Intelligent thermal simulation method for monolithic three-dimensional integrated chip

Abstract

The invention relates to an intelligent thermal simulation method for a monolithic three-dimensional integrated chip, belongs to the technical field of monolithic three-dimensional integrated chips, and solves the problems of long thermal simulation period and large operation memory space in the prior art. The thermal simulation method comprises the following steps: layering a monolithic three-dimensional integrated chip, and constructing a coordinate system of the monolithic three-dimensional integrated chip and physical equivalent information of each layer; measuring and acquiring a plurality of training samples; measuring the coordinate points of the heat source and the power of each heat source coordinate point at different time points to form heat source power data; according to the plurality of training samples, in combination with the physical equivalent information and the heat source power data of each layer, training a physical neural network (PINN) to obtain a thermal simulation model of the monolithic three-dimensional integrated chip; and inputting all the coordinate points of the monolithic three-dimensional integrated chip and all the prediction time points of each coordinate point into the thermal simulation model to obtain the prediction temperature of each coordinate point at each prediction time point. And rapid thermal simulation of the monolithic three-dimensional integrated chip is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of monolithic three-dimensional integrated chip design, and in particular to an intelligent thermal simulation method for a monolithic three-dimensional integrated chip. Background Art

[0002] Entering the post-Moore's era, device integration faces huge challenges in physical principles and process technology. Monolithic three-dimensional integrated chips have become a hot topic in the industry. This technology is expected to continue Moore's Law and develop higher-performance chip products with higher integration density, smaller area and larger bandwidth.

[0003] With the substantial increase in heterogeneous integration, power consumption density and signal transmission rate, thermal reliability issues have brought severe challenges to M3D (Monolithic 3-Dimensional Integration) technology, greatly affecting the performance and stability of M3D chips.

[0004] In the prior art, thermal simulation of M3D chips usually adopts numerical solution methods to predict the heat transport characteristics, which often faces problems such as long simulation cycle and large computing and storage volume, and it is difficult to meet the efficiency requirements of thermal design and management of M3D chips.

[0005] Therefore, there is an urgent need for a technical solution that can quickly perform thermal simulation on M3D chips. Summary of the invention

[0006] In view of the above analysis, an embodiment of the present invention aims to provide a single-chip three-dimensional integrated chip intelligent thermal simulation method to solve the problems of long thermal simulation cycle and large computing and storage capacity in the prior art.

[0007] An embodiment of the present invention provides a single-chip three-dimensional integrated chip intelligent thermal simulation method, the thermal simulation method comprising:

[0008] Acquire structural information of a monolithic three-dimensional integrated chip, layer the monolithic three-dimensional integrated chip, and construct a coordinate system of the monolithic three-dimensional integrated chip and physical equivalent information of each layer;

[0009] Measure and obtain the temperatures of multiple coordinate points of a single three-dimensional integrated chip at different time points, and use the temperature of each coordinate point at a time point as a training sample to obtain multiple training samples; measure the coordinate points of the heat source and the power of each heat source coordinate point at different time points to form heat source power data;

[0010] Based on multiple training samples, combined with the physical equivalent information of each layer and the heat source power data, the physical neural network PINN is trained to obtain a thermal simulation model of a single 3D integrated chip.

[0011] All coordinate points of the monolithic three-dimensional integrated chip and all predicted time points of each coordinate point are input into the thermal simulation model to obtain the predicted temperature of each coordinate point at each predicted time point.

[0012] Based on further improvement of the above thermal simulation method, the physical equivalent information of each layer includes one or more of the following:

[0013] Equivalent thermal conductivity;

[0014] Equivalent specific heat capacity;

[0015] Equivalent mass density.

[0016] Based on the further improvement of the above thermal simulation method, the physical neural network PINN is trained according to multiple training samples, combined with the physical equivalent information of each layer and all heat source power data, to obtain a thermal simulation model of a monolithic three-dimensional integrated chip, including:

[0017] Generate multiple internal coordinate training data and multiple boundary coordinate training data required for this iteration; each training data includes a coordinate point and a predicted time point;

[0018] Inputting a plurality of internal coordinate training data, a plurality of boundary coordinate training data and a plurality of training samples into the thermal simulation model respectively to obtain corresponding predicted temperature data;

[0019] Calculate the loss function value of this iteration based on the physical equivalent information of each layer, all heat source power data and the predicted temperature data;

[0020] It is determined whether the loss function value of this iteration meets the loss threshold; if it meets the loss threshold, the training process is terminated to obtain a trained thermal simulation model; otherwise, the next iteration process is started.

[0021] Based on the further improvement of the above thermal simulation method, the physical neural network calculates the predicted temperature data through the following formula:

[0022]

[0023] Among them, T i represents the predicted temperature corresponding to each input data, k represents the number of hidden layers, Q represents the width of the hidden layer, σ represents the activation function, Represents neural network composite operation, NN k (z (k-1) ) indicates the k-th layer neural network operation. represents the weight matrix of the kth layer of the neural network, represents the bias parameter vector of the kth layer of the neural network, z (k-1) represents the temperature distribution vector of the k-1th layer of the neural network, Represents the temperature distribution value of the Q-th width of the k-1th layer of the neural network.

[0024] Based on the further improvement of the above thermal simulation method, the loss function of the thermal simulation model is obtained through the following steps:

[0025] Determining the heat conduction equation and boundary condition equation of the monolithic three-dimensional integrated chip, and determining the heat conduction loss function and boundary condition loss function;

[0026] Set a first weight value, a second weight value, and a third weight value corresponding to the heat conduction loss function, the boundary condition loss function, and the training sample loss function;

[0027] The loss function of the thermal simulation model is obtained according to the heat conduction loss function, the boundary condition loss function, the training sample loss function, the first weight value, the second weight value and the third weight value.

[0028] Based on the further improvement of the above thermal simulation method, the heat conduction equation is:

[0029]

[0030] Among them, ρ(x, y, z, t) represents the equivalent mass density of the coordinate point (x, y, z) at time point t, C p (x, y, z, t) represents the equivalent specific heat capacity of the coordinate point (x, y, z) at time point t, T(x, y, z, t) represents the predicted temperature of the coordinate point (x, y, z) at time point t, k(x, y, z, t) represents the equivalent thermal conductivity of the coordinate point (x, y, z) at time point t, ▽ represents the Nebra operator, M represents the number of heat sources inside the monolithic three-dimensional integrated chip, S represents the heat source, g s (x0, y0, z0, t) represents the heat source power of the heat source S with coordinate point (x0, y0, z0) at time point t.

[0031] Based on the further improvement of the above thermal simulation method, the boundary condition equation is:

[0032]

[0033] Wherein, h represents the convection heat transfer coefficient, f(x, y, z, t) represents the ambient temperature of the boundary coordinate point (x, y, z) at time point t, and n is the normal vector of the boundary coordinate point (x, y, z).

[0034] Based on the further improvement of the above thermal simulation method, the loss function of the thermal simulation model is:

[0035] Loss = w1L1 + w2L2 + w3L3;

[0036]

[0037] Among them, Loss, L1, L2, L2 represent the loss function of the thermal simulation model, the heat conduction loss function, the boundary condition loss function, and the training sample loss function, respectively; w1, w2, w3 represent the first weight value, the second weight value, and the third weight value, respectively; N1, N2, N3 represent the number of internal coordinate points, the number of boundary coordinate points, and the number of training samples, respectively; T i0 Represents the temperature of the coordinate point in the training sample, T i represents the predicted temperature, f i Represents the ambient temperature of the i-th boundary coordinate point at time point t.

[0038] Based on the further improvement of the above thermal simulation method, after obtaining the predicted temperature of each coordinate point at each time point, the thermal simulation method further includes:

[0039] The coordinate range of the monolithic three-dimensional integrated chip is numerically discretized according to the preset coordinate optimization step and time optimization step, and the coefficient matrix of each coordinate point at each time point is determined;

[0040] Based on the least squares method, the projection matrix of each coordinate point at each time point is determined according to the coefficient matrix of each coordinate point at each time point;

[0041] The predicted temperature is updated according to the projection matrix of each coordinate point at each time point to obtain the final predicted temperature.

[0042] Based on the further improvement of the above thermal simulation method, the coefficient matrix of each coordinate point at each time point is:

[0043]

[0044] Among them, A represents the coefficient matrix of each coordinate point at each time point, ρ represents the equivalent mass density at each coordinate point, and C p represents the equivalent specific heat capacity at each coordinate point, Δt represents the time optimization step, Δx, Δy, Δz represent the coordinate optimization step of the X-axis, Y-axis, and Z-axis respectively, i, j, k represent the sequence number of each coordinate point in the X-axis, Y-axis, and Z-axis after numerical discretization, and κ represents the equivalent thermal conductivity after numerical discretization.

[0045] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0046] 1. By layering the monolithic 3D integrated chip and measuring a small number of training samples, the PINN is trained based on the training samples to obtain a thermal simulation model of the monolithic 3D integrated chip. The thermal simulation of the monolithic 3D integrated chip is performed according to the thermal simulation model, and the simulation data of the monolithic 3D integrated chip can be quickly obtained;

[0047] 2. In the process of training PINN, the use of randomly generated internal coordinate training data and boundary coordinate training data can increase the number of coordinates required for loss function calculation, improve the accuracy of the loss function value, and make the thermal simulation data generated by the trained thermal simulation model more reliable;

[0048] 3. Setting the corresponding first weight value, second weight value and third weight value for the heat conduction loss function, boundary condition loss function and training sample loss function can reasonably measure the loss value of physical laws and experimental data, and improve the authenticity of thermal simulation data generated by the thermal simulation model.

[0049] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0051] Figure 1 A schematic diagram of a process flow of a single-chip three-dimensional integrated chip intelligent thermal simulation method provided by an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of the structure of a monolithic three-dimensional integrated chip provided by an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of the composition structure of a layered monolithic three-dimensional integrated chip provided by an embodiment of the present invention;

[0054] Figure 4 A schematic diagram of the structure of a PINN provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0056] A specific embodiment of the present invention discloses a single-chip three-dimensional integrated chip intelligent thermal simulation method, such as Figure 1 As shown, the thermal simulation method includes:

[0057] Step S1: obtaining structural information of a monolithic three-dimensional integrated chip, layering the monolithic three-dimensional integrated chip, and constructing a coordinate system of the monolithic three-dimensional integrated chip and physical equivalent information of each layer;

[0058] Step S2: measuring and obtaining the temperatures of multiple coordinate points of the monolithic three-dimensional integrated chip at different time points, taking the temperature of each coordinate point at a time point as a training sample, and obtaining multiple training samples; measuring the coordinate points of the heat source, and the power of each heat source coordinate point at different time points, to form heat source power data;

[0059] Step S3: According to a plurality of training samples, combined with the physical equivalent information of each layer and the heat source power data, the physical neural network PINN is trained to obtain a thermal simulation model of a single three-dimensional integrated chip;

[0060] Step S4: input all coordinate points of the monolithic three-dimensional integrated chip and all predicted time points of each coordinate point into the thermal simulation model to obtain the predicted temperature of each coordinate point at each predicted time point.

[0061] Specifically, Figure 2 As shown, the composition structure of the monolithic three-dimensional integrated chip includes core particle 1 (Tier1), core particle 2 (Tier2), MIV (interlayer via) and handle body layer (Handle Bulk), and ILD (SiO2). In addition, the monolithic three-dimensional integrated chip also includes various other materials, and the structure is relatively complex.

[0062] Specifically, Figure 1 As shown, in step S1, the fine microstructure information of the monolithic three-dimensional integrated chip is extracted through the design information of the monolithic three-dimensional integrated chip, such as the GDS layout file and the process file, and the monolithic three-dimensional integrated chip is layered. The number of layers can be reasonably set. If the accuracy required for thermal simulation is high, more layers are used. After the layering, the structural information of each layer in the monolithic three-dimensional integrated chip is determined.

[0063] Specifically, Figure 3 As shown, the monolithic three-dimensional integrated chip after layering.

[0064] It is worth noting that since each layer contains a large number of materials of different sizes and structures, it is very difficult to accurately extract the physical properties of each layer of material, and it is also extremely difficult to accurately calculate the heat transfer characteristics of each material. Therefore, the present invention equates the physical properties of each layer and determines the physical equivalent information of each layer.

[0065] In order to facilitate thermal simulation of a monolithic three-dimensional integrated chip, a coordinate system of the monolithic three-dimensional integrated chip is constructed so that the coordinate range of each layer can be determined.

[0066] Specifically, Figure 3 As shown, the coordinate system of the monolithic three-dimensional integrated chip can be a three-dimensional rectangular coordinate system, with the center of the monolithic three-dimensional integrated chip as the origin, and three directions selected on the basis of the origin as the X, Y, and Z axes.

[0067] Preferably, the physical equivalent information of each layer includes one or more of the following:

[0068] Equivalent thermal conductivity;

[0069] Equivalent specific heat capacity;

[0070] Equivalent mass density.

[0071] Specifically, thermal conductivity is a measure of a material's ability to conduct heat; for a monolithic three-dimensional integrated chip, specific heat capacity is an important thermophysical parameter because it affects the accumulation and dissipation of heat during chip operation.

[0072] Specifically, determine the equivalent thermal conductivity of each layer and the total volume of each layer; determine the volumes of multiple design graphics included in each layer and the thermal conductivity, specific heat capacity and mass density of each design graphic, and determine the average thermal conductivity, average specific heat capacity and average mass density as the equivalent thermal conductivity, equivalent specific heat capacity and equivalent mass density of each layer.

[0073] Specifically, Figure 2 and Figure 3 As shown, each layer includes a plurality of small rectangular design patterns and other areas. The other areas are composed of the same material. Different design patterns represent different materials, such as Figure 2 The small rectangular blocks in the circuit can be made of metal copper or other interconnect materials such as metal tungsten.

[0074] The volume, thermal conductivity, specific heat capacity and mass density of each design graphic, as well as the volume, thermal conductivity, specific heat capacity and mass density of other areas are determined from the GDS layout file and process file of the monolithic three-dimensional integrated chip, so as to calculate the equivalent thermal conductivity, equivalent specific heat capacity and equivalent mass density of each layer.

[0075] Specifically, Figure 1 As shown, in step S1, the layers of the monolithic three-dimensional integrated chip that needs to be thermally simulated, as well as the coordinate information and equivalent physical information of each layer are determined. Specifically, in step S2, multiple training samples are obtained through measurement to train the thermal simulation model.

[0076] Specifically, in step S2, multiple coordinate points can be selected as reference points for measurement, and different time points can be set at the same time, and the temperature of a coordinate point at a time point can be used as a training sample, so as to obtain multiple training samples through measurement. At the same time, combined with the coordinate system of the monolithic three-dimensional integrated chip, the coordinate points of the heat sources of the monolithic three-dimensional integrated chip and the power of each heat source at different time points are determined as heat source power data.

[0077] Specifically, Figure 1 As shown, in step S3, the physical neural network PINN is trained according to the multiple training samples obtained in step S2, and the trained PINN is used as the thermal simulation model of the monolithic three-dimensional integrated chip.

[0078] Preferably, the physical neural network PINN is trained based on multiple training samples, combined with the physical equivalent information of each layer and all heat source power data, to obtain a thermal simulation model of a monolithic three-dimensional integrated chip, including:

[0079] Generate multiple internal coordinate training data and multiple boundary coordinate training data required for this iteration; each training data includes a coordinate point and a predicted time point;

[0080] Inputting a plurality of internal coordinate training data, a plurality of boundary coordinate training data and a plurality of training samples into the thermal simulation model respectively to obtain corresponding predicted temperature data;

[0081] Calculate the loss function value of this iteration based on the physical equivalent information of each layer, all heat source power data and the predicted temperature data;

[0082] It is determined whether the loss function value of this iteration meets the loss threshold; if it meets the loss threshold, the training process is terminated to obtain a trained thermal simulation model; otherwise, the next iteration process is started.

[0083] Specifically, when training PINN, multiple iterations are required. In each iteration, multiple internal coordinate training data and multiple boundary coordinate training data required for each iteration need to be generated, where the training data includes coordinate points and prediction time points. Specifically, the coordinate points required in each iteration can be randomly selected, and the prediction time points can also be randomly set; the amount of data of internal coordinate points and the number of boundary coordinate points are reasonably set. It can be understood that the more the amount of training data, the higher the accuracy of the thermal simulation model after training, and the training resources and time required will also increase synchronously.

[0084] Specifically, the data of multiple training samples and the internal coordinate training data and boundary coordinate training data generated in each iteration process are input into PINN, and the predicted temperature data corresponding to each data can be obtained.

[0085] Preferably, the physical neural network calculates the predicted temperature data using the following formula:

[0086]

[0087] Among them, T i represents the predicted temperature corresponding to each input data, k represents the number of hidden layers, Q represents the width of the hidden layer, σ represents the activation function, Represents neural network composite operation, NN k (z (k-1) ) indicates the k-th layer neural network operation. represents the weight matrix of the kth layer of the neural network, represents the bias parameter vector of the kth layer of the neural network, z (k-1) represents the temperature distribution vector of the k-1th layer of the neural network, Represents the temperature distribution value of the Q-th width of the k-1th layer of the neural network.

[0088] Specifically, Figure 4 As shown, each time the data input into PINN is a coordinate point and a time point, a predicted temperature is obtained through PINN.

[0089] Specifically, the PINN includes an input layer for receiving coordinate points (x, y, z) and time points t, k hidden layers in the middle and an output layer, and the output layer outputs the predicted temperature.

[0090] Specifically, the input coordinate point and time point are used as the initial z (0) , step by step through the calculation of multiple hidden layers, the final predicted temperature is obtained.

[0091] Specifically, the loss function value of this iteration is calculated based on the physical equivalent information of each layer, all heat source power data and the predicted temperature data.

[0092] Preferably, the loss function of the thermal simulation model is obtained by the following steps:

[0093] Determining the heat conduction equation and boundary condition equation of the monolithic three-dimensional integrated chip, and determining the heat conduction loss function and boundary condition loss function;

[0094] Set a first weight value, a second weight value, and a third weight value corresponding to the heat conduction loss function, the boundary condition loss function, and the training sample loss function;

[0095] The loss function of the thermal simulation model is obtained according to the heat conduction loss function, the boundary condition loss function, the training sample loss function, the first weight value, the second weight value and the third weight value.

[0096] Specifically, for thermal simulation of a monolithic three-dimensional integrated chip, the thermal simulation process needs to satisfy the physical laws of heat conduction and cannot exceed the boundary of the monolithic three-dimensional integrated chip, based on which the thermal conduction equation and boundary condition equation of the monolithic three-dimensional integrated chip are set.

[0097] Specifically, the heat conduction equation of a monolithic three-dimensional integrated chip can be:

[0098]

[0099] Among them, ρ(x, y, z, t) represents the equivalent mass density of the coordinate point (x, y, z) at time point t, C p (x, y, z, t) represents the equivalent specific heat capacity of the coordinate point (x, y, z) at time point t, T(x, y, z, t) represents the predicted temperature of the coordinate point (x, y, z) at time point t, κ(x, y, z, t) represents the equivalent thermal conductivity of the coordinate point (x, y, z) at time point t, ▽ represents the Nebra operator, M represents the number of heat sources inside the monolithic three-dimensional integrated chip, g s (x0, y0, z0, t) represents the power of the heat source S with coordinates (x0, y0, z0,) at time point t.

[0100] Specifically, the number of heat sources M is determined according to the single-chip three-dimensional integrated chip to be simulated, that is, the heat source power data measured in step S2. Figure 2 As shown, the heat source can be core particle 1 and core particle 2.

[0101] Specifically, the boundary condition equation is:

[0102]

[0103] Wherein, h represents the convection heat transfer coefficient, f(x, y, z, t) represents the ambient temperature of the boundary coordinate point (x, y, z) at time point t, and n is the normal vector of the boundary coordinate point (x, y, z).

[0104] Specifically, the loss function of the thermal simulation model is:

[0105] Loss = w1L1 + w2L2 + w3L3;

[0106]

[0107] Among them, Loss, L1, L2, L2 represent the loss function of the thermal simulation model, the heat conduction loss function, the boundary condition loss function, and the training sample loss function, respectively; w1, w2, w3 represent the first weight value, the second weight value, and the third weight value, respectively; N1, N2, N3 represent the number of internal coordinate points, the number of boundary coordinate points, and the number of training samples, respectively; Ti0 Represents the temperature of the coordinate point in the training sample, T i represents the predicted temperature of the i-th coordinate point, f i Represents the ambient temperature of the i-th boundary coordinate point at time point t.

[0108] Specifically, the heat conduction loss function and the boundary condition loss function are determined by the heat conduction equation and the boundary condition equation, and the training sample loss function is determined at the same time. The PINN provided by the embodiment of the present invention includes a heat conduction loss function, a boundary condition loss function and a training sample loss function, and N1, N2, and N3 represent the number of internal coordinate points, the number of boundary coordinate points, and the number of training samples, respectively, and can be reasonably set according to the situation.

[0109] w1, w2, and w3 represent the first weight value, the second weight value, and the third weight value, respectively. During the training process, they can be reasonably adjusted to enable rapid training to obtain a thermal simulation model.

[0110] Specifically, in the multiple iterations of training PINN, after each iteration, the loss function value of each iteration is calculated by Loss = w1L1 + w2L2 + w3L3, combined with the pre-set loss threshold. If the loss function value of this iteration meets the loss threshold, the training process ends and the trained thermal simulation model is obtained; if the loss function value of this iteration does not meet the pre-set loss threshold, the next iteration process is performed until the loss function value at the end of each iteration meets the loss threshold, and the thermal simulation model is obtained.

[0111] Specifically, at each iteration, the weight matrix and bias parameter vector of the PINN can be updated by the error back propagation algorithm, thereby optimizing the PINN. Exemplarily, the weight and bias parameters are adjusted by calling the Adam optimization algorithm.

[0112] Specifically, Figure 1 As shown, in step S4, after the thermal simulation model is trained, all coordinate points of the monolithic three-dimensional integrated chip and all predicted time points of each coordinate point are input into the thermal simulation model to obtain the predicted temperature of each coordinate point at each predicted time point.

[0113] Preferably, after obtaining the predicted temperature of each coordinate point at each time point, the thermal simulation method further comprises:

[0114] The coordinate range of the monolithic three-dimensional integrated chip is numerically discretized according to the preset coordinate optimization step and time optimization step, and the coefficient matrix of each coordinate point at each time point is determined;

[0115] Based on the least squares method, the projection matrix of each coordinate point at each time point is determined according to the coefficient matrix of each coordinate point at each time point;

[0116] The predicted temperature is updated according to the projection matrix of each coordinate point at each time point to obtain the final predicted temperature.

[0117] Specifically, through the above steps S1, S2, S3 and S4, the predicted temperature of each coordinate point in the monolithic three-dimensional integrated chip at each time point can be obtained. In order to further improve the fitting accuracy of the predicted temperature, the predicted temperature of step S4 is further processed.

[0118] Specifically, the coordinate range of the monolithic three-dimensional integrated chip set in step S1 is discretized by a preset coordinate optimization step, and the time required for thermal simulation is discretized by a preset time optimization step, so that the coordinate range is in a grid shape and the time is discrete.

[0119] Preferably, the coefficient matrix of each coordinate point at each time point is:

[0120]

[0121] Among them, A represents the coefficient matrix of each coordinate point at each time point, ρ represents the equivalent mass density at each coordinate point, and C p represents the equivalent specific heat capacity at each coordinate point, Δt represents the time optimization step, Δx, Δy, Δz represent the coordinate optimization step of the X-axis, Y-axis, and Z-axis respectively, i, j, k represent the sequence number of each coordinate point in the X-axis, Y-axis, and Z-axis after numerical discretization, and κ represents the equivalent thermal conductivity after numerical discretization.

[0122] Specifically, determine the grid corresponding to each coordinate point in the coordinate range after numerical discretization, that is, the sequence number i, j, k of each coordinate point on the X-axis, Y-axis, and Z-axis, and use i, j, k to represent each coordinate point. Determine the coefficient matrix of each coordinate point at each time point:

[0123]

[0124] Specifically, the projection matrix of each coordinate point at each time point is:

[0125] I A T A(AA T ) -1 A;

[0126] Where I represents the identity matrix.

[0127] Specifically, the predicted temperature is updated to obtain the final predicted temperature:

[0128] T i :=[IA T A(AA T ) -1 A]T i ;

[0129] Among them, T i Indicates the final predicted temperature.

[0130] It is worth noting that by performing projection correction on the predicted temperature of each coordinate point output by the thermal simulation model at each time point, the problem of large prediction errors of local coordinate points that may be caused by minimizing the average error is eliminated, thereby improving the accuracy of predicted temperature.

[0131] Compared with the prior art, an intelligent thermal simulation method for a monolithic three-dimensional integrated chip provided by an embodiment of the present invention layers the monolithic three-dimensional integrated chip, measures and obtains a small amount of training samples, trains PINN based on the training samples to obtain a thermal simulation model of the monolithic three-dimensional integrated chip, and performs thermal simulation on the monolithic three-dimensional integrated chip according to the thermal simulation model, so as to quickly obtain simulation data of the monolithic three-dimensional integrated chip; at the same time, in the process of training PINN, by using randomly generated internal coordinate training data and boundary coordinate training data, the number of coordinates required for calculating the loss function can be increased, the accuracy of the loss function value can be improved, and the thermal simulation data generated by the trained thermal simulation model can be more reliable; and the corresponding first weight value, second weight value and third weight value are set for the heat conduction loss function, the boundary condition loss function and the training sample loss function, so as to reasonably measure the loss value of the physical law and the experimental data, and improve the authenticity of the thermal simulation data generated by the thermal simulation model.

[0132] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0133] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for intelligent thermal simulation of a monolithic three-dimensional integrated chip, characterized in that: The thermal simulation method comprises: Acquire structural information of a monolithic three-dimensional integrated chip, layer the monolithic three-dimensional integrated chip, and construct a coordinate system of the monolithic three-dimensional integrated chip and physical equivalent information of each layer; Measure and obtain the temperatures of multiple coordinate points of a single three-dimensional integrated chip at different time points, and use the temperature of each coordinate point at a time point as a training sample to obtain multiple training samples; measure the coordinate points of the heat source and the power of each heat source coordinate point at different time points to form heat source power data; Based on multiple training samples, combined with the physical equivalent information of each layer and the heat source power data, the physical neural network PINN is trained to obtain a thermal simulation model of a single 3D integrated chip. All coordinate points of the monolithic three-dimensional integrated chip and all predicted time points of each coordinate point are input into the thermal simulation model to obtain the predicted temperature of each coordinate point at each predicted time point.

2. The thermal simulation method according to claim 1, characterized in that: The physical equivalent information of each layer includes one or more of the following: Equivalent thermal conductivity; Equivalent specific heat capacity; Equivalent mass density.

3. The thermal simulation method according to claim 2, characterized in that: The method trains the physical neural network PINN based on multiple training samples, combined with the physical equivalent information of each layer and all heat source power data, to obtain a thermal simulation model of a single three-dimensional integrated chip, including: Generate multiple internal coordinate training data and multiple boundary coordinate training data required for this iteration; each training data includes a coordinate point and a predicted time point; Inputting a plurality of internal coordinate training data, a plurality of boundary coordinate training data and a plurality of training samples into the thermal simulation model respectively to obtain corresponding predicted temperature data; Calculate the loss function value of this iteration based on the physical equivalent information of each layer, all heat source power data and the predicted temperature data; It is determined whether the loss function value of this iteration meets the loss threshold; if it meets the loss threshold, the training process is terminated to obtain a trained thermal simulation model; otherwise, the next iteration process is started.

4. The thermal simulation method according to claim 3, characterized in that: The physical neural network calculates the predicted temperature data using the following formula: Among them, T i represents the predicted temperature corresponding to each input data, k represents the number of hidden layers, Q represents the width of the hidden layer, σ represents the activation function, Represents neural network composite operation, NN k (z (k-1) ) indicates the k-th layer neural network operation. represents the weight matrix of the kth layer of the neural network, represents the bias parameter vector of the Kth layer of the neural network, z (k-1) represents the temperature distribution vector of the k-1th layer of the neural network, Represents the temperature distribution value of the Q-th width of the k-1th layer of the neural network.

5. The thermal simulation method according to claim 3, characterized in that: The loss function of the thermal simulation model is obtained by the following steps: Determining the heat conduction equation and boundary condition equation of the monolithic three-dimensional integrated chip, and determining the heat conduction loss function and boundary condition loss function; Set a first weight value, a second weight value, and a third weight value corresponding to the heat conduction loss function, the boundary condition loss function, and the training sample loss function; The loss function of the thermal simulation model is obtained according to the heat conduction loss function, the boundary condition loss function, the training sample loss function, the first weight value, the second weight value and the third weight value.

6. The thermal simulation method according to claim 5, characterized in that: The heat conduction equation is: Among them, ρ(x, y, z, t) represents the equivalent mass density of the coordinate point (x, y, z) at time point t, C p (x, y, z, t) represents the equivalent specific heat capacity of the coordinate point (x, y, z) at time point t, T(x, y, z, t) represents the predicted temperature of the coordinate point (x, y, z) at time point t, κ(x, y, z, t) represents the equivalent thermal conductivity of the coordinate point (x, y, z) at time point t, represents the Nebra operator, M represents the number of heat sources inside the monolithic three-dimensional integrated chip, S represents the heat source, g s (x0, y0, z0, t) represents the heat source power of the heat source S with coordinate point (x0, y0, z0) at time point t.

7. The thermal simulation method according to claim 6, characterized in that: The boundary condition equation is: Wherein, h represents the convection heat transfer coefficient, f(x, y, z, t) represents the ambient temperature of the boundary coordinate point (x, y, z) at time point t, and n is the normal vector of the boundary coordinate point (x, y, z).

8. The thermal simulation method according to claim 7, characterized in that: The loss function of the thermal simulation model is: Loss = w1L1 + w2L2 + w3L3; Among them, Loss, L1, L2, L2 represent the loss function of the thermal simulation model, the heat conduction loss function, the boundary condition loss function, and the training sample loss function, respectively; w1, w2, w3 represent the first weight value, the second weight value, and the third weight value, respectively; N1, N2, N3 represent the number of internal coordinate points, the number of boundary coordinate points, and the number of training samples, respectively; T i0 Represents the temperature of the coordinate point in the training sample, T i represents the predicted temperature, f i Represents the ambient temperature of the i-th boundary coordinate point at time t.

9. The thermal simulation method according to claim 1, characterized in that: After obtaining the predicted temperature of each coordinate point at each time point, the thermal simulation method further includes: The coordinate range of the monolithic three-dimensional integrated chip is numerically discretized according to the preset coordinate optimization step and time optimization step, and the coefficient matrix of each coordinate point at each time point is determined; Based on the least squares method, the projection matrix of each coordinate point at each time point is determined according to the coefficient matrix of each coordinate point at each time point; The predicted temperature is updated according to the projection matrix of each coordinate point at each time point to obtain the final predicted temperature.

10. The thermal simulation method according to claim 9, characterized in that: The coefficient matrix of each coordinate point at each time point is: Among them, A represents the coefficient matrix of each coordinate point at each time point, ρ represents the equivalent mass density at each coordinate point, and C p represents the equivalent specific heat capacity at each coordinate point, Δt represents the time optimization step, Δx, Δy, Δz represent the coordinate optimization step of the X-axis, Y-axis, and Z-axis respectively, i, j, k represent the sequence number of each coordinate point in the X-axis, Y-axis, and Z-axis after numerical discretization, and κ represents the equivalent thermal conductivity after numerical discretization.

Images