factors under the current topology, and normalize the multiphysics distribution data to construct a conditional feature vector; S2, the conditional feature vector is injected into the pre-trained latent diffusion model, and single-step denoising inference is performed in the low-dimensional latent space to predict the latent variable feature increments that characterize the evolution trend of the topological structure. S3, the decoder part of the variational autoencoder is called to reconstruct the latent variable feature increment and restore the material allocation factor differential evolution in physical space; S4. The material allocation factor differential evolution quantity is superimposed with the material allocation factor at the current moment to update the topology configuration, and the updated topology configuration is fed back to the physics solver for the next round of multiphysics calculation until the preset convergence condition is met.

2. The method as described in claim 1, characterized in that, S1 includes: A pseudo-3D simplified model is used to perform fluid-thermal coupling simulation, and the temperature field, fluid velocity field and material distribution factor of the current iteration step are collected as the original multiphysics field distribution data. The temperature field and the fluid velocity field are respectively subjected to normalization based on global extrema to generate dimensionless temperature field distribution data and velocity field distribution data, wherein the material allocation factor remains unchanged from its original value. The dimensionless temperature field distribution data, the dimensionless velocity field distribution data, and the original material allocation factor are concatenated into a multi-channel tensor, and then mapped into a high-dimensional conditional feature vector through a conditional encoder.

3. The method as described in claim 2, characterized in that, The normalization process based on global extrema is performed on the temperature field and the fluid velocity field, respectively, including: Traverse the three-dimensional data matrix with height h, width w, and time t under the current topology. Determine the maximum value in the three-dimensional data matrix. and minimum value ; Using formula For each pixel in the temperature field and fluid velocity field, a point-by-point calculation is performed to obtain the normalized physical quantity field. ,in Represents the field of physical quantities in the SI unit system. .

4. The method as described in claim 1, characterized in that, The S2 includes: The conditional feature vectors are used as strong constraints and injected into the denoising network of the latent diffusion model through a cross-attention mechanism. Inside the denoising network, the latent space features are dynamically adjusted using an adaptive layer normalization mechanism, wherein the scaling factor and offset factor are dynamically generated by the conditional feature vector and the current noise time step through a small multilayer perceptron. A single-step denoising operation is performed based on the adjusted latent space features to predict the noise estimate at the current noise time step, and the noise estimate is transformed into a latent variable feature increment that characterizes the evolution trend of the topological structure.

5. The method as described in claim 1, characterized in that, The S3 includes: During the inference phase, the encoder part of the variational autoencoder is skipped, and the latent variable feature increments output by the latent diffusion model are received directly. The latent variable feature increments are input into a pre-trained multi-layer convolutional decoder to perform upsampling and feature restoration operations, mapping the abstract feature increments in the low-dimensional latent space back to the high-dimensional pixel space. The output is the differential evolution of the material allocation factor in the restored physical space, wherein the edge smoothness of the differential evolution is guaranteed by the reparameterization technique used in the training phase of the variational autoencoder.

6. The method as described in claim 1, characterized in that, The step of residually superimposing the differential evolution of the material allocation factor with the material allocation factor at the current moment to update the topology configuration includes: Read the material allocation factor distribution of the current iteration step and the differential evolution of the material allocation factor output by the variational autoencoder decoder; Perform residual addition operation to superimpose the material allocation factor differential evolution onto the material allocation factor distribution of the current iteration step to generate the updated topology for the next iteration step; The updated topology is fed back as a new geometric boundary condition to the pseudo-3D physics solver, triggering the next round of multiphysics calculations to obtain new temperature and velocity field distributions, thus forming a closed-loop self-iterative process of physics calculations and AI incremental inference.

7. The method as described in claim 1, characterized in that, The preset convergence condition includes at least one of the following: The updated topology corresponds to a chip with a maximum temperature lower than a preset threshold, a material allocation factor change rate between two adjacent iterations less than a preset tolerance value, or a cumulative iteration count reaching a preset maximum iteration round. When any of the preset convergence conditions is met, the iteration process is terminated and the final microchannel topology configuration is output. The final microchannel topology configuration is used to guide the fabrication of the physical entity of the microfluidic channel in the silicon-based bonding and packaging process.

8. A topology optimization device for embedded chip cooling, characterized in that, include: The data acquisition and normalization module is used to acquire multiphysics distribution data and material allocation factors under the current topology, and to normalize the multiphysics distribution data to construct a conditional feature vector. The latent space denoising inference module is used to inject the conditional feature vector into the pre-trained latent diffusion model, perform single-step denoising inference in the low-dimensional latent space, and predict the latent variable feature increments that characterize the evolution trend of the topological structure. The incremental reconstruction mapping module is used to call the decoder part of the variational autoencoder to reconstruct the feature increment of the latent variable and restore the differential evolution of the material allocation factor in physical space. The topology iteration update module is used to superimpose the differential evolution of the material allocation factor with the material allocation factor at the current moment to update the topology configuration, and feed the updated topology configuration back to the physics solver for the next round of multiphysics calculation until the preset convergence condition is met.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the topology optimization method for embedded chip cooling as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a topology optimization method for embedded chip cooling as described in any one of claims 1-7.

Images