Diamond substrate heat sink sheet packaging quality intelligent evaluation method and system

Abstract

The application discloses a diamond substrate heat sink piece packaging quality intelligent evaluation method and system, belongs to the technical field of thermal analysis, in order to solve the problem of low evaluation accuracy and not deep fusion of acoustic-thermal signals in the evaluation of diamond substrate heat sink piece packaging quality, the diamond substrate heat sink piece packaging quality intelligent evaluation method includes obtaining the transient thermal reflection attenuation sequence and the ultrasonic echo sequence of the diamond heat sink piece packaging structure; a discrete time exponential decay model corresponding to the transient thermal reflection attenuation sequence is constructed; a thermal time constant vector is generated; a spatial reflectivity difference sequence is generated; an acoustic-thermal cross attention weight matrix is constructed through matrix multiplication; an interface thermal resistance distribution matrix of the diamond heat sink piece packaging structure is generated; the matrix elements in the interface thermal resistance distribution matrix are compared with a preset failure thermal resistance threshold value; a structure failure early warning instruction containing all the defect node position information is output; the application improves the accuracy of diamond substrate heat sink piece packaging quality evaluation.

Description

Technical Field

[0001] This invention relates to the field of thermal analysis technology, specifically to an intelligent evaluation method and system for the packaging quality of diamond substrate heat sink sheets. Background Technology

[0002] With the rapid development of semiconductor devices towards high power and high integration, diamond substrates, with their ultra-high thermal conductivity, excellent mechanical properties and good electrical compatibility, have become the core heat sink material in the packaging of high-end power devices and optoelectronic devices. The packaging quality directly determines the heat dissipation efficiency, operational stability and service life of the device. The core evaluation indicators for the packaging quality of diamond substrate heat sinks are the distribution of interfacial thermal resistance and the density of interfacial bonding. If defects such as poor soldering, micro-gaps, and residual stress concentration occur during the packaging process, the interfacial thermal resistance will increase sharply, the heat of the device cannot be dissipated in time, and thus the device will overheat and fail, causing serious economic losses.

[0003] Currently, the quality assessment methods for diamond substrate heat sink packaging are mainly divided into two categories: thermal testing and acoustic testing. Thermal testing can detect thermal resistance-related parameters, but it struggles to accurately characterize interfacial bonding density and is easily affected by ambient temperature and the dispersion of testing instruments, leading to significant assessment errors. Acoustic testing can reflect the interfacial bonding state, but it cannot correlate with thermal conductivity characteristics, making it difficult to accurately quantify interfacial thermal resistance. Existing technologies have attempted to combine thermal and acoustic signals, but these methods often employ simple feature superposition, failing to establish the intrinsic correlation between the two signals and fully exploiting their complementary information, resulting in insufficient assessment accuracy and difficulty in identifying minute defects. Furthermore, existing assessment methods do not consider the impact of diamond substrate surface roughness, residual stress, and residual stress in the packaging structure on the assessment results, and lack precise correction mechanisms for the assessment results, further reducing the reliability and versatility of the assessment.

[0004] To address the above issues, a smart evaluation method and system for the packaging quality of diamond substrate heat sink sheets is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent evaluation method and system for the packaging quality of diamond substrate heat sink sheets. By using this device, the problem of insufficient fusion of acoustic and thermal signals and low evaluation accuracy in the packaging quality evaluation of diamond substrate heat sink sheets mentioned above is solved.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent evaluation of the packaging quality of diamond substrate heat sink sheets, comprising: Obtain the transient thermal reflection attenuation sequence and ultrasonic echo sequence of the diamond heat sink packaging structure; Construct a discrete-time exponential decay model corresponding to the transient thermal reflection decay sequence; based on the discrete-time exponential decay model, obtain multiple sets of fitting residual values; when the fitting residual value is lower than a preset residual tolerance threshold, extract the corresponding decay coefficient value and generate a thermal time constant vector. Extract the amplitude envelope curve of the ultrasonic echo sequence; perform differential calculation on the amplitude envelope curve according to a preset spatial step size to generate a spatial reflectivity difference sequence; generate an acoustic impedance gradient vector based on the spatial reflectivity difference sequence and a preset local impedance constant. Using the acoustic impedance gradient vector as the query condition and the thermal time constant vector as the key sequence and value sequence, an acoustic-thermal cross-attention weight matrix is ​​constructed through matrix multiplication. The interface thermal resistance distribution matrix of the diamond heat sink packaging structure is generated by performing a Hadamard product operation based on the acoustic-thermal cross-attention weight matrix and the preset Kapitza thermal resistance reference matrix. The matrix elements in the interface thermal resistance distribution matrix are compared with a preset failure thermal resistance threshold; when a matrix element is greater than the failure thermal resistance threshold, the corresponding grid coordinates are marked as defect nodes; and a structural failure early warning command containing the location information of all the defect nodes is output.

[0007] Furthermore, using the acoustic impedance gradient vector as a query condition and the thermal time constant vector as a key sequence and a value sequence, a cross-attention weight matrix for acoustic and thermal processes is constructed through matrix multiplication, including: Multiply the acoustic impedance gradient vector by the first weight mapping matrix to generate the first query vector; The thermal time constant vector is multiplied by the second weight mapping matrix and the third weight mapping matrix respectively to generate the first key value vector and the first value vector respectively; Calculate the dot product of the first query vector and the first key value vector, and divide the dot product result by the scaling factor; Apply the normalized exponential function to the result after dividing by the scaling factor to generate an intermediate probability distribution matrix; The product of the intermediate probability distribution matrix and the first value vector is calculated to obtain the acoustic-thermal cross-attention weight matrix.

[0008] Furthermore, the Kapitza thermal resistance reference matrix is ​​obtained in advance in the following manner: Obtain the phonon transport parameters of the diamond layer and the Debye temperature constant of the solder layer in the diamond heat sink packaging structure; Based on the phonon mismatch model, the theoretical thermal interface conductivity limit between the diamond layer and the solder layer is calculated. The theoretical thermal interface conductivity limit is calculated by reciprocal and then filled with values ​​according to a preset evaluation pixel grid to generate the Kapiza thermal resistance reference matrix.

[0009] Furthermore, before obtaining multiple sets of fitted residual values ​​based on the discrete-time exponential decay model, the method further includes: Perform an inverse Laplace transform on the transient thermal reflection attenuation sequence to generate a continuous time constant spectrum; Calculate the peak half-width parameter in the continuous time constant spectrum; When the peak half-width at half-maximum parameter is less than the preset instrument dispersion threshold, the corresponding spectral range is retained for constructing the discrete-time exponential decay model.

[0010] Further, the amplitude envelope curve of the ultrasound echo sequence is extracted, including: Perform a Hilbert transform on the ultrasonic echo sequence to generate orthogonal analytic signals; Calculate the magnitude sequence of the orthogonal analytic signal; A sequence of modulus values ​​corresponding to the acoustic wave transit time window between the diamond lattice and the underlying solder is extracted and used as the amplitude envelope curve.

[0011] Furthermore, based on the phonon mismatch model, the theoretical thermal interface conductivity limit between the diamond layer and the solder layer is calculated, including: Obtain the surface roughness parameters of the diamond; Calculate the microscopic scattering attenuation coefficient of the phonon transport parameter under the diamond surface roughness parameter; The microscopic scattering attenuation coefficient and the Debye temperature constant are input into the phonon mismatch model and multiplied algebraically to obtain the theoretical thermal interface conductivity limit.

[0012] Furthermore, a structural failure early warning command containing the location information of all the aforementioned defect nodes is output, including: An eight-neighbor connected component search algorithm is applied to all the defect nodes to generate discrete defect patch regions. Calculate the sum of pixel areas and geometric eccentricity for each of the defective patch regions; Based on the sum of the pixel areas and the value of the geometric eccentricity, a rating tag is extracted from a preset rating lookup table, and a structural failure warning instruction containing the rating tag is generated and output.

[0013] Furthermore, before calculating the product of the intermediate probability distribution matrix and the first value vector, the method further includes: Based on the acoustic impedance gradient vector, the residual stress field distribution inside the diamond heat sink packaging structure is obtained by inversion using the acoustic elasticity theory equation. The residual stress field distribution is input as a constraint condition into the first key value vector, and stress-weighted correction is performed on the first key value vector.

[0014] Furthermore, after obtaining the transient thermal reflection attenuation sequence and ultrasonic echo sequence of the diamond heat sink packaging structure, the process also includes: Obtain the X-ray diffraction rocking curve data of the diamond heat sink packaging structure and extract the corresponding full width at half maximum (FWHM) broadening value; The half-width expansion value is concatenated with the thermal time constant vector to generate a fused feature vector. The fused feature vector is input into the pre-trained interface thermal resistance prediction model to correct the initial value of the acoustic-thermal cross-attention weight matrix.

[0015] A smart evaluation system for the packaging quality of diamond substrate heat sink wafers includes: The signal acquisition module acquires the transient thermal reflection attenuation sequence and ultrasonic echo sequence of the diamond heat sink packaging structure; The time constant vector generation module constructs a discrete-time exponential decay model corresponding to the transient thermal reflection decay sequence; based on the discrete-time exponential decay model, it obtains multiple sets of fitting residual values; when the fitting residual value is lower than a preset residual tolerance threshold, it extracts the corresponding decay coefficient value and generates a thermal time constant vector. The acoustic impedance gradient vector generation module extracts the amplitude envelope curve of the ultrasonic echo sequence; performs differential calculation on the amplitude envelope curve according to a preset spatial step size to generate a spatial reflectivity difference sequence; and generates an acoustic impedance gradient vector based on the spatial reflectivity difference sequence and a preset local impedance constant. The acoustic-thermal cross-attention weight matrix construction module uses the acoustic impedance gradient vector as the query condition and the thermal time constant vector as the key sequence and value sequence, and constructs the acoustic-thermal cross-attention weight matrix through matrix multiplication. The interface thermal resistance distribution matrix generation module performs a Hadamard product operation based on the acoustic-thermal cross-attention weight matrix and the preset Kapitza thermal resistance reference matrix to generate the interface thermal resistance distribution matrix of the diamond heat sink packaging structure. The failure early warning module compares the matrix elements in the interface thermal resistance distribution matrix with a preset failure thermal resistance threshold; when a matrix element is greater than the failure thermal resistance threshold, the corresponding grid coordinates are marked as defect nodes; and outputs a structural failure early warning command containing the location information of all the defect nodes.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention obtains transient thermal reflection attenuation sequences and ultrasonic echo sequences, extracts thermal time constant vectors and acoustic impedance gradient vectors respectively, and constructs an acoustic-thermal cross-attention weight matrix through matrix multiplication to establish the intrinsic correlation between acoustic and thermal signals. It fully explores the complementary information of the two signals, solves the problem of insufficient evaluation accuracy caused by the simple superposition of acoustic and thermal signals, and can accurately identify minute defects, such as tiny gaps and local cold welds, thus improving the calculation accuracy of interface thermal resistance distribution. 2. By preprocessing the transient thermal reflection attenuation sequence with inverse Laplace transform, an effective spectral range is selected for constructing a discrete-time exponential decay model, reducing the impact of instrument discreteness on the extraction of the thermal time constant vector. The amplitude envelope curve of the ultrasonic echo sequence is extracted using Hilbert transform, improving the accuracy of acoustic impedance gradient vector extraction. Simultaneously, the key vector is corrected using the residual stress field distribution, and the initial value of the acoustic-thermal cross-attention weight matrix is ​​corrected by combining X-ray diffraction rocking curve data, further improving the reliability and anti-interference capability of the evaluation results. 3. Based on the phonon mismatch model, the microscopic scattering attenuation coefficient is calculated by combining the diamond surface roughness parameters. The theoretical thermal interface conductivity limit is corrected so that the Kapitza thermal resistance reference matrix is ​​more in line with the actual contact state between the diamond substrate and the solder layer. This solves the evaluation deviation problem caused by the current reference matrix not considering the influence of surface roughness, and improves the pertinence and accuracy of the interface thermal resistance distribution matrix calculation. 4. By dynamically combining parameters such as diamond substrate thickness and solder type, the construction process of the acoustic-thermal cross-attention weight matrix is ​​optimized. At the same time, the initial value of the calculation is corrected by introducing fused feature vectors, so that the evaluation method can be adapted to the quality evaluation of diamond substrate heat sinks with different specifications and structures, thus improving its versatility. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the present invention. Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Please refer to Figure 1 An intelligent evaluation method for the packaging quality of a diamond substrate heat sink, as described in this embodiment of the invention, includes the following steps: Step 101: Obtain the transient thermal reflection attenuation sequence and ultrasonic echo sequence of the diamond heat sink package structure. Specifically, the transient thermal reflection attenuation sequence of the diamond heat sink package structure is acquired through a transient thermal reflection test system. This system uses a femtosecond pulsed laser as the excitation source, with a wavelength set to 800nm ​​and a pulse width not exceeding 100fs. A continuous semiconductor laser is used as the probe light. The reflected light signal is received by a photodetector and converted into an electrical signal. After sampling by a data acquisition card, the transient thermal reflection attenuation sequence is generated. The ultrasonic echo sequence is acquired through an ultrasonic detection system. A high-frequency line-focusing probe is used as the ultrasonic probe, with a center frequency not lower than 50MHz. High thermal conductivity silicone oil is used as the coupling agent to ensure effective sound energy transmission. The ultrasonic echo signal is amplified and filtered by a preamplifier and then sampled by the data acquisition system to generate the ultrasonic echo sequence. The transient thermal reflection attenuation sequence refers to the continuous numerical sequence of the intensity of reflected light decaying with time within a preset time window. The ultrasonic echo sequence refers to the continuous acoustic signal sequence formed by the reflection of ultrasonic signals between the diamond substrate and the solder layer, and between the solder layer and the substrate. Step 102: Construct a discrete-time exponential decay model corresponding to the transient thermal reflection decay sequence; based on the discrete-time exponential decay model, obtain multiple sets of fitting residual values; when the fitting residual value is lower than the preset residual tolerance threshold, extract the corresponding decay coefficient value and generate a thermal time constant vector; specifically, the discrete-time exponential decay model adopts a multi-exponential superposition form, and the expression is: y(t)=ΣA n e^(-t / τ n ), where y(t) is the amplitude of the transient thermal reflection decay sequence at time t, A n Let τ be the attenuation amplitude of the nth heat transfer path. n Let be the thermal time constant of the nth heat transfer path, where n is the number of heat transfer paths, preset to 3-5 based on the number of encapsulation layers. The transient thermal reflection attenuation sequence is fitted to this model using the least squares method. The residual value after each fitting is calculated, and the root mean square error (RMSE) is used to calculate the residual value. The specific calculation process is as follows: first, calculate the difference between the actual value and the model-fitted value at each sampling point; then, square each difference; calculate the arithmetic mean of all squared differences; finally, take the square root of this arithmetic mean. The preset residual tolerance threshold is set according to the accuracy of the detection instrument, typically between 0.05 and 0.1. When the fitted residual value is lower than this threshold, the fitting is considered valid, and the corresponding attenuation coefficient 1 / τ is extracted. n Arranged in order of heat transfer path to form a thermal time constant vector; the thermal time constant vector is a one-dimensional numerical vector that characterizes the thermal response speed of different heat transfer paths such as diamond substrate, solder layer, and substrate, and the vector dimension is consistent with the number of heat transfer paths. Step 103: Extract the amplitude envelope curve of the ultrasonic echo sequence; perform differential calculation on the amplitude envelope curve according to the preset spatial step size to generate a spatial reflectivity difference sequence; generate an acoustic impedance gradient vector based on the spatial reflectivity difference sequence and the preset local impedance constant; specifically, the amplitude envelope curve of the ultrasonic echo sequence is extracted using an envelope detection algorithm, the carrier component of the ultrasonic signal is removed, and the amplitude change reflecting the interface reflection characteristics is retained; the preset spatial step size is set according to the scanning accuracy of the ultrasonic probe, usually 0.1-0.5mm, and the first-order differential calculation is performed on the amplitude envelope curve according to this step size. The specific calculation process is as follows: for each sampling point on the amplitude envelope curve, the amplitude of the sampling point is subtracted from the amplitude of the previous sampling point, and the difference obtained is the adjacent scan value. The amplitude difference between the plotted points is arranged sequentially to form a spatial reflectivity difference sequence. The preset local impedance constant is based on the inherent impedance parameters of the diamond substrate and the solder layer. The specific calculation process is as follows: first, obtain the acoustic impedance values ​​of the diamond substrate and the solder layer, add the two acoustic impedance values, and then divide by 2. The result is the local impedance constant. Divide each element in the spatial reflectivity difference sequence by the local impedance constant. Arrange the results sequentially to form an acoustic impedance gradient vector. The acoustic impedance gradient vector is a one-dimensional numerical vector that characterizes the change in bonding density at different locations of the packaging interface. The larger the value of the vector element, the more drastic the change in bonding density at the corresponding location, which may indicate a defect. Step 104: Using the acoustic impedance gradient vector as the query condition and the thermal time constant vector as the key-value sequence and value sequence, construct the acoustic-thermal cross-attention weight matrix through matrix multiplication. Specifically, the acoustic impedance gradient vector is used as the query vector in the attention mechanism, and the thermal time constant vector is used as both the key vector and value vector. First, perform the first matrix multiplication operation to calculate the product of the query vector and the transpose of the key vector. The specific calculation process is as follows: the query vector is a matrix with 1 row and multiple columns, and the transpose of the key vector is a matrix with multiple rows and 1 column. Multiply each element of the query vector by each corresponding element of the transpose of the key vector, and then add all the multiplication results to obtain a scalar result. Then... The scalar result is normalized. The specific calculation process is as follows: First, calculate the exponent values ​​of all scalar results, and divide each exponent value by the sum of all exponent values ​​to obtain the normalized result. Finally, perform a second matrix multiplication operation, multiplying the normalized result by the value vector. The specific calculation process is as follows: The normalized result is a scalar, and the scalar is multiplied by each element of the value vector. The results are arranged in order to form the acoustic-thermal cross-attention weight matrix. The acoustic-thermal cross-attention weight matrix is ​​a two-dimensional matrix that characterizes the correlation strength between the acoustic impedance gradient feature and the thermal time constant feature. The larger the matrix element value, the stronger the correlation between the acoustic and thermal features at the corresponding position, and the greater the contribution to the calculation of interface thermal resistance. Step 105: Perform a Hadamard product operation based on the acoustic-thermal cross-attention weight matrix and the preset Kapitza thermal resistance reference matrix to generate the interface thermal resistance distribution matrix of the diamond heat sink packaging structure. Specifically, the preset Kapitza thermal resistance reference matrix is ​​a reference thermal resistance matrix calculated in advance based on the material properties of the diamond substrate and the solder layer. Its spatial resolution is consistent with the ultrasonic scanning resolution and transient thermal reflection test resolution. The Hadamard product operation is the pairwise multiplication of corresponding elements of two matrices of the same dimension. The specific calculation process is as follows: find the elements with the same position in the acoustic-thermal cross-attention weight matrix and the Kapitza thermal resistance reference matrix, multiply the two elements, and arrange the results of multiplying all corresponding elements according to the original matrix position to obtain the interface thermal resistance distribution matrix. The interface thermal resistance distribution matrix is ​​a two-dimensional matrix that characterizes the interface thermal resistance at different positions of the diamond substrate heat sink packaging structure. The matrix element value is the interface thermal resistance value at the corresponding position, with the unit being K / W. Step 106: Compare the matrix elements in the interface thermal resistance distribution matrix with the preset failure thermal resistance threshold; when a matrix element is greater than the failure thermal resistance threshold, mark the corresponding grid coordinates as a defect node; output a structural failure early warning command containing the location information of all defect nodes; specifically, the preset failure thermal resistance threshold is set according to the application scenario of the diamond substrate heat sink, and for high-power device packaging, it is usually 50-100K / W; compare each element in the interface thermal resistance distribution matrix with the threshold one by one, and when the element value is greater than the threshold, it is determined that there is a packaging defect at the corresponding position, and the grid coordinates of the position are recorded, which are the defect nodes; organize the grid coordinates of all defect nodes into a location information list, combine it with the preset early warning level, generate a structural failure early warning command and output it; the structural failure early warning command refers to the command information containing the defect location, defect quantity and early warning level, which can be sent to the monitoring terminal or control system to guide subsequent packaging quality rectification.

[0020] In one implementation, the acquisition of the transient thermal reflection attenuation sequence can adopt a synchronous triggering mode, so that the transient thermal reflection test and ultrasonic detection are performed simultaneously, ensuring that the spatial positions of the two sequences correspond one-to-one and improving the accuracy of subsequent feature fusion.

[0021] In another implementation, the preset residual tolerance threshold can be dynamically adjusted according to the complexity of the packaging structure. For multi-layer packaging structures, the threshold can be appropriately increased to a value of 0.08-0.1 to reduce the fitting difficulty; for simple packaging structures, the threshold can be decreased to a value of 0.05-0.07 to improve the extraction accuracy of the thermal time constant vector.

[0022] By combining transient thermal reflection and ultrasonic testing, thermal and acoustic features are extracted respectively. The intrinsic relationship between the two features is established through the acoustic-thermal cross-attention mechanism, fully exploring complementary information and solving the problem of insufficient evaluation accuracy of single detection methods. The interface thermal resistance distribution matrix enables precise defect location, and the failure warning command provides a clear basis for packaging quality control, thereby improving the evaluation accuracy of diamond substrate heat sink packaging quality.

[0023] In some embodiments, the acoustic impedance gradient vector is used as the query condition, and the thermal time constant vector is used as the key sequence and value sequence. The acoustic-thermal cross-attention weight matrix is ​​constructed by matrix multiplication, including the following steps: Step 201: Multiply the acoustic impedance gradient vector by the first weight mapping matrix to generate the first query vector; specifically, the first weight mapping matrix is ​​a weight matrix pre-trained using multiple sets of encapsulated sample data, and its dimension is (d _q d _in ), where d _in Let d be the dimension of the acoustic impedance gradient vector. _q The preset dimension of the first query vector is typically 64-256; the specific calculation process of matrix multiplication is as follows: the acoustic impedance gradient vector is a 1-row vector d _in The column matrix, the first weight mapping matrix is ​​d _in line d _q The matrix of columns is obtained by multiplying each element of the acoustic impedance gradient vector by all elements of the corresponding row of the first weight mapping matrix, and then summing the results of the multiplications in each column to obtain a 1-row d matrix. _q The matrix of columns is the first query vector; the first query vector is the acoustic impedance gradient feature vector after weight mapping, which is used to query the correlation with the thermal time constant feature in the attention mechanism.

[0024] Step 202: Multiply the thermal time constant vector by the second weight mapping matrix and the third weight mapping matrix respectively to generate the first key-value vector and the first value vector; specifically, the second weight mapping matrix and the third weight mapping matrix are both pre-trained weight matrices, both with dimension (d). _k d _in' ), where d _in' Let d be the dimension of the thermal time constant vector. _k Let be the preset dimensions of the first key-value vector and the first value vector, and be the same as the dimension d of the first query vector. _q To maintain consistency; the specific calculation process of matrix multiplication is as follows: the thermal time constant vector is a 1-row d _in' The column matrix, the second weight mapping matrix, and the third weight mapping matrix are all d. _in' line d _kThe matrix of columns is obtained by multiplying each element of the thermal time constant vector by all elements of the corresponding row of the second weight mapping matrix, and then summing the results of the multiplications of each column to obtain a 1-row d matrix. _k The first key value vector of the column; using the same calculation method, multiply the thermal time constant vector by the third weight mapping matrix to obtain row d. _k The first value vector of the column; the first key value vector is used to calculate the association strength with the first query vector, and the first value vector is used to generate the final attention weight matrix.

[0025] Step 203: Calculate the dot product of the first query vector and the first key-value vector, and divide the dot product result by the scaling factor; specifically, the first query vector is a 1-row vector d _q A matrix of columns, with the first key-value vector being row 1 (d). _q The specific calculation process for the dot product of the column matrix is ​​as follows: multiply the first element of the first query vector by the first element of the first key value vector, multiply the second element by the second element, and so on, summing all the multiplication results to obtain a scalar value; the scaling factor is d. _q The square root of d is calculated as follows: _q The result of the square root operation is the scaling factor. Dividing the scalar value of the dot product by the scaling factor yields the scaled result. This operation is used to alleviate the normalization distortion problem caused by the excessively large dot product result. The scaling factor is used to perform normalization preprocessing on the correlation strength to ensure that the output of the subsequent normalization exponential function is within a reasonable range. Step 204: Apply the normalized exponential function to the result after dividing by the scaling factor to generate an intermediate probability distribution matrix. Specifically, the normalized exponential function is used to convert the scaled dot product result into a probability value, making the result between 0 and 1. This probability value represents the association probability between the first query vector and the first key value vector. The specific calculation process of the normalized exponential function is as follows: first calculate the exponential value of the scaled dot product result, and then divide the exponential value by the sum of the exponential values ​​themselves. The result is the probability value. Since both the first query vector and the first key value vector are one-dimensional vectors, the generated intermediate probability distribution matrix is ​​a 1×1 matrix, and the matrix elements are the association probability values. The intermediate probability distribution matrix is ​​a probability matrix that represents the association strength of acoustic and thermal features. The larger the element value, the stronger the association between the two features. Step 205: Calculate the product of the intermediate probability distribution matrix and the first value vector to obtain the acoustic-thermal cross-attention weight matrix; specifically, the intermediate probability distribution matrix is ​​a 1×1 matrix, and the first value vector is a 1×d matrix. _k The specific calculation process of matrix multiplication for a column matrix is ​​as follows: multiply the unique element of the intermediate probability distribution matrix by each element of the first value vector, and arrange the results in order to form a 1×d matrix. _kThe acoustic-thermal cross-attention weight matrix of the column; if the thermal time constant vector corresponds to multiple heat transfer paths, it can be expanded into a multi-dimensional weight matrix through batch operation to ensure that the spatial resolution of the weight matrix is ​​consistent with the interface thermal resistance distribution matrix.

[0026] In one implementation, the first, second, and third weight mapping matrices can be obtained by training a deep learning model. The training samples are acoustic and thermal signal data of diamond substrate heat sinks with different packaging quality. The parameters are optimized with the goal of minimizing the error in the interface thermal resistance calculation.

[0027] The above technical solution unifies the dimensions and enhances the features of acoustic and thermal features through a weight mapping matrix. By combining scaling factors and normalized exponential functions to optimize the correlation strength calculation, the constructed acoustic and thermal cross-attention weight matrix can accurately represent the intrinsic correlation of acoustic and thermal features. This solves the problem of poor fusion effect caused by the simple superposition of existing acoustic and thermal features, and provides reliable support for the accurate calculation of the interface thermal resistance distribution matrix.

[0028] In some embodiments, the Kapitza thermal resistance reference matrix is ​​obtained in advance by means of the following steps: Step 301: Obtain the phonon transport parameters of the diamond layer and the Debye temperature constant of the solder layer in the diamond heat sink packaging structure. Specifically, the phonon transport parameters of the diamond layer include the phonon mean free path and phonon velocity, which are measured by a low-temperature thermal conductivity testing system. The test temperature is set to 77K (liquid nitrogen temperature) to eliminate the influence of thermal excitation on phonon transport. The Debye temperature constant of the solder layer is obtained by querying a preset database based on the solder material composition. For example, the Debye temperature constant of Sn-Ag-Cu solder is usually 200-250K. The phonon transport parameters are the core parameters characterizing the thermal conductivity of the diamond layer, and the Debye temperature constant is the key parameter characterizing the phonon vibration characteristics of the solder layer.

[0029] Step 302: Based on the phonon mismatch model, calculate the theoretical thermal interface conductivity limit between the diamond layer and the solder layer. Specifically, the classical acoustic mismatch model is selected, and the phonon transmission parameters of the diamond layer and the Debye temperature constant of the solder layer are input into the model. The specific calculation process is as follows: First, based on the phonon velocities of the diamond layer and the solder layer, calculate the phonon transmission coefficient and reflection coefficient at the diamond-solder interface; then, combine the phonon mean free path of the diamond layer and the Debye temperature constant of the solder layer to calculate the phonon scattering probability at the interface; finally, based on the transmission coefficient, reflection coefficient, and scattering probability, calculate the theoretical thermal interface conductivity limit. The theoretical thermal interface conductivity limit refers to the maximum thermal conductivity between the diamond layer and the solder layer under ideal conditions without defects or impurities, and the unit is W / (m·K).

[0030] Step 303: Perform a reciprocal operation on the theoretical thermal interface conductivity limit and fill the values ​​according to the preset evaluation pixel grid to generate a Kapitza thermal resistance reference matrix. Specifically, perform a reciprocal operation on the theoretical thermal interface conductivity limit. The specific calculation process is as follows: divide 1 by the theoretical thermal interface conductivity limit, and the result is the theoretical Kapitza thermal resistance value, because thermal resistance is equal to the reciprocal of conductivity. The preset evaluation pixel grid is consistent with the spatial resolution of ultrasonic scanning and transient thermal reflection testing, usually 0.1mm × 0.1mm. Fill each pixel of the grid with the theoretical Kapitza thermal resistance value to generate a Kapitza thermal resistance reference matrix with the same size as the evaluation area. The Kapitza thermal resistance reference matrix is ​​a reference matrix that characterizes the interface thermal resistance distribution under ideal packaging conditions. It is used to fuse with the acoustic-thermal cross-attention weight matrix to calculate the actual interface thermal resistance distribution.

[0031] In one implementation, if the diamond substrate has a polycrystalline structure, the phonon transport parameters of different grain regions can be measured separately, the theoretical thermal interface conductivity limit of the corresponding region can be calculated, and a non-uniform Kapitza thermal resistance reference matrix can be generated to improve the specificity of the reference matrix.

[0032] The above technical solution calculates the Kapitza thermal resistance reference matrix based on the phonon mismatch model and the inherent parameters of the material, ensuring the scientific nature and accuracy of the reference matrix. This solves the problem that the existing reference matrix is ​​too simple and deviates greatly from the actual packaging state, providing a reliable reference for the accurate calculation of interface thermal resistance distribution.

[0033] In some embodiments, a method for intelligently evaluating the packaging quality of a diamond substrate heat sink includes the following steps before obtaining multiple sets of fitted residual values ​​based on a discrete-time exponential decay model: Step 401: Perform an inverse Laplace transform on the transient thermal reflection attenuation sequence to generate a continuous time constant spectrum. Specifically, the transient thermal reflection attenuation sequence is processed using a numerical inverse Laplace transform method, such as the Stehfest algorithm. The specific calculation process is as follows: First, the time domain signal of the transient thermal reflection attenuation sequence is discretized to obtain discrete time points and corresponding signal amplitudes. Then, according to the formula of the Stehfest algorithm, the weight coefficient corresponding to each time constant is calculated. Each weight coefficient is multiplied by the signal amplitude at the corresponding time point, and then all the multiplication results are added together to obtain the continuous spectrum in the time constant domain, i.e., the continuous time constant spectrum. The horizontal axis of this spectrum is the thermal time constant, and the vertical axis is the amplitude of the corresponding time constant, representing the contribution of the heat transfer path corresponding to different thermal time constants. The continuous time constant spectrum refers to the continuous distribution curve of the thermal time constant, which can clearly present the thermal response characteristics of different heat transfer paths.

[0034] Step 402: Calculate the peak half-width at half-maximum (HWHM) parameter in the continuous time constant spectrum. Specifically, first, identify all peaks in the continuous time constant spectrum, with each peak corresponding to a heat transfer path. For each peak, first find its maximum amplitude and calculate half of that maximum amplitude to obtain the half-peak value. Then, find the two time constant coordinates corresponding to the amplitude equal to the half-peak value on the spectrum, one to the left of the peak and the other to the right. Calculate the difference between these two coordinates. Specifically, subtract the value of the left coordinate from the value of the right coordinate. The result is the HWHM parameter of the peak. The HWHM parameter is an indicator of the sharpness of the peak. The smaller the HWHM, the more concentrated the thermal time constant of the corresponding heat transfer path, and the higher the signal purity.

[0035] Step 403: When the peak half-width at half maximum (HWHM) parameter is less than the preset instrument dispersion threshold, the corresponding spectral range is retained for constructing the discrete-time exponential decay model. Specifically, the preset instrument dispersion threshold is set according to the accuracy of the transient thermal reflection test system, and is usually set to 0.1–0.2 ns. The HWHM parameter of each peak is compared with the instrument dispersion threshold. When the peak HWHM parameter is less than the threshold, it is determined that the signal in the corresponding spectral range is less affected by the instrument dispersion and has higher signal purity. This range is retained for constructing the discrete-time exponential decay model. For spectral ranges with HWHM parameters greater than the threshold, it is determined that the signal interference is greater and they are discarded.

[0036] In one implementation, the inverse Laplace transform can be accelerated by GPU to improve processing efficiency, making it suitable for large-scale online detection scenarios; the instrument dispersion threshold can be dynamically adjusted according to the calibration results of the testing instrument to ensure the accuracy of spectrum range screening.

[0037] By employing the above technical solution, the transient thermal reflection decay sequence is preprocessed using inverse Laplace transform, and a high-purity spectral range is selected for model construction. This reduces the impact of instrument discreteness on the extraction of the thermal time constant vector, solves the problems of insufficient fitting accuracy and severe interference in existing models, and improves the accuracy of thermal time constant vector extraction.

[0038] In some embodiments, a smart evaluation method for the packaging quality of a diamond substrate heat sink includes the following steps: extracting the amplitude envelope curve of an ultrasonic echo sequence. Step 501: Perform a Hilbert transform on the ultrasonic echo sequence to generate an orthogonal analytic signal. Specifically, a discrete Hilbert transform is used to process the ultrasonic echo sequence, which is a real signal. The specific calculation process is as follows: Perform a Fourier transform on the discrete signal of the ultrasonic echo sequence to obtain a frequency domain signal; perform a sign function weighting on the frequency domain signal, that is, multiply the positive frequency part by j, multiply the negative frequency part by -j, and keep the zero frequency part unchanged; perform an inverse Fourier transform on the weighted frequency domain signal to obtain the corresponding orthogonal analytic signal, which is a complex signal; the real part of the analytic signal is the original ultrasonic echo sequence, and the imaginary part is the Hilbert transform result of the original signal; the core function of the Hilbert transform is to provide a 90° phase shift for the original signal, which facilitates the subsequent extraction of the amplitude envelope; the orthogonal analytic signal is a complex signal that contains the amplitude and phase information of the original signal, and can more comprehensively reflect the characteristics of the ultrasonic echo.

[0039] Step 502: Calculate the magnitude sequence of the orthogonal analytic signal. Specifically, for each sampling point of the orthogonal analytic signal, the sampling point is a complex signal containing a real part and an imaginary part. The specific calculation process of the magnitude is as follows: squaring the real part of the sampling point and squaring the imaginary part, adding the two squaring results, and then taking the square root of the sum to obtain the magnitude of the sampling point. Arranging the magnitudes of all sampling points in chronological order generates the magnitude sequence. The magnitude sequence is the amplitude change sequence of the ultrasonic echo sequence, eliminating the interference of phase information and retaining only the amplitude characteristics.

[0040] Step 503: Extract the modulus sequence segment corresponding to the acoustic wave transit time window between the diamond lattice and the underlying solder as the amplitude envelope curve; specifically, first calculate the transit time of the acoustic wave between the diamond layer and the solder layer. The specific calculation process is as follows: obtain the thickness of the diamond layer, the thickness of the solder layer, and the propagation speed of the acoustic wave in the diamond layer and the propagation speed of the acoustic wave in the solder layer respectively; divide the thickness of the diamond layer by the propagation speed of the acoustic wave in the diamond layer to obtain the propagation time of the acoustic wave in the diamond layer; divide the thickness of the solder layer by the propagation speed of the acoustic wave in the solder layer to obtain the propagation time of the acoustic wave in the solder layer; compare the two propagation times. Add the transit time of the sound wave between the diamond layer and the solder layer. Using this transit time as the center, set a time window. The window width is typically 1.2-1.5 times the transit time. The specific calculation process is as follows: multiply the transit time by 1.2 or 1.5 to obtain the window width. The start time of the time window is the transit time minus half the window width, and the end time is the transit time plus half the window width. Extract the modulus sequence segment within this time window, which is the amplitude envelope curve. The sound wave transit time window refers to the time range corresponding to the reflection of the ultrasonic signal at the diamond-solder interface. The modulus sequence within this range can accurately reflect the interface bonding state.

[0041] In one implementation, the Hilbert transform can be indirectly implemented using the fast Fourier transform, improving computational efficiency; the width of the time window can be dynamically adjusted according to the thickness of the packaging structure to ensure that the extracted modulus sequence can completely cover the interface reflection signal.

[0042] The above technical solution uses Hilbert transform to extract the amplitude envelope curve of the ultrasonic echo sequence, and combines the sound wave transit time window to screen effective signal segments, removing interference from irrelevant signals, solving the problems of low accuracy and high noise in existing envelope extraction methods, and improving the accuracy of acoustic impedance gradient vector extraction. In some embodiments, the theoretical thermal interface conductivity limit between the diamond layer and the solder layer is calculated based on the phonon mismatch model, including the following steps: Step 601: Obtain diamond surface roughness parameters. Specifically, the surface of the diamond substrate is scanned using an atomic force microscope (AFM). The scanning range is set to 10 μm × 10 μm, and the scanning resolution is not less than 0.1 nm. Surface roughness parameters, including arithmetic mean deviation and root mean square deviation, are extracted from the scanning results. The arithmetic mean deviation is usually taken as the core parameter, with a value range of 0.5-5 nm. The specific calculation process for the arithmetic mean deviation is as follows: the absolute value of the difference between the height value of each point on the scanned diamond surface and the average height value is taken, and the arithmetic mean of all absolute values ​​is calculated to obtain the arithmetic mean deviation. The specific calculation process for the root mean square deviation is as follows: the difference between the height value of each point on the scanned diamond surface and the average height value is squared, the arithmetic mean of all squared values ​​is calculated, and the square root of the arithmetic mean is taken to obtain the root mean square deviation. Diamond surface roughness parameters are indicators characterizing the flatness of the diamond surface and directly affect the contact state between the diamond and the solder layer and the phonon transmission efficiency.

[0043] Step 602: Calculate the microscopic scattering attenuation coefficient of phonon transport parameters under diamond surface roughness parameters. Specifically, based on the Maxwell-Boltzmann distribution theory and combined with the diamond surface roughness parameters, a phonon scattering model is established. The specific calculation process is as follows: calculate the collision probability of phonons on the diamond surface according to the diamond surface roughness parameters; calculate the phonon velocity distribution using the Maxwell-Boltzmann distribution; calculate the scattering probability of phonons on the diamond surface based on the collision probability and velocity distribution; calculate the microscopic scattering attenuation coefficient based on the relationship between the scattering probability and the phonon mean free path, specifically by subtracting the scattering probability from 1 to obtain the microscopic scattering attenuation coefficient. The value range of this coefficient is 0.8-0.95. The greater the surface roughness, the greater the scattering probability, the smaller the attenuation coefficient, and the more severe the phonon scattering. The microscopic scattering attenuation coefficient is an index characterizing the degree of phonon scattering on the diamond surface and is used to correct the phonon transport parameters.

[0044] Step 603: Input the microscopic scattering attenuation coefficient and the Debye temperature constant into the phonon mismatch model and perform algebraic multiplication to obtain the theoretical thermal interface conductivity limit. The specific calculation process is as follows: First, multiply the microscopic scattering attenuation coefficient by the phonon mean free path of the diamond layer to obtain the corrected phonon mean free path; then multiply the microscopic scattering attenuation coefficient by the phonon velocity of the diamond layer to obtain the corrected phonon velocity; input the corrected phonon mean free path, the corrected phonon velocity, and the Debye temperature constant of the solder layer into the phonon mismatch model; in the phonon mismatch model, perform algebraic multiplication of the corrected phonon mean free path, the corrected phonon velocity, and the Debye temperature constant in sequence to obtain the theoretical thermal interface conductivity limit considering the influence of surface roughness.

[0045] In one implementation, if a coating layer exists on the diamond surface, the thickness and roughness parameters of the coating layer can be measured additionally and incorporated into the calculation of the microscopic scattering attenuation coefficient, thereby further improving the accuracy of the theoretical thermal interface conductivity limit.

[0046] By combining the above technical solutions with the calculation of the microscopic scattering attenuation coefficient based on the diamond surface roughness parameters, the theoretical thermal interface conductivity limit is corrected. This solves the calculation deviation problem caused by the failure of the existing phonon mismatch model to consider the influence of surface roughness, making the Kapitza thermal resistance reference matrix more consistent with the actual surface state of the diamond substrate and improving the accuracy of the interface thermal resistance distribution calculation.

[0047] In some embodiments, a method for intelligently evaluating the packaging quality of a diamond substrate heat sink includes the following steps: Outputting a structural failure early warning command containing information on the locations of all defect nodes. Step 701: Apply the eight-neighbor connected component search algorithm to all defective nodes to generate discrete defect patch regions. Specifically, the execution process of the eight-neighbor connected component search algorithm is as follows: Taking each defective node as the center, determine the eight adjacent nodes in the directions of top, bottom, left, right, and diagonal; determine whether each adjacent node is a defective node one by one. If an adjacent node is also a defective node, it is determined that the adjacent node and the center node belong to the same connected component; traverse all defective nodes and cluster the interconnected defective nodes into a defect patch region, finally generating multiple discrete defect patch regions; a defect patch region refers to a continuous region composed of multiple adjacent defective nodes, representing concentrated defects in the encapsulation structure.

[0048] Step 702: Calculate the total pixel area and geometric eccentricity of each defect patch region. Specifically, the calculation process for the total pixel area is as follows: First, count the number of defect nodes, i.e., the number of pixels, contained in each defect patch region; obtain the preset pixel grid size, usually 0.1mm × 0.1mm, and calculate the area of ​​a single pixel, i.e., 0.1mm multiplied by 0.1mm, to obtain 0.01mm²; multiply the number of defect nodes by the area of ​​a single pixel to obtain the total pixel area of ​​the defect patch region, in mm². The calculation process for the geometric eccentricity is as follows: First, find the smallest circumscribed ellipse of the defect patch region and measure the length of the major axis and the length of the minor axis of the ellipse; divide the length of the major axis by the length of the minor axis, and the resulting ratio is the geometric eccentricity; the geometric eccentricity ranges from 0 to 1. The closer it is to 1, the thinner and longer the defect patch; the closer it is to 0, the closer the defect patch is to a circle. The total pixel area represents the size of the defect, and the geometric eccentricity represents the shape characteristics of the defect.

[0049] Step 703: Extract rating tags from a preset rating lookup table based on the sum of pixel areas and geometric eccentricity, and generate and output a structural failure warning command containing the rating tags. Specifically, the preset rating lookup table is pre-defined according to the size and shape of the defect. For example: a first-level warning is given if the sum of pixel areas is less than 0.1 mm² and the eccentricity is less than 0.5, i.e., a minor defect; a second-level warning is given if the sum of pixel areas is greater than or equal to 0.1 mm² and less than 1 mm², and the eccentricity is greater than or equal to 0.5 and less than 0.8, i.e., a moderate defect; and a third-level warning is given if the sum of pixel areas is greater than or equal to 1 mm² and the eccentricity is greater than or equal to 0.8, i.e., a severe defect. Match the corresponding rating tags in the lookup table according to the sum of pixel areas and geometric eccentricity of each defect patch area. After organizing the positions, sizes, and rating tags of all defect patches, generate and output a structural failure warning command.

[0050] In one implementation, the eight-neighbor connected domain search algorithm can be implemented using recursive search or iterative search. Iterative search is more suitable for clustering large-scale defect nodes, improving processing efficiency. The preset rating lookup table can be dynamically adjusted according to the application scenario of diamond substrate heat sink to meet the quality control needs of different industries.

[0051] By using the above technical solution, cluster analysis is performed on defective nodes, and graded early warning is given based on defect size and shape. This solves the problem that existing early warning methods only focus on the location of defects and do not consider the severity of defects, making the early warning instructions more targeted and providing a clear priority basis for packaging quality rectification.

[0052] In some embodiments, a smart evaluation method for the packaging quality of a diamond substrate heat sink includes the following steps before calculating the product of the intermediate probability distribution matrix and the first value vector: Step 801: Based on the acoustic impedance gradient vector, the residual stress field distribution inside the diamond heat sink packaging structure is obtained by inversion using the acoustoelastic theory equation. Specifically, the acoustoelastic theory equation uses the generalized Hooke's law combined with the relationship between sound velocity and stress, expressed as v = v0(1 + Kσ), where v is the propagation speed of ultrasound in the packaging structure, v0 is the sound velocity in the stress-free state, K is the acoustoelastic coefficient, and σ is the residual stress. The specific inversion process is as follows: extract the value of the ultrasonic propagation speed v from the acoustic impedance gradient vector; obtain the pre-measured sound velocity v0 and acoustoelastic coefficient K in the stress-free state; substitute v, v0, and K into the acoustoelastic theory equation, transforming it to obtain σ = (v / v0-1) / K, and calculate the residual stress value inside the packaging structure using this formula; arrange the residual stress values ​​corresponding to each grid according to the position of the evaluation pixel grid to generate a residual stress field distribution matrix; the residual stress field distribution refers to a two-dimensional matrix characterizing the magnitude and direction of residual stress at different positions of the packaging structure. Residual stress affects the correlation of acoustic and thermal characteristics, and thus affects the accuracy of interface thermal resistance calculation.

[0053] Step 802: Input the residual stress field distribution as a constraint condition into the first key value vector and perform stress-weighted correction on the first key value vector. The specific calculation process is as follows: extract the stress value corresponding to the first key value vector from the residual stress field distribution matrix; convert the stress value into a weighting coefficient, the conversion rule is that the greater the stress, the smaller the weighting coefficient, and the value range of the weighting coefficient is 0.5-1. The specific conversion method is: subtract the ratio of the stress value to the maximum residual stress value from 1, multiply by 0.5, and add 0.5 to obtain the weighting coefficient; perform element-wise multiplication operation on the corresponding element of the first key value vector, that is, multiply the weighting coefficient by the first element of the first key value vector to obtain the first element after correction, and so on, to complete the stress-weighted correction of the first key value vector. The purpose of stress-weighted correction is to reduce the interference of the residual stress concentration area on the acoustic and thermal characteristic correlation calculation, so that the first key value vector is more in line with the actual encapsulation state.

[0054] In one implementation, the acoustoelastic coefficient K can be obtained through experimental testing. A standard sample made of the same material as the encapsulation structure is selected, different stresses are applied, the corresponding ultrasonic propagation velocities are measured, and the K value is obtained by fitting. The inversion of the residual stress field distribution can be optimized using the least squares method to improve the inversion accuracy. By using the above technical solution, combined with the acoustoelastic theory to invert the residual stress field distribution, the first bond value vector is weighted and corrected, which solves the problem of correlation calculation deviation caused by the failure to consider the influence of residual stress in the construction of the existing attention weight matrix, improves the accuracy of the acoustic-thermal cross-attention weight matrix, and thus improves the calculation accuracy of the interface thermal resistance distribution.

[0055] In some embodiments, a method for intelligently evaluating the packaging quality of a diamond substrate heat sink sheet, after obtaining the transient thermal reflection attenuation sequence and ultrasonic echo sequence of the diamond heat sink sheet packaging structure, further includes the following steps: 1. Obtain X-ray diffraction rocking curve data of the diamond heat sink packaging structure and extract the corresponding full width at half maximum (FWHM) broadening value. Specifically, the diamond substrate heat sink is tested using an X-ray diffraction testing system. The X-ray wavelength is CuKα rays with a wavelength of 1.54 Å, the scanning angle range is 2θ = 40°-100°, and the scanning speed is 0.5° / min. The X-ray diffraction rocking curve is obtained, and the FWHM broadening value of the curve is calculated using the Lorentz fitting method. The specific calculation process is as follows: the rocking curve is fitted with a Lorentz function. Obtain the fitted curve; find the peak amplitude of the fitted curve, calculate half of the peak amplitude, and obtain the half-peak value; find the two angular coordinates on the fitted curve corresponding to the amplitude equal to the half-peak value, one is the angle to the left of the peak, and the other is the angle to the right of the peak; subtract the left angle from the right angle, and the difference is the full width at half maximum (FWHM) value; this value characterizes the degree of distortion of the diamond lattice, and the greater the lattice distortion, the greater the FWHM value; X-ray diffraction rocking curve data is the core data characterizing the integrity of the diamond lattice, and the FWHM value is positively correlated with the degree of lattice distortion.

[0056] Step 902: Concatenate the half-height width expansion value and the thermal time constant vector to generate a fused feature vector. The specific calculation process is as follows: convert the half-height width expansion value (a scalar) into a one-dimensional vector, which is a 1x1 matrix; the thermal time constant vector is a 1x1 matrix. _in' A matrix of columns; concatenate two vectors along their feature dimensions, that is, append the 1-row, 1-column vector of half-height and width expansion values ​​to the right of the thermal time constant vector, resulting in a 1-row (d... _in' +1) The fused feature vector of the column; the fused feature vector contains information on thermal conductivity characteristics, i.e., thermal time constant, and lattice integrity, i.e., half-width at half-maximum (FWHM) broadening value, which can more comprehensively characterize the packaging quality.

[0057] Step 903: Input the fused feature vector into the pre-trained interface thermal resistance prediction model to correct the initial value of the acoustic-thermal cross-attention weight matrix. Specifically, the pre-trained interface thermal resistance prediction model adopts a deep learning model, such as MLP or CNN. The training samples are fused feature vectors of diamond substrate heat sinks with different lattice distortion degrees and different packaging quality, and the corresponding actual values ​​of interface thermal resistance. The model is trained with the goal of minimizing the prediction error. The generated fused feature vector is input into the model to obtain the interface thermal resistance prediction value. The correction coefficient is calculated. The specific calculation process is as follows: divide the interface thermal resistance prediction value by the preset ideal interface thermal resistance value to obtain the correction coefficient. Multiply the correction coefficient by the initial value of the acoustic-thermal cross-attention weight matrix to obtain the corrected initial value, so that the initial value is closer to the ideal calculation range.

[0058] In one implementation, X-ray diffraction testing can be performed simultaneously with transient thermal reflection testing and ultrasonic testing to ensure that the spatial location of the test data corresponds one-to-one; the interface thermal resistance prediction model can adopt an online incremental learning method to continuously optimize the model parameters based on new test data and improve the correction accuracy.

[0059] By introducing X-ray diffraction rocking curve data and constructing a fused feature vector, the initial value of the attention weight matrix is ​​corrected by combining it with the interface thermal resistance prediction model. This solves the problem of insufficient accuracy of the weight matrix caused by unreasonable initial value setting in the existing calculation, and further improves the calculation accuracy of the interface thermal resistance distribution matrix.

[0060] Example 2: Please refer to Figure 2 An intelligent evaluation system for the packaging quality of a diamond substrate heat sink, as described in this embodiment of the invention, includes: Signal acquisition module: Used to acquire transient thermal reflection attenuation sequence and ultrasonic echo sequence of diamond heat sink package structure; this module integrates transient thermal reflection test unit and ultrasonic detection unit. The transient thermal reflection test unit includes femtosecond pulse laser, continuous semiconductor laser, photodetector and data acquisition card, and the ultrasonic detection unit includes high-frequency line focusing probe, preamplifier, filtering circuit and data acquisition system; the two test units adopt synchronous triggering mechanism to ensure that the spatial positions of the two acquired sequences correspond one-to-one, and the data sampling frequency is not less than 1MHz to ensure the integrity and accuracy of the signal; The time constant vector generation module is used to construct a discrete-time exponential decay model corresponding to the transient thermal reflection decay sequence; based on the discrete-time exponential decay model, multiple sets of fitting residual values ​​are obtained; when the fitting residual value is lower than the preset residual tolerance threshold, the corresponding decay coefficient value is extracted to generate a thermal time constant vector; this module has a built-in model fitting unit and a residual judgment unit. The model fitting unit uses the least squares method to achieve multi-exponential fitting, and the residual judgment unit has a preset residual tolerance threshold that can be dynamically adjusted according to the complexity of the encapsulation structure. It also integrates an inverse Laplace transform preprocessing unit to improve the extraction accuracy of the thermal time constant vector. Acoustic impedance gradient vector generation module: used to extract the amplitude envelope curve of the ultrasonic echo sequence; perform differential calculation on the amplitude envelope curve according to the preset spatial step size to generate a spatial reflectivity difference sequence; generate an acoustic impedance gradient vector based on the spatial reflectivity difference sequence and the preset local impedance constant; this module has built-in envelope extraction unit, differential calculation unit and vector generation unit. The envelope extraction unit is implemented using Hilbert transform. The differential calculation unit can adjust the spatial step size according to the ultrasonic scanning accuracy. The local impedance constant can be preset and dynamically updated according to the encapsulation material parameters. The acoustic-thermal cross-attention weight matrix construction module uses the acoustic impedance gradient vector as a query condition and the thermal time constant vector as a key-value sequence and value sequence to construct the acoustic-thermal cross-attention weight matrix through matrix multiplication. This module has built-in weight mapping unit, dot product calculation unit, normalization unit, and matrix multiplication unit. The weight mapping unit stores the pre-trained first, second, and third weight mapping matrices and can be updated online. It also integrates residual stress correction unit and initial value correction unit to optimize the construction accuracy of the weight matrix. Interface thermal resistance distribution matrix generation module: This module is used to perform Hadamard product operations based on the acoustic-thermal cross-attention weight matrix and the preset Kapitza thermal resistance reference matrix to generate the interface thermal resistance distribution matrix of the diamond heat sink packaging structure. The module has a built-in reference matrix storage unit and a Hadamard product operation unit. The reference matrix storage unit pre-stores Kapitza thermal resistance reference matrices for different packaging structures, which can be called or updated according to the packaging material parameters. The operation unit supports large-scale matrix parallel operation to improve processing efficiency. Failure early warning module: This module compares the matrix elements in the interface thermal resistance distribution matrix with a preset failure thermal resistance threshold; when a matrix element is greater than the failure thermal resistance threshold, the corresponding grid coordinates are marked as a defect node; and it outputs a structural failure early warning command containing the location information of all defect nodes. This module has built-in threshold comparison unit, defect node marking unit, cluster analysis unit, and early warning command generation unit. The cluster analysis unit uses an eight-neighbor connected component search algorithm, and the early warning command generation unit can generate graded early warning commands according to the severity of the defects, and supports real-time output and storage of commands.

[0061] In one implementation, the system also includes a data storage module and a human-computer interaction module. The data storage module is used to store the collected raw signals, intermediate calculation results, and final evaluation results. The human-computer interaction module is used for parameter setting, result display, and command issuance, and supports online real-time monitoring and historical data query.

[0062] By employing the aforementioned technical solutions, based on the collaborative acquisition of transient thermal reflection and ultrasonic detection, the thermal time constant vector and acoustic impedance gradient vector are extracted separately. An intrinsic correlation between these two features is established through an acoustic-thermal cross-attention mechanism, fully leveraging complementary information and addressing the issues of insufficient accuracy and superficial fusion of acoustic and thermal features in single detection methods. Through multi-level optimization techniques such as Laplace inverse transform preprocessing, Hilbert transform envelope extraction, residual stress correction, and lattice distortion information fusion, the accuracy of feature extraction and weight matrix construction is improved. A Kapitza thermal resistance benchmark matrix is ​​constructed based on a phonon mismatch model combined with surface roughness parameters, ensuring the targeted nature of interface thermal resistance distribution calculations. Accurate defect location and severity assessment are achieved through eight-neighborhood connected domain clustering and hierarchical early warning. An integrated intelligent assessment system is constructed to achieve fully automated assessment, adapting to the needs of industrial online testing. In summary, this invention effectively solves the industry problems of low accuracy, poor reliability, weak versatility, and low intelligence level in existing diamond substrate heat sink packaging quality assessment, improving the accuracy, practicality, and efficiency of packaging quality assessment, and providing reliable technical support for the packaging quality control of diamond substrate heat sinks.

[0063] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligently evaluating the packaging quality of diamond substrate heat sink sheets, characterized in that, include: Obtain the transient thermal reflection attenuation sequence and ultrasonic echo sequence of the diamond heat sink packaging structure; Construct a discrete-time exponential decay model corresponding to the transient thermal reflection decay sequence; Based on the discrete-time exponential decay model, multiple sets of fitting residual values ​​are obtained; when the fitting residual value is lower than the preset residual tolerance threshold, the corresponding decay coefficient value is extracted to generate a thermal time constant vector. Extract the amplitude envelope curve of the ultrasonic echo sequence; perform differential calculation on the amplitude envelope curve according to a preset spatial step size to generate a spatial reflectivity difference sequence; generate an acoustic impedance gradient vector based on the spatial reflectivity difference sequence and a preset local impedance constant. Using the acoustic impedance gradient vector as the query condition and the thermal time constant vector as the key sequence and value sequence, an acoustic-thermal cross-attention weight matrix is ​​constructed through matrix multiplication. The interface thermal resistance distribution matrix of the diamond heat sink packaging structure is generated by performing a Hadamard product operation based on the acoustic-thermal cross-attention weight matrix and the preset Kapitza thermal resistance reference matrix. The matrix elements in the interface thermal resistance distribution matrix are compared with a preset failure thermal resistance threshold; when the matrix element is greater than the failure thermal resistance threshold, the corresponding grid coordinates are marked as defect nodes. Output a structural failure early warning command containing the location information of all the aforementioned defect nodes.

2. The intelligent evaluation method for the packaging quality of a diamond substrate heat sink as described in claim 1, characterized in that, Using the acoustic impedance gradient vector as the query condition and the thermal time constant vector as the key sequence and value sequence, an acoustic-thermal cross-attention weight matrix is ​​constructed through matrix multiplication, including: Multiply the acoustic impedance gradient vector by the first weight mapping matrix to generate the first query vector; The thermal time constant vector is multiplied by the second weight mapping matrix and the third weight mapping matrix respectively to generate the first key value vector and the first value vector respectively; Calculate the dot product of the first query vector and the first key value vector, and divide the dot product result by the scaling factor; Apply the normalized exponential function to the result after dividing by the scaling factor to generate an intermediate probability distribution matrix; The product of the intermediate probability distribution matrix and the first value vector is calculated to obtain the acoustic-thermal cross-attention weight matrix.

3. The intelligent evaluation method for the packaging quality of a diamond substrate heat sink as described in claim 1, characterized in that, The Kapiza thermal resistance reference matrix is ​​obtained in advance in the following manner: Obtain the phonon transport parameters of the diamond layer and the Debye temperature constant of the solder layer in the diamond heat sink packaging structure; Based on the phonon mismatch model, the theoretical thermal interface conductivity limit between the diamond layer and the solder layer is calculated. The theoretical thermal interface conductivity limit is calculated by reciprocal and then filled with values ​​according to a preset evaluation pixel grid to generate the Kapiza thermal resistance reference matrix.

4. The intelligent evaluation method for the packaging quality of a diamond substrate heat sink as described in claim 1, characterized in that, Before obtaining multiple sets of fitted residual values ​​based on the discrete-time exponential decay model, the process also includes: Perform an inverse Laplace transform on the transient thermal reflection attenuation sequence to generate a continuous time constant spectrum; Calculate the peak half-width parameter in the continuous time constant spectrum; When the peak half-width at half-maximum parameter is less than the preset instrument dispersion threshold, the corresponding spectral range is retained for constructing the discrete-time exponential decay model.

5. The intelligent evaluation method for the packaging quality of a diamond substrate heat sink as described in claim 1, characterized in that, Extracting the amplitude envelope curve of the ultrasound echo sequence includes: Perform a Hilbert transform on the ultrasonic echo sequence to generate orthogonal analytic signals; Calculate the magnitude sequence of the orthogonal analytic signal; A sequence of modulus values ​​corresponding to the acoustic wave transit time window between the diamond lattice and the underlying solder is extracted and used as the amplitude envelope curve.

6. The intelligent evaluation method for the packaging quality of a diamond substrate heat sink as described in claim 3, characterized in that, Based on the phonon mismatch model, the theoretical thermal interface conductivity limit between the diamond layer and the solder layer is calculated, including: Obtain the surface roughness parameters of the diamond; Calculate the microscopic scattering attenuation coefficient of the phonon transport parameter under the diamond surface roughness parameter; The microscopic scattering attenuation coefficient and the Debye temperature constant are input into the phonon mismatch model and multiplied algebraically to obtain the theoretical thermal interface conductivity limit.

7. The intelligent evaluation method for the packaging quality of a diamond substrate heat sink as described in claim 1, characterized in that, Output a structural failure early warning command containing the location information of all the aforementioned defect nodes, including: An eight-neighbor connected component search algorithm is applied to all the defect nodes to generate discrete defect patch regions. Calculate the sum of pixel areas and geometric eccentricity for each of the defective patch regions; Based on the sum of the pixel areas and the value of the geometric eccentricity, a rating tag is extracted from a preset rating lookup table, and a structural failure warning instruction containing the rating tag is generated and output.

8. The intelligent evaluation method for the packaging quality of a diamond substrate heat sink as described in claim 2, characterized in that, Before calculating the product of the intermediate probability distribution matrix and the first value vector, the method further includes: Based on the acoustic impedance gradient vector, the residual stress field distribution inside the diamond heat sink packaging structure is obtained by inversion using the acoustic elasticity theory equation. The residual stress field distribution is input as a constraint condition into the first key value vector, and stress-weighted correction is performed on the first key value vector.

9. The intelligent evaluation method for the packaging quality of a diamond substrate heat sink as described in claim 1, characterized in that, After obtaining the transient thermal reflection attenuation sequence and ultrasonic echo sequence of the diamond heat sink packaging structure, the following is also included: Obtain the X-ray diffraction rocking curve data of the diamond heat sink packaging structure and extract the corresponding full width at half maximum (FWHM) broadening value; The half-width expansion value is concatenated with the thermal time constant vector to generate a fused feature vector. The fused feature vector is input into the pre-trained interface thermal resistance prediction model to correct the initial value of the acoustic-thermal cross-attention weight matrix.

10. A smart evaluation system for the packaging quality of a diamond substrate heat sink, applied to the smart evaluation method for the packaging quality of a diamond substrate heat sink as described in any one of claims 1-9, characterized in that, include: The signal acquisition module acquires the transient thermal reflection attenuation sequence and ultrasonic echo sequence of the diamond heat sink packaging structure; The time constant vector generation module constructs a discrete-time exponential decay model corresponding to the transient thermal reflection decay sequence; based on the discrete-time exponential decay model, it obtains multiple sets of fitting residual values; when the fitting residual value is lower than a preset residual tolerance threshold, it extracts the corresponding decay coefficient value and generates a thermal time constant vector. The acoustic impedance gradient vector generation module extracts the amplitude envelope curve of the ultrasonic echo sequence; performs differential calculation on the amplitude envelope curve according to a preset spatial step size to generate a spatial reflectivity difference sequence; and generates an acoustic impedance gradient vector based on the spatial reflectivity difference sequence and a preset local impedance constant. The acoustic-thermal cross-attention weight matrix construction module uses the acoustic impedance gradient vector as the query condition and the thermal time constant vector as the key sequence and value sequence, and constructs the acoustic-thermal cross-attention weight matrix through matrix multiplication. The interface thermal resistance distribution matrix generation module performs a Hadamard product operation based on the acoustic-thermal cross-attention weight matrix and the preset Kapitza thermal resistance reference matrix to generate the interface thermal resistance distribution matrix of the diamond heat sink packaging structure. The failure early warning module compares the matrix elements in the interface thermal resistance distribution matrix with a preset failure thermal resistance threshold; when the matrix element is greater than the failure thermal resistance threshold, the corresponding grid coordinates are marked as defect nodes. Output a structural failure early warning command containing the location information of all the aforementioned defect nodes.

Images