Method, device and equipment for thermal field prediction of photonic integrated circuit and storage medium

Abstract

Embodiments of the present application provide a photonic integrated circuit thermal field prediction method, device, equipment and storage medium, belonging to the technical field of three-dimensional thermal simulation. The method comprises: obtaining a target heat source power map of a target simulation model and a target spatial coordinate of a target query point; inputting the target heat source power map and the target spatial coordinate into a preset neural network, the preset neural network comprising a target branch network and a target trunk network; encoding the target heat source power map through the target branch network to obtain a target encoded heat source feature, and encoding the target spatial coordinate through the target trunk network to obtain a target encoded space feature; performing feature mapping on the target encoded heat source feature and the target encoded space feature to obtain a target thermal field of the target query point. The embodiments of the present application can improve the thermal field prediction efficiency of the photonic integrated circuit.

Description

Technical Field

[0001] This application relates to the field of three-dimensional thermal simulation technology, and in particular to a method, apparatus, device and storage medium for predicting the thermal field of photonic integrated circuits. Background Technology

[0002] With the widespread application of photonic integrated circuits (PICs) in data communication, high-performance computing, and neural network acceleration, their integration density and scale continue to grow, making thermal management a key factor affecting their performance and reliability. To identify and suppress thermal crosstalk effects in photonic integrated circuits, high-resolution, three-dimensional steady-state thermal simulations are required. Traditional numerical methods, such as the finite difference method (FDM) and the finite element method (FEM), predict the thermal field through fine mesh generation and solving complex partial differential equations of thermal diffusion. However, solving these equations requires significant time and computational resources, resulting in low efficiency in thermal field prediction. Summary of the Invention

[0003] The main objective of this application is to provide a method, apparatus, device, and storage medium for predicting the thermal field of photonic integrated circuits, aiming to improve the efficiency of thermal field prediction for photonic integrated circuits.

[0004] To achieve the above objectives, a first aspect of this application proposes a method for predicting the thermal field of a photonic integrated circuit, the method comprising: Obtain the target heat source power map of the target simulation model and the target spatial coordinates of the target query point; wherein, the target simulation model is a simulation model of a photonic integrated circuit, and both the target simulation model and the target query point are located in a preset simulation area; The target heat source power map and the target spatial coordinates are input into a preset neural network, which includes a target branch network and a target backbone network. The target heat source power map is encoded through the target branch network to obtain the target encoded heat source features, and the target spatial coordinates are encoded through the target backbone network to obtain the target encoded spatial features. The target thermal field of the target query point is obtained by performing feature mapping on the target encoded heat source features and the target encoded spatial features.

[0005] In some embodiments, the target encoded heat source feature includes a heat source feature vector, the target encoded spatial feature includes a spatial feature vector, and the step of performing feature mapping on the target encoded heat source feature and the target encoded spatial feature to obtain the target heat field of the target query point includes: Multiply the heat source feature vector by the corresponding spatial feature vector to obtain the initial temperature distribution; The initial temperature distribution is integrated to obtain the target thermal field.

[0006] In some embodiments, the preset neural network is trained according to the following steps: For each iteration of the first training phase, the first sample heat source power map and the first sample query point spatial coordinates are randomly generated. The initial neural network is trained in corresponding iterations based on the heat source power map of the first sample and the spatial coordinates of the first sample to obtain the reference neural network; For each iteration of the second training phase, the second sample heat source power map and the second sample query point's second sample spatial coordinates are randomly generated. The reference neural network is trained in corresponding iterations based on the second sample heat source power map and the second sample spatial coordinates to obtain the preset neural network.

[0007] In some embodiments, the initial neural network includes an initial branch network and an initial backbone network, and the step of training the initial neural network for corresponding iterations based on the first sample heat source power map and the first sample spatial coordinates to obtain a reference neural network includes: The first sample heat source power map is encoded by the initial branch network to obtain sample encoded heat source features, and the sample encoded spatial features are encoded by the initial backbone network based on the first sample spatial coordinates. By performing feature mapping on the sample encoded heat source features and the sample encoded spatial features, the sample thermal field of the first sample query point is obtained; Based on the sample thermal field, the preset thermal conductivity at the first sample query point, the preset pseudo heat source representation, the preset unit normal vector, the preset convective heat transfer coefficient, and the preset ambient temperature, the loss is calculated to obtain the target loss. The network parameters of the initial neural network in the corresponding iteration round are adjusted according to the target loss to obtain the reference neural network.

[0008] In some embodiments, the step of encoding based on the first sample space coordinates through the initial backbone network to obtain sample coding space features includes: Perform a Fourier transform on the first sample spatial coordinates or the offset spatial coordinates to obtain the transformed spatial coordinates; wherein, the offset spatial coordinates are spatial coordinates obtained by offsetting the first sample spatial coordinates as dummy variables; The transformed spatial coordinates are encoded using the initial backbone network to obtain the sample encoded spatial features.

[0009] In some embodiments, the step of calculating the target loss based on the sample thermal field, a preset thermal conductivity at the first sample query point, a preset pseudo-heat source representation, a preset unit normal vector, a preset convective heat transfer coefficient, and a preset ambient temperature includes: Calculate the steady-state residual loss based on the sample thermal field, the preset thermal conductivity, and the preset pseudo-heat source representation; Calculate the boundary residual loss based on the sample thermal field, the preset thermal conductivity, the preset unit normal vector, the preset convective heat transfer coefficient, and the preset ambient temperature; The target loss is obtained by weighted summation of the steady-state residual loss and the boundary residual loss.

[0010] In some embodiments, the preset pseudo-heat source is calculated according to the following steps: Obtain the normalization constant and the vertical distribution function; The preset pseudo-heat source representation is obtained by multiplying the normalization constant, the power diagram of the first sample heat source, and the vertical distribution function.

[0011] To achieve the above objectives, a second aspect of this application provides a thermal field prediction device for photonic integrated circuits, the device comprising: The acquisition module is used to acquire the target heat source power map of the target simulation model and the target spatial coordinates of the target query point; wherein, the target simulation model is a simulation model of a photonic integrated circuit, and both the target simulation model and the target query point are located in a preset simulation area; The input module is used to input the target heat source power map and the target spatial coordinates into a preset neural network, the preset neural network including a target branch network and a target backbone network; The encoding module is used to encode the target heat source power map through the target branch network to obtain the target encoded heat source features, and to encode the target spatial coordinates through the target backbone network to obtain the target encoded spatial features; The feature mapping module is used to perform feature mapping on the target encoded heat source features and the target encoded space features to obtain the target heat field of the target query point.

[0012] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the thermal field prediction method for photonic integrated circuits described in the first aspect.

[0013] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the thermal field prediction method for photonic integrated circuits described in the first aspect.

[0014] The thermal field prediction method, device, electronic device, and computer-readable storage medium for photonic integrated circuits proposed in this application predict the thermal field of a target query point by acquiring the target heat source power map of the target simulation model and the target spatial coordinates of the target query point. To address the low efficiency of traditional numerical methods for thermal field prediction, a neural network is introduced. The target heat source power map and target spatial coordinates are input into a preset neural network for thermal field prediction, thereby improving the prediction efficiency. The preset neural network includes a target branch network and a target backbone network. The target heat source power map is encoded through the target branch network to encode the heat source field, obtaining the target encoded heat source features. The target spatial coordinates are encoded through the target backbone network to encode spatial points, obtaining the target encoded spatial features. Feature mapping is performed on the target encoded heat source features and the target encoded spatial features to map the heat source field to a temperature field, thus obtaining the target thermal field of the target query point. Attached Figure Description

[0015] Figure 1 This is a flowchart of the thermal field prediction method for photonic integrated circuits provided in the embodiments of this application; Figure 2 This is a flowchart of the training process of the preset neural network provided in the embodiments of this application; Figure 3 yes Figure 2 The flowchart of step S220 in the text; Figure 4 yes Figure 3 The flowchart of step S310 in the process; Figure 5 yes Figure 3 The flowchart of step S330 in the text; Figure 6 This is a flowchart of the calculation process for the preset pseudo-heat source representation provided in the embodiments of this application; Figure 7 yes Figure 1 The flowchart of step S140 in the middle; Figure 8This is a schematic diagram of the thermal field prediction device for photonic integrated circuits provided in the embodiments of this application; Figure 9 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0019] With the widespread application of photonic integrated circuits in data communication, high-performance computing, and neural network acceleration, their integration density and scale continue to grow, making thermal management a key factor affecting their performance and reliability. In many photonic integrated systems, thermal modulation mechanisms are used to control the phase of optical signals, such as in reconfigurable Mach-Zehnder interferometer structures, where microheaters induce refractive index changes to achieve optical path manipulation. However, while these microheaters generate the desired thermal field, they also diffuse heat to adjacent waveguides or optical devices, causing unintended optical phase shifts and severely reducing system stability. This phenomenon is called thermal crosstalk, which is particularly pronounced in complex layouts and high-density designs.

[0020] To identify and suppress thermal crosstalk effects in the design process, high-resolution, three-dimensional steady-state thermal simulations of photonic integrated circuits are required. Traditional numerical methods, such as the finite difference method and the finite element method, obtain high-fidelity temperature distributions through fine mesh generation and solving complex thermal diffusion partial differential equations. However, these methods require resolving the thermal diffusion partial differential equations on the complete three-dimensional geometry, consuming significant time and computational resources. Furthermore, the simulation process is not reusable for each design change, resulting in low design iteration efficiency. Therefore, traditional numerical methods are ill-suited for real-time optimization and large-scale simulation tasks in photonic integrated circuits, leading to low efficiency in thermal field prediction.

[0021] Based on this, embodiments of this application provide a method for predicting the thermal field of photonic integrated circuits, a device for predicting the thermal field of photonic integrated circuits, an electronic device, and a computer-readable storage medium, aiming to improve the efficiency of thermal field prediction for photonic integrated circuits.

[0022] The thermal field prediction method, device, electronic device, and computer-readable storage medium for photonic integrated circuits provided in this application are specifically described through the following embodiments. First, the thermal field prediction method for photonic integrated circuits in this application is described.

[0023] The thermal field prediction method for photonic integrated circuits provided in this application relates to the field of three-dimensional thermal simulation technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the thermal field prediction method for photonic integrated circuits, but is not limited to the above forms.

[0024] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0025] Figure 1 This is an optional flowchart of the thermal field prediction method for photonic integrated circuits provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S110 to S140.

[0026] Step S110: Obtain the target heat source power map of the target simulation model and the target spatial coordinates of the target query point; wherein, the target simulation model is a simulation model of a photonic integrated circuit, and both the target simulation model and the target query point are located in a preset simulation area; Step S120: Input the target heat source power map and target spatial coordinates into a preset neural network. The preset neural network includes a target branch network and a target backbone network. Step S130: Encode the target heat source power map through the target branch network to obtain the target encoded heat source features, and encode the target spatial coordinates through the target backbone network to obtain the target encoded spatial features; Step S140: Perform feature mapping on the target encoded heat source features and the target encoded spatial features to obtain the target heat field of the target query point.

[0027] In step S110 of some embodiments, regarding the high-density micro-heater distribution and cross-layer thermal coupling in the three-dimensional thermal simulation of photonic integrated circuits, this embodiment models the steady-state heat conduction process as a numerical solution problem of the heat diffusion equation, i.e., the steady-state heat conduction partial differential equation. The goal is to quickly predict the steady-state temperature distribution, i.e., the thermal field, of the entire chip space of the photonic integrated circuit under given heat source distribution and chip geometry boundary conditions, providing a foundation for subsequent design optimization and thermal reliability analysis. Specifically, the target heat source power map of the target simulation model and the target spatial coordinates of the target query point are obtained. The target simulation model is a simulation model of the photonic integrated circuit. Both the target simulation model and the target query point are located in a preset simulation region. The target query point is a spatial coordinate point within the preset simulation region, and the target spatial coordinates are the three-dimensional spatial coordinates of the target query point within the preset simulation region. The preset simulation region is the computational region used for three-dimensional steady-state thermal simulation of the photonic integrated circuit. Considering the multilayer heterogeneous material structure in the photonic integrated circuit, the preset simulation region can be modeled as a cuboid computational domain with a multilayer heterogeneous structure. The heat sources of the photonic integrated circuit are concentrated in a specific functional layer containing a micro-heater array within the cuboid computational domain. Robin-type (convection) boundary conditions are applied to all six outer surfaces of the chip, serving as the chip geometry boundary conditions for photonic integrated circuits to simulate heat exchange with the environment. The target heat source power map is a discretized two-dimensional planar power map, representing the lateral power distribution of the heater array. The size of the target heat source power map can be adjusted according to actual conditions; for example, it can be a 120×120 two-dimensional power map.

[0028] If the target spatial coordinates of the target query point are represented as The heat field of the target query point is then represented as .

[0029] In step S120 of some embodiments, thermal simulation predicts the temperature field (thermal field) and its derived thermal properties and thermal reliability indicators by numerically solving the partial differential equation of thermal diffusion. To reduce computational and time resource consumption and ensure the reusability of the thermal simulation process, this application introduces a preset neural network. The target heat source power map and target spatial coordinates are input into the preset neural network for thermal field prediction, addressing the problem that traditional numerical methods cannot adapt to real-time optimization and large-scale simulation tasks in photonic integrated circuits. The preset neural network includes a target branch network and a target backbone network. The target branch network and the target backbone network have the same output dimension, such as both outputting a latent feature representation with a dimension of 256. The target branch network is used to encode the heat source power map, and the target backbone network is used to process the spatial coordinate points (x, y, z). The target branch network is implemented in two variants. The first uses the standard ResNet-18 structure, while the second inserts a self-attention module after Stage 2 of the ResNet-18. This self-attention module captures long-range thermal diffusion correlations and contains four attention heads, each with an output dimension of 32, thus forming a Transformer-enhanced branch network. The introduction of residual connections and self-attention mechanisms enhances the modeling capability for complex thermal field patterns. The target backbone network is implemented as a residual fully connected network.

[0030] Please see Figure 2 In some embodiments, the training process of the preset neural network in step S120 may include, but is not limited to, steps S210 to S240: Step S210: For each iteration of the first training phase, randomly generate the first sample heat source power map and the first sample query point spatial coordinates. Step S220: Train the initial neural network for corresponding iterations based on the first sample heat source power map and the first sample spatial coordinates to obtain the reference neural network; Step S230: For each iteration of the second training phase, randomly generate the second sample heat source power map and the second sample query point's second sample spatial coordinates. Step S240: Train the reference neural network for corresponding iterations based on the second sample heat source power map and the second sample spatial coordinates to obtain the preset neural network.

[0031] In step S210 of some embodiments, during the training process, all models are trained using a loss function driven by physical information, without using any supervised temperature data. The optimization process adopts a two-stage strategy, namely, the training stage includes a first training stage and a second training stage, with the first training stage preceding the second training stage. For each iteration of the first training stage, a two-dimensional planar power map is randomly generated to obtain a first sample heat source power map, and spatial coordinate points are randomly sampled in a preset simulation area to obtain a first sample query point and its first sample spatial coordinates. The first sample spatial coordinates are the three-dimensional spatial coordinates of the first sample query point. The first sample query point is a coordinate point located inside or on the boundary of the preset simulation area, i.e., the first sample query point is an internal point or a boundary point. The number of first sample query points can be set according to the actual situation. For example, 30,000 internal points and 5,000 boundary points can be sampled in each iteration for evaluating the loss term.

[0032] In step S220 of some embodiments, the first training phase uses the Adam optimizer with a learning rate of The maximum number of iterations is 20,000. In each iteration of the first training phase, the initial neural network obtained from the previous iteration is trained based on the first sample heat source power map and the first sample spatial coordinates generated in that iteration, until the last iteration is completed, resulting in the reference neural network. The initial neural network includes an initial branch network and an initial backbone network. The target branch network is trained from the initial branch network, and the target backbone network is trained from the initial backbone network.

[0033] In step S230 of some embodiments, for each iteration of the second training phase, a two-dimensional planar power map is randomly generated to obtain a second sample heat source power map. Spatial coordinate points are then randomly sampled within a preset simulation region to obtain a second sample query point and its second sample spatial coordinates. Compared to the first sample query point, the second sample query point is also a coordinate point located within or on the boundary of the preset simulation region. The second sample spatial coordinates are the three-dimensional spatial coordinates of the second sample query point.

[0034] In step S240 of some embodiments, the second training phase uses the AdamW optimizer with a learning rate of The upper limit for the number of iterations is 80,000. In each iteration of the second training phase, the reference neural network obtained from the previous iteration is trained based on the second sample heat source power map and the second sample spatial coordinates generated in that iteration, until the training of the last iteration is completed, and the preset neural network is obtained.

[0035] Steps S210 to S240 above employ a two-stage training strategy, which can accelerate model convergence and improve model performance. Furthermore, the sampling points are dynamically updated in each iteration round to enhance training diversity and model generalization ability.

[0036] Please see Figure 3 In some embodiments, step S220 may include, but is not limited to, steps S310 to S340: Step S310: Encode the power map of the first sample heat source through the initial branch network to obtain the sample encoded heat source features; and encode the spatial features of the sample based on the spatial coordinates of the first sample through the initial backbone network. Step S320: Perform feature mapping on the sample encoded heat source features and sample encoded spatial features to obtain the sample heat field of the first sample query point; Step S330: Calculate the loss based on the sample thermal field, the preset thermal conductivity at the first sample query point, the preset pseudo-heat source representation, the preset unit normal vector, the preset convective heat transfer coefficient, and the preset ambient temperature to obtain the target loss. Step S340: Adjust the network parameters of the initial neural network in the corresponding iteration round according to the target loss to obtain the reference neural network.

[0037] In step S310 of some embodiments, the first sample heat source power map is input into the initial branch network for heat source field encoding to obtain sample-encoded heat source features. The first sample spatial coordinates are input into the initial backbone network for spatial point encoding to obtain sample-encoded spatial features.

[0038] The initial branch network can use a convolutional residual network (ResNet), or a convolutional residual network (ResNet-18) can be combined with a multi-head self-attention module to enhance the modeling ability of non-smooth heatmaps and achieve efficient encoding of the heat source field (power field). The encoding process of the convolutional residual network can be represented as: , in, Indicates the convolutional residual network of the th The input of each network layer; Indicates the first Each network layer.

[0039] The initial backbone network consists of sequentially connected linear mapping layers, multiple residual fully connected modules, and a linear projection layer. The residual fully connected modules are multi-layer residual sensing structures (Res-MLP), each containing 128 neurons with tanh activation functions for efficient encoding of spatial coordinates. The output dimension of the initial backbone network is consistent with that of the initial branch networks. The encoding process of the residual fully connected modules can be represented as follows: , in, Indicates the initial backbone network. The input of each network layer; and Indicates the first Two convolutional modules are cascaded in a network layer; shortcut represents a residual connection.

[0040] It should be noted that, in addition to the currently used residual multilayer perceptron, convolutional ResNet, and Transformer structures, other neural network architectures can be used as alternatives for the backbone and branch networks. For example, graph neural networks are suitable for modeling local thermal coupling relationships in irregular device structures, sparse convolutional networks help improve the modeling efficiency for sparse heat source representations, and SIREN networks can better express high-frequency gradient changes in the thermal field. These alternative structures can further improve the expressive power and generalization performance of thermal simulation models in different scenarios.

[0041] In step S320 of some embodiments, the sample-encoded heat source features output by the branch network and the sample-encoded spatial features output by the backbone network are fused through bilinear interaction to predict the temperature distribution of the first sample query point, thus obtaining the sample thermal field. The sample-encoded heat source features include multi-dimensional sample heat source feature vectors, and the sample-encoded spatial features include multi-dimensional sample space feature vectors. The feature dimensions of the sample-encoded heat source features and the sample-encoded spatial features are the same. For each dimension, the sample heat source feature vector is multiplied by the corresponding dimension's sample space feature vector to obtain intermediate features. The intermediate features of each dimension are summed to obtain the sample thermal field of the first sample query point.

[0042] In step S330 of some embodiments, the preset thermal conductivity is the material thermal conductivity when establishing a steady-state heat conduction model, and its value is... mW represents milliwatts, mm represents millimeters, and K represents Kelvin. The preset pseudo-heat source is represented as a pre-modeled volumetric heat source function. The preset unit normal vector is the outward-pointing unit normal vector at the spatial coordinate point. The preset convective heat transfer coefficient is a macroscopic measure of the strength of convection-conduction coupling, and its value is... The preset ambient temperature is the actual temperature around the chip, which is 298.15K, or 25℃.

[0043] To optimize model parameters, the target loss is calculated based on the sample thermal field of the first sample query point, the preset thermal conductivity of the first sample query point, the preset pseudo heat source representation, the preset unit normal vector, the preset convective heat transfer coefficient, and the preset ambient temperature, so as to guide the model optimization process based on the target loss.

[0044] In step S340 of some embodiments, the target loss is minimized, and the network parameters of the initial neural network for the corresponding iteration are updated until the model training for the last iteration is completed, resulting in a reference neural network. It should be noted that the model training process in the second training phase can refer to the model training process in the first training phase of steps S310 to S340, and will not be repeated here.

[0045] Through the above steps S310 to S340, the coding process of the heat source field can be optimized, thereby improving the accuracy of heat field prediction.

[0046] Please see Figure 4 In some embodiments, step S310 may include, but is not limited to, steps S410 to S420: Step S410: Perform a Fourier transform on the first sample spatial coordinates or the offset spatial coordinates to obtain the transformed spatial coordinates; wherein, the offset spatial coordinates are the spatial coordinates obtained by offsetting the first sample spatial coordinates as dummy variables; Step S420: Encode the transformation space coordinates through the initial backbone network to obtain the sample encoded space features.

[0047] In step S410 of some embodiments, a Fourier transform is performed on the spatial coordinates of the first sample to obtain the transformed spatial coordinates. To efficiently solve the gradient of the target loss during model training, a zero-coordinate offset automatic differentiation strategy is introduced. Its core idea is to treat the spatial point offset as a dummy variable. The gradient is obtained through nested inverse automatic differentiation. Specifically, the spatial coordinates of the first sample can be shifted as dummy variables to obtain the shifted spatial coordinates, and then the shifted spatial coordinates are subjected to Fourier transform to obtain the transformed spatial coordinates. The Fourier transform can expand the spatial coordinates into 21-dimensional input features, and the frequency of the Fourier transform can be selected as 2π, 4π, or 6π.

[0048] If the coordinates of the first sample space are represented as Then the transformed spatial coordinates are expressed as , This represents a dummy variable.

[0049] In step S420 of some embodiments, the transformed spatial coordinates are input into the initial backbone network for spatial point encoding to obtain sample encoded spatial features.

[0050] It should be noted that the training process in this embodiment is based on physically-informed model training. To efficiently calculate the spatial derivative in solving the physical loss, a zero-coordinate offset automatic differentiation strategy is introduced. This strategy avoids explicit loop structures and significantly reduces the size of the computation graph, thereby improving differentiation efficiency. Specifically, the perturbation function is defined as follows: , in, Indicated based on virtual variables Power diagram of the i-th sample heat source The j-th sample query point The predicted thermal field of the sample; This represents the power graph of the heat source in the i-th sample. The mapping function between the offset spatial coordinates of the j-th sample query point and the sample thermal field.

[0051] The objective derivative can be expressed as: , in, This represents the partial derivative of the sample thermal field with respect to the sample query point; This represents a dummy variable.

[0052] Construct scalar root function The scalar root function is expressed as: , in, This represents the weighting coefficient between the heat source power map of the i-th sample and the query point of the j-th sample.

[0053] By using scalar root functions to perform nested differentiation of the target derivative, the target derivative is transformed into: .

[0054] Through steps S410 to S420, the coordinates of spatial points can be encoded to predict the thermal field based on the encoded spatial features. Furthermore, by introducing a zero-coordinate offset strategy, the memory and computational overhead of automatic differentiation can be reduced, improving training efficiency and numerical stability.

[0055] Please see Figure 5 In some embodiments, step S330 may include, but is not limited to, steps S510 to S530: Step S510: Calculate the steady-state residual loss based on the sample thermal field, preset thermal conductivity, and preset pseudo-heat source representation; Step S520: Calculate the boundary residual loss based on the sample thermal field, preset thermal conductivity, preset unit normal vector, preset convective heat transfer coefficient, and preset ambient temperature. Step S530: The steady-state residual loss and the boundary residual loss are weighted and summed to obtain the target loss.

[0056] In step S510 of some embodiments, to ensure that the temperature prediction conforms to the steady-state heat conduction equation, a physical residual loss based on physical information is introduced during the training process. For each batch of training samples, a set of sampling points is collected at the inner domain and the boundary. This method evaluates physical residual terms such as steady-state partial differential equation (PDE) residuals and boundary condition loss terms based on this point set. Specifically, the steady-state PDE residuals are evaluated based on the sample thermal field, preset thermal conductivity, and preset pseudo-heat source representation at the sample query point, yielding the steady-state residual loss for the first sample query point. The formula for calculating the steady-state residual loss can be expressed as: , in, This represents the steady-state residual loss; This represents the j-th sample query point; Indicates the preset thermal conductivity; Indicates the thermal field of the sample; Represents the gradient; This indicates a preset pseudo-heat source representation.

[0057] It should be noted that the steady-state residual loss is calculated based on internal points. The selection of internal points employs an importance sampling strategy, prioritizing points that significantly influence the thermal field prediction, such as those near high-power heat sources, those with drastic changes in material thermal conductivity, or those with large temperature gradients. This improves the training efficiency and physical convergence of the residual constraints. These points are more sensitive to temperature changes, providing stronger supervisory signals and effectively promoting the model's learning of local strongly nonlinear thermal diffusion behavior.

[0058] In step S520 of some embodiments, Robin boundary residual term evaluation is performed based on the sample thermal field, preset thermal conductivity, preset unit normal vector, preset convective heat transfer coefficient, and preset ambient temperature to obtain the boundary residual loss of the first sample query point. The first sample query point is the boundary point when calculating the boundary residual loss. The formula for calculating the boundary residual loss is expressed as: , in, Represents boundary residual loss; n represents the preset unit normal vector; h represents the preset convective heat transfer coefficient; This indicates the preset ambient temperature.

[0059] In step S530 of some embodiments, the mean of the squares of the steady-state residual loss for each first sample query point is calculated to obtain a first mean. The mean of the squares of the boundary residual loss for each first sample query point is calculated to obtain a second mean. The first mean and the second mean are weighted and calculated to obtain the target loss. The formula for calculating the target loss is expressed as: , in, Indicates target loss; and These represent the first mean and its weight coefficient, respectively; and These represent the second mean and its weighting coefficient, respectively.

[0060] It should be noted that during the training process driven by physical information, a conflict-free incremental gradient training strategy (ConFIG) can be adopted. By eliminating gradient conflicts during the multi-objective loss optimization process, it ensures that the dot product between the final update direction and the gradient direction of each sub-loss term is positive, thereby achieving consistent optimization speed among loss terms.

[0061] Through the above steps S510 to S530, physical loss can be obtained, and model parameters can be optimized based on physical loss to achieve efficient encoding of heat source field and spatial coordinate points.

[0062] Please see Figure 6 In some embodiments, the calculation process for the preset pseudo-heat source representation may include, but is not limited to, steps S610 to S620: Step S610: Obtain the normalization constant and the vertical distribution function; Step S620: Multiply the normalization constant, the power diagram of the first sample heat source, and the vertical distribution function to obtain the preset pseudo heat source representation.

[0063] In step S610 of some embodiments, considering the multilayer heterogeneous material structure in the photonic integrated circuit, a true three-dimensional heat source is expressed. The high dimensionality leads to significant computational complexity during training. Therefore, this application employs a pseudo-3D heat source strategy, using a pseudo-3D heat source expression to model the volumetric heat source term and modeling the volumetric heat source function as a two-dimensional horizontal heat source diagram. With vertical distribution function The product of these two factors, to simultaneously consider the lateral distribution and vertical stratification characteristics of the heat injection, takes the following form: , in, This is a preset pseudo-heat source representation; This is a normalization constant to ensure that the total power remains consistent. This represents the heat source power diagram, which is a normalized two-dimensional planar power diagram used to indicate the lateral power distribution of the heater array; It is the vertical distribution function, i.e., the normalized longitudinal profile function, used to reflect the intralayer response of different process layers to heat injection; Represents three-dimensional spatial coordinates.

[0064] The specific form of the normalization constant is: , in, Indicates total power; express The plane integral region covered; express The range of integration; It represents the multiplication operation.

[0065] To balance physical realism and network trainability, Using various Lorentz-type functions, it can be expressed as: , in, Z-axis spatial coordinates; Indicates the center coordinates of the z-axis; Indicates shape parameters.

[0066] It should be noted that, in the pseudo-three-dimensional representation of heat sources, in addition to the proposed Lorentz-type function form, other distribution functions, such as Gaussian distribution, uniform distribution, or custom piecewise functions, can also be used to flexibly adapt to the actual vertical distribution characteristics of heat sources in different photonic chip structures, thereby achieving wider applicability and the ability to model thermal diffusion behavior under different manufacturing conditions.

[0067] When representing data, point cloud distribution can be used to sparsely sample and represent heat sources and temperature fields, thereby significantly reducing input dimensions and memory overhead, further reducing computational resource consumption and improving model inference efficiency.

[0068] Specifically, obtain the normalization constant C and the vertical distribution function. .

[0069] In step S620 of some embodiments, the normalization constant C and the first sample heat source power diagram are... and vertical distribution function Multiplying these together yields the pre-defined pseudo-heat source representation.

[0070] Steps S610 to S620 above, by introducing a pseudo-three-dimensional heat source modeling method, can effectively compress the simulation input space dimension while maintaining the interlayer heat diffusion modeling capability and thermal coupling accuracy.

[0071] In step S130 of some embodiments, after model training is completed and optimal parameters are loaded, the target heat source power map is encoded using a target branch network to obtain target encoded heat source features. The target spatial coordinates of the target query point are then encoded using a target backbone network to obtain target encoded spatial features. Thermal field inference is then performed based on the target encoded heat source features and the target encoded spatial features. The target encoded heat source features include multi-dimensional heat source feature vectors, and the target encoded spatial features include multi-dimensional spatial feature vectors. The target encoded heat source features and the target encoded spatial features have the same feature dimensions.

[0072] Please see Figure 7 In some embodiments, step S140 may include, but is not limited to, steps S710 to S720.

[0073] Step S710: Multiply the heat source feature vector and the corresponding spatial feature vector to obtain the initial temperature distribution; Step S720: Integrate the initial temperature distribution to obtain the target thermal field.

[0074] In step S710 of some embodiments, to adapt to chip-level thermal modeling requirements, the target heat source power map is... Input to the target branch network, and obtain the output of the target branch network. Each dimension of heat source feature vectors constitutes a vector group. , This represents the feature vector of the heat source in the k-th dimension. The target space coordinates... Input is fed into the target backbone network, and the output of the target backbone network is obtained. Spatial feature vectors of various dimensions; these spatial feature vectors constitute a vector group. , This represents the spatial feature vector of the k-th dimension. For each dimension, the heat source feature vector... and corresponding spatial feature vectors Multiplying these values ​​yields the initial temperature distribution of the target query point in that dimension.

[0075] In step S720 of some embodiments, the initial temperature distributions of each dimension are summed to obtain the target query point. The predicted temperature value. If only one target query point exists, this predicted temperature value is used as the target thermal field. The target thermal field is a three-dimensional temperature field. The formula for calculating the predicted temperature value is expressed as: , in, Indicates the target query point Predicted temperature value at the location; This represents a structure-aware operator used to visualize the power map of a target heat source. and target query point The target space coordinates are mapped to the temperature field; Indicates feature dimension; This represents the feature vector of the heat source in the k-th dimension; This represents the spatial feature vector of the k-th dimension.

[0076] If there are multiple target query points, then for each target query point Batch execution of steps S710 to S720 can quickly obtain a three-dimensional temperature field. .

[0077] Through the above steps S710 to S720, the temperature field at any query point in the preset simulation area can be quickly predicted, improving the efficiency of thermal field prediction.

[0078] To address the thermal crosstalk challenges caused by high-density microheater arrays and cross-layer thermal coupling, this application employs a pseudo-3D heat source representation. A volumetric heat source is constructed by multiplying a 2D heater power map by a normalized longitudinal profile function, effectively compressing the input dimension while preserving interlayer diffusion information. Simultaneously, automatic differentiation using zero-coordinate offset is employed to synchronously acquire spatial derivatives during forward inference, reducing the computational overhead of partial differential equation loss assessment and significantly improving the efficiency and numerical stability of residual evaluation. This application, while maintaining prediction accuracy, reduces GPU memory overhead and training time, significantly accelerating the thermal analysis process of photonic integrated circuits.

[0079] Please see Figure 8 This application also provides a thermal field prediction device for photonic integrated circuits, which can implement the above-described thermal field prediction method for photonic integrated circuits. The thermal field prediction device for photonic integrated circuits includes: The acquisition module 810 is used to acquire the target heat source power map of the target simulation model and the target spatial coordinates of the target query point; wherein, the target simulation model is a simulation model of a photonic integrated circuit, and both the target simulation model and the target query point are located in a preset simulation area; The input module 820 is used to input the target heat source power map and target spatial coordinates into a preset neural network, which includes a target branch network and a target backbone network. The encoding module 830 is used to encode the target heat source power map through the target branch network to obtain the target encoded heat source features, and to encode the target spatial coordinates through the target backbone network to obtain the target encoded spatial features; The feature mapping module 840 is used to perform feature mapping on the target encoded heat source features and the target encoded space features to obtain the target heat field of the target query point.

[0080] The specific implementation of the thermal field prediction device for the photonic integrated circuit is basically the same as the specific implementation of the thermal field prediction method for the photonic integrated circuit described above, and will not be repeated here.

[0081] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described thermal field prediction method for photonic integrated circuits. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0082] Please see Figure 9 , Figure 9 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 910 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 920 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 920 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 920 and called by the processor 910 to execute the thermal field prediction method of the photonic integrated circuit in the embodiments of this application. The input / output interface 930 is used to implement information input and output; The communication interface 940 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 950 transmits information between various components of the device (e.g., processor 910, memory 920, input / output interface 930, and communication interface 940); The processor 910, memory 920, input / output interface 930 and communication interface 940 are connected to each other within the device via bus 950.

[0083] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described thermal field prediction method for photonic integrated circuits.

[0084] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0085] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0086] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0087] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0088] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0089] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0090] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0091] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0092] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0093] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0094] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for predicting the thermal field of a photonic integrated circuit, characterized in that, The method includes: Obtain the target heat source power map of the target simulation model and the target spatial coordinates of the target query point; wherein, the target simulation model is a simulation model of a photonic integrated circuit, and both the target simulation model and the target query point are located in a preset simulation area; The target heat source power map and the target spatial coordinates are input into a preset neural network, which includes a target branch network and a target backbone network. The target heat source power map is encoded through the target branch network to obtain the target encoded heat source features, and the target spatial coordinates are encoded through the target backbone network to obtain the target encoded spatial features. The target thermal field of the target query point is obtained by performing feature mapping on the target encoded heat source features and the target encoded spatial features.

2. The method according to claim 1, characterized in that, The target encoded heat source feature includes a heat source feature vector, and the target encoded spatial feature includes a spatial feature vector. The step of performing feature mapping on the target encoded heat source feature and the target encoded spatial feature to obtain the target heat field of the target query point includes: Multiply the heat source feature vector by the corresponding spatial feature vector to obtain the initial temperature distribution; The initial temperature distribution is integrated to obtain the target thermal field.

3. The method according to claim 1, characterized in that, The preset neural network is trained according to the following steps: For each iteration of the first training phase, the first sample heat source power map and the first sample query point spatial coordinates are randomly generated. The initial neural network is trained in corresponding iterations based on the heat source power map of the first sample and the spatial coordinates of the first sample to obtain the reference neural network; For each iteration of the second training phase, the second sample heat source power map and the second sample query point's second sample spatial coordinates are randomly generated. The reference neural network is trained in corresponding iterations based on the second sample heat source power map and the second sample spatial coordinates to obtain the preset neural network.

4. The method according to claim 3, characterized in that, The initial neural network includes an initial branch network and an initial backbone network. The step of training the initial neural network iteratively for corresponding rounds based on the first sample heat source power map and the first sample spatial coordinates to obtain a reference neural network includes: The first sample heat source power map is encoded by the initial branch network to obtain sample encoded heat source features, and the sample encoded spatial features are encoded by the initial backbone network based on the first sample spatial coordinates. By performing feature mapping on the sample encoded heat source features and the sample encoded spatial features, the sample thermal field of the first sample query point is obtained; Based on the sample thermal field, the preset thermal conductivity at the first sample query point, the preset pseudo heat source representation, the preset unit normal vector, the preset convective heat transfer coefficient, and the preset ambient temperature, the loss is calculated to obtain the target loss. The network parameters of the initial neural network in the corresponding iteration round are adjusted according to the target loss to obtain the reference neural network.

5. The method according to claim 4, characterized in that, The step of encoding based on the first sample space coordinates through the initial backbone network to obtain sample encoding space features includes: Perform a Fourier transform on the first sample spatial coordinates or the offset spatial coordinates to obtain the transformed spatial coordinates; wherein, the offset spatial coordinates are spatial coordinates obtained by offsetting the first sample spatial coordinates as dummy variables; The transformed spatial coordinates are encoded using the initial backbone network to obtain the sample encoded spatial features.

6. The method according to claim 4, characterized in that, The loss calculation based on the sample thermal field, the preset thermal conductivity at the first sample query point, the preset pseudo-heat source representation, the preset unit normal vector, the preset convective heat transfer coefficient, and the preset ambient temperature yields the target loss, including: Calculate the steady-state residual loss based on the sample thermal field, the preset thermal conductivity, and the preset pseudo-heat source representation; Calculate the boundary residual loss based on the sample thermal field, the preset thermal conductivity, the preset unit normal vector, the preset convective heat transfer coefficient, and the preset ambient temperature; The target loss is obtained by weighted summation of the steady-state residual loss and the boundary residual loss.

7. The method according to claim 4, characterized in that, The preset pseudo-heat source is calculated according to the following steps: Obtain the normalization constant and the vertical distribution function; The preset pseudo-heat source representation is obtained by multiplying the normalization constant, the power diagram of the first sample heat source, and the vertical distribution function.

8. A thermal field prediction device for photonic integrated circuits, characterized in that, The device includes: The acquisition module is used to acquire the target heat source power map of the target simulation model and the target spatial coordinates of the target query point; wherein, the target simulation model is a simulation model of a photonic integrated circuit, and both the target simulation model and the target query point are located in a preset simulation area; The input module is used to input the target heat source power map and the target spatial coordinates into a preset neural network, the preset neural network including a target branch network and a target backbone network; The encoding module is used to encode the target heat source power map through the target branch network to obtain the target encoded heat source features, and to encode the target spatial coordinates through the target backbone network to obtain the target encoded spatial features; The feature mapping module is used to perform feature mapping on the target encoded heat source features and the target encoded space features to obtain the target heat field of the target query point.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images