TSV packaging defect detection method and system

Abstract

The invention discloses a TSV packaging defect detection method and system, and the method comprises the steps: carrying out the multi-dimensional excitation scanning of a TSV packaging structure through employing an eddy current pulse thermal imaging device, and generating a three-dimensional eddy current field distribution feature matrix; inputting the three-dimensional eddy current field distribution characteristic matrix into a topological manifold decomposition module for defect characteristic separation, and constructing three characteristic components of a structure deformation topological ring, an interface fracture chain and a thermal stress abnormal curved surface based on a continuous coherence analysis method; performing cross-domain fusion processing of a dynamic heterogeneous neural network on the three feature components, and generating a defect topology fingerprint spectrum through a double-path attention gating mechanism; and inputting the defect topology fingerprint into a multi-scale entropy evaluation module for defect evolution simulation, and outputting a quantitative detection report containing defect geometric parameters, failure probability and reliability threshold. According to the embodiment of the invention, a comprehensive defect detection system can be constructed, and the accuracy and efficiency of defect identification are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of packaging monitoring, and in particular to a method and system for detecting TSV packaging defects. Background Art

[0002] In modern electronic devices, through-silicon via (TSV) technology, as an integrated circuit packaging technique, has been widely used due to its superior electrical performance and space-saving features. TSV technology can effectively reduce inter-chip interconnect latency, increase signal transmission rates, and reduce package size. However, with the advancement of packaging technology, the complexity and high integration density of TSV structures have increased the probability of defects during the manufacturing process, particularly packaging defects at the micro-nanoscale. If these defects are not discovered and addressed promptly, they can lead to decreased device performance and reliability, and even serious failure.

[0003] Traditional defect detection methods typically rely on optical imaging, X-ray imaging, or ultrasonic testing. These methods suffer from insufficient resolution, low sensitivity, and significant impact on material properties when detecting tiny defects. Furthermore, existing technologies are limited in their ability to handle the coupling between different physical fields, making it impossible to fully capture the multidimensional characteristics of defects, resulting in inaccurate defect identification or missed detections. Summary of the Invention

[0004] The purpose of the present invention is to provide a TSV package defect detection method and system to address the deficiencies in the prior art, to build a comprehensive defect detection system, and to improve the accuracy and efficiency of defect identification.

[0005] An embodiment of the present application provides a method for detecting TSV packaging defects, the method comprising: An eddy current pulse thermal imaging device is used to perform multi-dimensional excitation scanning on the TSV packaging structure. Based on an adaptive phase encoding algorithm, the original electromagnetic echo signal is subjected to multi-physics field coupling demodulation processing to generate a three-dimensional eddy current field distribution feature matrix. The multi-dimensional excitation scanning achieves cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads. The three-dimensional eddy current field distribution feature matrix is ​​input into the topological manifold decomposition module for defect feature separation. Based on the continuous homology analysis method, three feature components are constructed: structural deformation topological ring, interface fracture chain, and thermal stress abnormal surface. The topological manifold decomposition module uses dynamic topological reconstruction technology to eliminate multiple scattering noise and extract the geometric invariance characteristics of submicron defects. The three feature components are subjected to cross-domain fusion processing using a dynamic heterogeneous neural network, and a defect topology fingerprint is generated through a dual-path attention gating mechanism. The cross-domain fusion processing adopts a hybrid architecture of spatiotemporal causal convolutional network and graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy; The defect topology fingerprint map is input into a multi-scale entropy evaluation module to simulate defect evolution. The nonlinear mapping relationship between the defect extension path and the degradation of electrothermal performance is solved based on the simulated annealing algorithm, and a quantitative detection report containing defect geometric parameters, failure probability and reliability threshold is output. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy model and generates a feedback control signal to optimize the scanning parameter configuration.

[0006] Optionally, the eddy current pulse thermal imaging device is used to perform multi-dimensional excitation scanning on the TSV packaging structure, and the original electromagnetic echo signal is subjected to multi-physics field coupling demodulation processing based on the adaptive phase encoding algorithm to generate a three-dimensional eddy current field distribution feature matrix, wherein the multi-dimensional excitation scanning realizes cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads, including: Based on the dielectric constant distribution of the TSV packaging structure, a frequency combination sequence of orthogonal polarization electromagnetic fields is dynamically generated. The transverse electric field component and the longitudinal magnetic field component are alternately applied through the electromagnetic field emission array to stimulate the coupled response of surface eddy currents and deep thermal effects, and obtain the original electromagnetic echo signal. Performing dual-domain synchronous sampling processing on the original electromagnetic echo signal, using nonlinear pulse compression technology in the time domain to eliminate multipath interference, and extracting phase mutation characteristics of non-stationary signals through fractional Fourier transform in the frequency domain to generate a time-frequency joint distribution map; Based on the time-frequency joint distribution map, a multi-physics field coupling equation is constructed, and the alternating direction multiplier method ADMM is used to jointly solve the electromagnetic-heat conduction equation to obtain the vector distribution data of the three-dimensional eddy current density field and the temperature gradient field; The vector distribution data is input into a phase encoding matrix generator, and a chaotic sequence is used to dynamically encrypt and map the spatial coordinates to generate a three-dimensional eddy current field distribution characteristic matrix with spatial continuity and anti-aliasing characteristics.

[0007] Optionally, the three-dimensional eddy current field distribution feature matrix is ​​input into a topological manifold decomposition module for defect feature separation, and three feature components, namely, a structural deformation topological ring, an interface fracture chain, and a thermal stress anomaly surface, are constructed based on a continuous homology analysis method. The topological manifold decomposition module uses a dynamic topological reconstruction technique to eliminate multiple scattering noise and extract geometric invariance features of submicron defects, including: Performing super-pixel region division on the three-dimensional eddy current field distribution feature matrix, constructing a critical point topological skeleton of the three-dimensional gradient flow using discrete Morse theory, and generating an initial persistent homology feature map; Dynamic topology reconstruction technology is used to filter noise from the persistent homology feature graph. By calculating the lifetime entropy of each topological ring, the connection radius threshold of the Vietoris-Rips complex is dynamically adjusted to eliminate pseudo-topological structures caused by multiple scattering. Based on the Hodge decomposition theorem, the filtered topological skeleton is decomposed into three orthogonal subspaces: the spiral gradient field component caused by structural deformation, the vortex circulation component caused by interface fracture, and the divergence anomaly component formed by thermal stress anomaly. The geometric invariance of each subspace component is optimized, and the evolution path of the curvature tensor is constrained by the Seiberg-Witten equation to generate structural deformation topological rings, interface fracture chains and thermal stress anomaly surfaces with differential homeomorphism invariance, and three characteristic components are obtained.

[0008] Optionally, the three feature components are subjected to cross-domain fusion processing of a dynamic heterogeneous neural network, and a defect topology fingerprint is generated through a dual-path attention gating mechanism. The cross-domain fusion processing adopts a hybrid architecture of a spatiotemporal causal convolutional network and a graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy, including: The structural deformation topological ring is input into the spatiotemporal causal convolutional network, and the deformation evolution characteristics in the time dimension are extracted through the dilated convolution kernel with memory gating to generate a feature tensor containing multi-scale spatiotemporal correlations; The interface fracture chain is input into a graph neural network, a dynamic adjacency matrix is ​​constructed based on the nanoscale spatial coordinates of the fracture nodes, and a graph feature representing the topological connectivity of the pore defects is generated using an edge weight propagation algorithm; Performing three-dimensional curvature parameterization processing on the thermal stress anomaly surface, matching the surface gradient distribution through a deformable convolution kernel, and extracting the characteristic vector of the thermal stress anomaly area; Implement an asymmetric weight sharing strategy for cross-network feature fusion to generate enhanced features that integrate cross-physics field characteristics; The enhanced features are input into a dual-path attention gating mechanism. The first path uses channel-weighted attention to screen key frequency band features, and the second path strengthens the defect area response through a spatial topological focusing mechanism, and finally outputs a defect topology fingerprint map with cross-scale correlation.

[0009] Optionally, the defect topology fingerprint map is input into a multi-scale entropy evaluation module to perform defect evolution simulation, and a nonlinear mapping relationship between the defect extension path and the electrothermal performance degradation is solved based on a simulated annealing algorithm, and a quantitative detection report including defect geometric parameters, failure probability, and reliability threshold is output. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy model and generates a feedback control signal to optimize the scanning parameter configuration, including: The defect topological fingerprint map is input into a multi-scale entropy value evaluation module, and an atomic-mesoscopic-macroscopic three-level entropy value calculation model is constructed based on the defect geometric parameters. At the atomic scale, the atomic displacement entropy of the interface defect area is calculated based on the molecular dynamics simulation results; at the mesoscale, the morphological entropy change of the crack propagation path is simulated based on the phase field method; at the macroscale, a joint entropy function of electromagnetic conductivity, thermal resistivity and mechanical strength degradation is established; The Hamiltonian of the simulated annealing algorithm is constructed based on the three-level entropy calculation model, the defect expansion path is mapped to the spin state combination of the Ising model, and the nonlinear mapping relationship between electrothermal performance degradation and entropy value change is solved through the quantum tunneling effect; Based on the Monte Carlo method, a set of defect evolution paths under different service conditions is generated. The weight parameters of the entropy calculation model are dynamically corrected by comparing the KL divergence of the measured entropy change value with the simulated value. The optimized parameter combination of polarization field intensity ratio and heat load application period is generated according to the modified entropy value calculation model, and a feedback control signal is formed to return the scanning parameter configuration.

[0010] Optionally, the implementation of an asymmetric weight sharing strategy to perform cross-network feature fusion to generate enhanced features that fuse cross-physical field characteristics includes: A cross-domain connection layer is established between the spatiotemporal causal convolutional network and the graph neural network. The high-dimensional feature channels of the convolutional network and the node embedding layer of the graph neural network share a weight matrix, and asymmetric mapping is achieved through different nonlinear activation functions. The feature tensor is concatenated with the graph features, and the geometric similarity matrix S of the nanopore defects is calculated through the cross-scale association module, where S = ξ(F_t^T·W_2·F_g), where F_t is the feature tensor, T is the matrix transpose, F_g is the graph feature, ξ is the Sigmoid function, and W_2 is the learnable shared weight; A Hadamard product is performed on the geometric similarity matrix and the thermal stress eigenvectors to generate enhanced features that incorporate cross-physics properties.

[0011] Another embodiment of the present application provides a TSV package defect detection system, the system comprising: A scanning module is used to perform multi-dimensional excitation scanning on the TSV packaging structure using an eddy current pulse thermal imaging device, perform multi-physics field coupling demodulation processing on the original electromagnetic echo signal based on an adaptive phase encoding algorithm, and generate a three-dimensional eddy current field distribution feature matrix, wherein the multi-dimensional excitation scanning achieves cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads; A construction module is used to input the three-dimensional eddy current field distribution feature matrix into a topological manifold decomposition module to separate defect features, and construct three feature components: structural deformation topological rings, interface fracture chains, and thermal stress anomaly surfaces based on a continuous homology analysis method. The topological manifold decomposition module uses dynamic topological reconstruction technology to eliminate multiple scattering noise and extract geometric invariance features of submicron defects; A fusion module for performing cross-domain fusion processing of the three feature components using a dynamic heterogeneous neural network to generate a defect topology fingerprint through a dual-path attention gating mechanism. The cross-domain fusion processing adopts a hybrid architecture of a spatiotemporal causal convolutional network and a graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy; The output module is used to input the defect topology fingerprint map into the multi-scale entropy value evaluation module to simulate the defect evolution, solve the nonlinear mapping relationship between the defect extension path and the electrothermal performance degradation based on the simulated annealing algorithm, and output a quantitative detection report containing defect geometric parameters, failure probability and reliability threshold. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy value model and generates a feedback control signal to optimize the scanning parameter configuration.

[0012] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0013] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0014] Compared with the prior art, the present invention provides a method for detecting TSV packaging defects, which uses an eddy current pulse thermal imaging device to perform multi-dimensional excitation scanning on the TSV packaging structure to generate a three-dimensional eddy current field distribution feature matrix; the three-dimensional eddy current field distribution feature matrix is ​​input into a topological manifold decomposition module for defect feature separation, and three feature components of structural deformation topological ring, interface fracture chain and thermal stress abnormal surface are constructed based on the continuous homology analysis method; the three feature components are subjected to cross-domain fusion processing of a dynamic heterogeneous neural network, and a defect topology fingerprint map is generated through a dual-path attention gating mechanism; the defect topology fingerprint map is input into a multi-scale entropy value evaluation module for defect evolution simulation, and a quantitative detection report including defect geometric parameters, failure probability and reliability threshold is output, thereby building a comprehensive defect detection system and improving the accuracy and efficiency of defect identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A hardware structure block diagram of a computer terminal for a TSV packaging defect detection method provided by an embodiment of the present invention; Figure 2 A schematic flow chart of a method for detecting TSV packaging defects provided by an embodiment of the present invention; Figure 3 A schematic structural diagram of a TSV packaging defect detection system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0017] The embodiment of the present invention first provides a method for detecting TSV packaging defects. The method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0018] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a TSV packaging defect detection method provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0019] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can enable a processor to execute any TSV package defect detection method.

[0020] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0021] The internal memory provides an environment for running the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any TSV packaging defect detection method.

[0022] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0023] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0024] See also Figure 2 , an embodiment of the present invention provides a method for detecting TSV packaging defects, which may include the following steps: S201, performing multi-dimensional excitation scanning on the TSV packaging structure using an eddy current pulse thermal imaging device, performing multi-physics field coupling demodulation processing on the original electromagnetic echo signal based on an adaptive phase encoding algorithm, and generating a three-dimensional eddy current field distribution feature matrix, wherein the multi-dimensional excitation scanning achieves cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads; This method uses an eddy current pulse thermal imaging device to perform multi-dimensional excitation scanning of the TSV packaging structure. By alternating the application of orthogonal polarization electromagnetic fields and gradient thermal loads, cross-modal excitation of interface defects is achieved. Based on an adaptive phase encoding algorithm, multi-physics field coupled demodulation is performed on the original electromagnetic echo signal to generate a three-dimensional eddy current field distribution feature matrix. This multi-physics field coupled excitation can simultaneously stimulate the response signals of surface and deep-seated defects, improving detection sensitivity. The adaptive phase encoding algorithm effectively suppresses noise interference and provides high-quality three-dimensional field distribution data for subsequent defect feature extraction.

[0025] Specifically, based on the material dielectric constant distribution of the TSV packaging structure, a frequency combination sequence of orthogonally polarized electromagnetic fields can be dynamically generated. The transverse electric field component and the longitudinal magnetic field component are alternately applied through the electromagnetic field emission array to stimulate the coupled response of surface eddy currents and deep thermal effects, thereby obtaining the original electromagnetic echo signal. The dynamic frequency combination optimizes the electromagnetic field excitation efficiency, and the orthogonal polarization design realizes the coordinated excitation of multiple physical fields, providing richer signal characteristics for defect detection.

[0026] The original electromagnetic echo signal is subjected to dual-domain synchronous sampling processing, nonlinear pulse compression technology is used in the time domain to eliminate multipath interference, and fractional-order Fourier transform is used in the frequency domain to extract the phase mutation characteristics of the non-stationary signal to generate a time-frequency joint distribution map; the dual-domain processing technology effectively improves the signal-to-noise ratio of the signal, and the fractional-order Fourier transform enhances the feature extraction capability of the non-stationary signal, providing more accurate time-frequency characteristics for subsequent analysis.

[0027] Based on the time-frequency joint distribution map, a multi-physics field coupling equation is constructed, and the alternating direction multiplier method ADMM is used to jointly solve the electromagnetic-heat conduction equation to obtain the vector distribution data of the three-dimensional eddy current density field and the temperature gradient field; the multi-physics field coupling solution realizes the coordinated analysis of electromagnetic and thermal signals, and the ADMM algorithm improves the solution efficiency of large-scale equation groups and provides accurate field distribution data for defect location.

[0028] The vector distribution data is input into a phase-encoded matrix generator, which dynamically encrypts the spatial coordinates using a chaotic sequence to generate a three-dimensional eddy current field distribution matrix with spatial continuity and anti-aliasing properties. Chaotic coding enhances the spatial resolution of the data, while anti-aliasing prevents signal distortion, providing a high-quality input matrix for subsequent topological analysis.

[0029] In practical applications, a specific implementation may include: Step 2.1: Multi-dimensional excitation scanning and signal excitation Based on the dielectric constant distribution of the TSV packaging structure, a frequency combination sequence of orthogonally polarized electromagnetic fields is dynamically generated. The transverse electric field component and the longitudinal magnetic field component are alternately applied through the electromagnetic field emission array to stimulate the coupled response of surface eddy currents and deep thermal effects, and obtain the original electromagnetic echo signal. The specific implementation is as follows: 1. Dynamic analysis of dielectric constant: A broadband dielectric spectrometer is used to measure the dielectric constants of each layer of TSV packaging material (e.g., the dielectric constant of the silicon interposer ε_r = 11.7, and the dielectric constant of the oxide layer ε_r = 3.9).

[0030] Construct a frequency-dielectric response curve based on the material properties and select the characteristic frequency corresponding to the dielectric loss peak (for example, silicon has the highest loss at 2.45 GHz).

[0031] 2. Orthogonal polarization field sequence generation: Transverse electric field (TE mode): Excitation is performed using an interdigitated electrode array, with a frequency range of 1-5 GHz and a step of 50 MHz. The electric field direction is parallel to the TSV axis.

[0032] Longitudinal magnetic field (TM mode): Excitation is performed using a toroidal coil array with a frequency range of 0.5-3 GHz and a magnetic field direction perpendicular to the TSV axis.

[0033] Alternating excitation strategy: Each polarization mode lasts for 10 μs, and the alternating period is 20 μs to ensure that the thermal-electric coupling effect is fully excited.

[0034] 3. Thermal load gradient loading: The infrared laser array scans the TSV surface in a checkerboard pattern with a gradient power density distribution (5 W / cm² in the center and 2 W / cm² at the edge).

[0035] The thermal excitation frequency is synchronized with the alternating period of the electromagnetic field to form thermal-electric synergistic modulation.

[0036] Output: A raw data set containing time-varying electromagnetic echo signals (sampling rate 10 GS / s) and thermal response curves (infrared camera frame rate 100 Hz).

[0037] Step 2.2: Dual-domain simultaneous sampling and signal demodulation The original electromagnetic echo signal is subjected to dual-domain synchronous sampling processing. Nonlinear pulse compression technology is used in the time domain to eliminate multipath interference. In the frequency domain, fractional Fourier transform is used to extract the phase mutation characteristics of the non-stationary signal to generate a time-frequency joint distribution map. The specific implementation is as follows: 1. Time domain nonlinear compression: An adaptive matched filter is designed to perform pulse compression on the echo signal based on the linear frequency modulation characteristics of the transmitted signal (bandwidth 2 GHz, pulse width 1 μs).

[0038] The minimum mean square error (MMSE) algorithm is used to suppress multipath interference (such as the substrate reflection signal attenuation of 20dB).

[0039] 2. Fractional Fourier Transform (FrFT): The signal energy distribution is calculated in the fractional domain with a rotation angle of α=0.35π, and the phase jumps caused by defects (jump amplitude > 5°) are separated.

[0040] Construct a time-frequency ridge tracking map and mark the time-frequency coordinates corresponding to the phase mutation points (such as t=3.2ms, f=2.1GHz).

[0041] 3. Joint time-frequency analysis: The time-domain compression signal is fused with the fractional-order domain features to generate a time-frequency energy density matrix with a resolution of 0.1mm×0.1GHz.

[0042] The potential defect signal is extracted by threshold segmentation (energy > background noise 3τ).

[0043] Output: joint time-frequency distribution map (matrix size 1000×500, unit dBm / Hz).

[0044] Step 2.3: Solving the multiphysics coupled equations Based on the time-frequency joint distribution map, a multi-physics coupling equation is constructed. The alternating direction multiplier method (ADMM) is used to jointly solve the electromagnetic-heat conduction equation to obtain the vector distribution data of the three-dimensional eddy current density field and temperature gradient field. The specific implementation is as follows: 1. Discretization of equations: Electromagnetic field equations: Maxwell's equations were discretized into a finite element mesh (element size 10 μm) with the boundary condition set as a perfectly matched layer (PML).

[0045] Heat conduction equation: Use the unsteady Fourier law with a time step of 1 μs and a spatial step of 50 μm.

[0046] 2. ADMM Iteration Process: Variable splitting: Decouple electromagnetic field variables (E, H) from temperature field variables (T).

[0047] Alternating solution: The temperature field is fixed and the electromagnetic field is solved using the FDTD method (convergence condition: residual < 1e-4).

[0048] The electromagnetic field is fixed and the temperature field is solved using the finite volume method (convergence condition: temperature change <0.1K / step).

[0049] Multiplier Update: Adjust the Lagrange multipliers based on electromagnetic-thermal coupling terms (such as Joule heating terms).

[0050] 3. Vector field reconstruction: Eddy current density field: Calculated by ▽×H to obtain a three-dimensional vector distribution (unit: A / m²).

[0051] Temperature gradient field: The three-dimensional heat flow vector (unit: K / m) is obtained by calculating ▽T.

[0052] Output: Vector datasets of eddy current density field (resolution 10 μm) and temperature gradient field (resolution 50 μm).

[0053] Step 2.4: Generate three-dimensional eddy current field characteristic matrix The vector distribution data is input into the phase encoding matrix generator, and the spatial coordinates are dynamically encrypted and mapped using a chaotic sequence to generate a three-dimensional eddy current field distribution characteristic matrix with spatial continuity and anti-aliasing characteristics. The specific implementation is as follows: 1. Chaotic sequence design: Logistic mapping is used to generate chaotic sequence: x_{n+1}=μx_n(1-x_n), μ=3.99, initial value x_0=0.1234.

[0054] The sequence length matches the number of spatial grids (e.g., 1000×1000×100), and a three-dimensional coordinate mapping table is generated after quantization.

[0055] 2. Phase encoding strategy: Map the eddy current density amplitude to the phase angle: φ=arctan(J_y / J_x), in the range [-π, π].

[0056] A chaotic sequence is used to dynamically scramble the phase angle to eliminate periodic artifacts.

[0057] 3. Anti-aliasing processing: A Kaiser window (β = 6) is applied in the frequency domain to suppress the Gibbs phenomenon.

[0058] Bilinear interpolation is used to ensure spatial continuity.

[0059] Output: three-dimensional eddy current field distribution characteristic matrix (size 1000×1000×100, data type float32).

[0060] Application scenario example: Silicon interposer TSV packaging defect detection Test object: A 7nm silicon interposer TSV structure containing 2000 copper-filled vias with a diameter of 5μm and a depth of 50μm. After packaging, interface microcracks and nanopores appeared.

[0061] Step 1: Multi-dimensional excitation scanning Frequency combination: TE mode main frequency 2.45GHz (matching silicon dielectric loss peak), TM mode main frequency 1.8GHz (matching oxide layer characteristics).

[0062] Thermal load configuration: laser power gradient loading, 5W / cm² in the center area (covering the TSV array) and 2W / cm² at the edge (covering the passivation layer).

[0063] Response signal: An abnormal echo signal was captured (amplitude exceeded the standard by 30%), and the infrared thermal image showed an abnormal local temperature rise (ΔT=3.2K).

[0064] Step 2: Dual-Domain Signal Processing Time domain compression: The matched filter compresses the 1μs pulse width to 10ns, improving the signal-to-noise ratio to 45dB and separating the direct wave from the defect reflection wave.

[0065] Fractional-order transform: A phase jump (Δφ=7.8°) was detected in the α=0.35π domain, corresponding to the defect location (x=1.2mm, y=0.8mm).

[0066] Time-frequency spectrum: Three energy anomaly areas are marked (frequency 2.1GHz / 2.4GHz / 2.7GHz), indicating defects of different depths.

[0067] Step 3: Multiphysics coupling solution Electromagnetic field distribution: The solution shows abnormal eddy current density at the edge of the TSV array (J_max=1.2e6 A / m² vs. average 8e5 A / m²).

[0068] Temperature field distribution: The direction of the heat flow vector in the defective area is disordered, forming a clear contrast with the intact area.

[0069] Step 4: Feature matrix generation Chaotic encryption: The physical coordinate (1.2, 0.8, 0.05) mm is mapped to the encrypted coordinate (0.763, 0.291, 0.418) through the Logistic sequence.

[0070] Phase encoding: The phase angle distribution in the defect area presents a spiral feature (φ gradually changes from -π / 2 to π / 2).

[0071] Anti-aliasing effect: After Kaiser window processing, the artifact energy is reduced to below -60dB.

[0072] Test results: Defect 1: Interface microcracks (8μm in length and 2μm in depth) located at the TSV-passivation layer interface, exceeding the reliability threshold by 23%.

[0073] Defect 2: Nanopores (diameter 200nm, density 5e8 / m³), resulting in a 15% decrease in conductivity.

[0074] Quantitative report: Failure probability is 0.17% (microcracks) and 0.09% (pores). It is recommended to optimize the electroplating process parameters.

[0075] Among them, the τ in the background noise 3τ is: Parameter meaning: τ (Sigma) represents the standard deviation, which is an indicator of data fluctuation in statistics.

[0076] Application scenario: In signal processing, the 3τ principle is used to identify abnormal signals. If the signal energy exceeds three standard deviations (3τ) of the background noise average, it is considered a valid defect signal.

[0077] Example: If the average background noise energy is -60dBm and the standard deviation τ = 2dBm, the threshold is set to -60 + 3×2 = -54dBm. Signals exceeding -54dBm are considered potential defects.

[0078] Electromagnetic field variables (E, H): E (electric field strength): Physical meaning: Describes the strength of the force exerted by the electric field on the charge, with the unit being volt / meter (V / m).

[0079] Application scenario: In TSV inspection, the distribution of E reflects the dielectric properties of the material, and E in the defective area will be abnormally enhanced or distorted.

[0080] H (magnetic field strength): Physical meaning: Describes the excitation effect of the magnetic field on the electric current, the unit is ampere / meter (A / m).

[0081] Application scenario: The curl of H (▽×H) is used to calculate eddy current density and locate deep defects.

[0082] ▽×H (magnetic field curl): Mathematical meaning: The curl operator ▽× represents the spatial rate of change of the magnetic field.

[0083] Physical meaning: According to Maxwell's equations, ▽ × H = J + D / t, where J is the current density and D is the electric displacement field.

[0084] Application scenario: By calculating the curl of H, the eddy current density distribution (unit: A / m²) is obtained for detecting cracks inside TSVs.

[0085] ▽T (temperature gradient): Physical meaning: The rate of change of temperature in space, the direction points to the direction of the fastest temperature increase.

[0086] Unit: Kelvin / meter (K / m).

[0087] Application scenario: Under thermal load excitation, the defect area may experience abnormal ▽T due to uneven heat conduction, such as a sudden increase in temperature gradient at the crack.

[0088] Logistic mapping parameters (x_{n+1}, μ, n, x_n): x_n: current iteration value, ranging from [0,1], representing the instantaneous state of the chaotic sequence.

[0089] x_{n+1}: The iterative value at the next moment, calculated by the formula x_{n+1} = μx_n(1-x_n).

[0090] μ: Control parameter (usually 3.99), which determines whether the system is in a chaotic state.

[0091] n: number of iterations (integer).

[0092] J_y, J_x (current density components): Physical meaning: The y- and x-direction components of current density in a rectangular coordinate system, with the unit ampere / square meter (A / m²).

[0093] Application scenario: In TSV defect detection, the vector synthesis of J_x and J_y can characterize the eddy current direction, such as when a crack causes the eddy current to detour (a sudden change in the J direction).

[0094] Kaiser window parameter β=6: β (Beta): Window function shape parameter that controls the trade-off between main lobe width and side lobe attenuation.

[0095] Physical meaning: When β=6, the sidelobe attenuation of the Kaiser window is about -60dB, and the mainlobe width is moderate, which is suitable for suppressing spectrum leakage.

[0096] Application example: After windowing the eddy current signal, the energy of the Gibbs oscillation (spectral artifact) is reduced from -40dB to below -60dB.

[0097] φ (phase angle): Physical meaning: Indicates the phase shift of the eddy current signal, range [-π, π], unit radian (rad).

[0098] Application scenario: Defects cause changes in the electromagnetic wave reflection path, resulting in a sudden change in the phase angle φ (for example, from -π / 2 to π / 2).

[0099] 5e8 / m³ (nanopore density): Scientific notation analysis: 5×10 8 pores / cubic meter, that is, there are 500,000,000 pores per cubic meter.

[0100] Practical significance: If the TSV volume is 1 mm³ (1e-9 m³), ​​it contains about 500 pores, which significantly affects conductivity.

[0101] S202: Inputting the three-dimensional eddy current field distribution feature matrix into a topological manifold decomposition module to separate defect features, and constructing three feature components: structural deformation topological rings, interface fracture chains, and thermal stress anomaly surfaces based on a continuous homology analysis method. The topological manifold decomposition module uses dynamic topological reconstruction technology to eliminate multiple scattering noise and extract geometric invariance features of submicron defects. The three-dimensional eddy current field distribution matrix is ​​input into the topological manifold decomposition module for defect feature separation. Based on the persistent coherence analysis method, three characteristic components are constructed: structural deformation topological rings, interface fracture chains, and thermal stress anomaly surfaces. The topological manifold decomposition module uses dynamic topological reconstruction technology to eliminate multiscattering noise. This topological manifold decomposition enables precise separation of defect features and can distinguish different types of defect patterns. Dynamic topological reconstruction technology effectively eliminates noise interference during the inspection process, improving the detection accuracy of submicron defects.

[0102] Specifically, the three-dimensional eddy current field distribution feature matrix can be divided into regions at the superpixel level, and the discrete Morse theory can be used to construct the critical point topological skeleton of the three-dimensional gradient flow to generate an initial persistent homological feature map; superpixel division improves the continuity of regional features, and discrete Morse theory realizes the precise extraction of topological features, establishing a reliable topological representation for defect identification.

[0103] Dynamic topological reconstruction technology is used to filter noise from the persistent coherent feature map. By calculating the lifetime entropy value of each topological ring, the connection radius threshold of the Vietoris-Rips complex is dynamically adjusted to eliminate pseudo-topological structures caused by multiple scattering. The dynamic reconstruction technology effectively suppresses noise interference, and the adaptive adjustment of the entropy value optimizes the extraction accuracy of the topological features, thereby improving the detection capability of submicron defects.

[0104] Based on the Hodge decomposition theorem, the filtered topological skeleton is decomposed into three orthogonal subspaces: the spiral gradient field component caused by structural deformation, the vortex circulation component caused by interface fracture, and the divergence anomaly component formed by thermal stress anomaly; the orthogonal decomposition realizes the separation of different types of defect features and provides a structured feature representation for subsequent targeted analysis.

[0105] By optimizing the geometric invariance of each subspace component and constraining the curvature tensor's evolution path via the Seiberg-Witten equation, we generate three characteristic components: a topological ring of structural deformation, an interface fracture chain, and a thermal stress anomaly surface that exhibit diffeomorphic invariance. This geometric invariance enhances the robustness of the features to varying detection conditions, while the diffeomorphic constraints ensure the mathematical rigor of the feature representation.

[0106] Step 3.1: Superpixel-level region division and initial topology skeleton construction 1. Superpixel Segmentation: SLIC algorithm (Simple Linear Iterative Clustering): The three-dimensional feature matrix is ​​divided into superpixel blocks of size 50×50×50μm³, and the variance of the eddy current density within each superpixel is <5%.

[0107] For example, a TSV area (1mm×1mm×0.1mm) is divided into 800 superpixels, where the superpixel density variance in the defective area suddenly increases to 12%.

[0108] 2. Application of discrete Morse theory: Gradient flow field calculation: Calculate the eddy current density gradient (▽J) for each superpixel and mark the critical points where the gradient direction suddenly changes (such as saddle points and extreme points).

[0109] Topological skeleton generation: Connect critical points to form three-dimensional streamlines and construct an initial skeleton network.

[0110] Example: Dense saddle points (200 per cubic millimeter, compared to <50 in normal areas) were detected in the interface crack region.

[0111] 3. Persistent homology feature map: Vietoris-Rips complex construction: With the critical point as the center and the initial connection radius r = 10 μm, a simplicial complex is generated.

[0112] Continuous calculation: record the survival time of topological rings (such as holes) as the radius increases.

[0113] Output: A feature map containing 50 persistent topological rings (rings with a survival time > 100 μm are considered true defects).

[0114] Step 3.2: Dynamic topology reconstruction and noise filtering Dynamic topology reconstruction technology is used to filter noise from the persistent homology feature graph. By calculating the lifetime entropy of each topological ring, the connection radius threshold of the Vietoris-Rips complex is dynamically adjusted to eliminate the pseudo-topological structure caused by multiple scattering. The specific implementation is as follows: 1. Calculation of survival time entropy: Definition: H = -Σ(p_i log p_i), where p_i is the survival probability of a topological ring in the i-th radius interval.

[0115] Dynamic threshold adjustment: When the entropy value H>2.0, it is judged as a noise ring, and the connection radius is gradually increased (from 10μm→15μm→20μm) until H<1.5.

[0116] 2. Pseudo-structure elimination: Multiple scattering identification: Topological rings generated by scattering noise typically exhibit short lifetime (<50 μm) and high density distribution (>5 rings / 100 μm³).

[0117] Example: After filtering, the number of topological rings in a certain area dropped from 120 to 35, and the pseudo-ring elimination rate was 71%.

[0118] 3. Optimized topology skeleton: Key parameters: final connection radius r = 18 μm, entropy value H = 1.3.

[0119] Output: Contains persistent topological rings corresponding to real defects (e.g., interface crack ring survival time 250μm).

[0120] Step 3.3: Hodge decomposition and feature component extraction Based on the Hodge decomposition theorem, the filtered topological skeleton is decomposed into three orthogonal subspaces: the spiral gradient field component caused by structural deformation, the vortex circulation component caused by interface fracture, and the divergence anomaly component formed by thermal stress anomaly. The specific implementation is as follows: 1. Hodge decomposition implementation: Spiral gradient field: Extracts the irrotational component (▽×F=0) related to structural deformation and characterizes the overall warpage of the TSV.

[0121] Vortex circulation component: Calculate the non-zero part of the curl (▽·F=0), which corresponds to the closed vortex ring of the interface rupture.

[0122] Divergence anomaly component: The remaining part (▽·F≠0) reflects the material diffusion anomaly caused by thermal stress.

[0123] 2. Geometric invariance optimization: Seiberg-Witten equation constraints: curvature tensor invariance is maintained through gauge fixing conditions (such as the Lorenz gauge ▽·A=0).

[0124] Example: After optimization, the curvature radius change rate of a crack's vortex circulation component is less than 0.1%, satisfying the differential homeomorphism invariance.

[0125] 3. Feature component output: Structural deformation topological ring: contains 5 main spiral rings (average diameter 80μm).

[0126] Interface fracture chains: 3 closed vortex chains (length 150-300 μm).

[0127] Thermal stress anomaly surface: 2 divergence anomaly areas (volume 0.02 mm³).

[0128] Application scenario example: Advanced packaging TSV interface defect detection Test object: The through-silicon via (TSV) array of a 5G communication chip, with a diameter of 10μm and a depth of 80μm, showed interface delamination and clusters of nanopores after packaging.

[0129] Step 1: Superpixel segmentation and topology skeleton construction 1. Data input: 3D eddy current field characteristic matrix (size 2000×2000×200, resolution 5μm).

[0130] 2. Superpixel division: The SLIC algorithm generates 1000 superpixels, among which the eddy current density variance of superpixels numbered 487-492 reaches 15% (<5% in normal areas).

[0131] 3. Critical point detection: 83 saddle points are marked at the delamination interface, forming a dense streamline network.

[0132] 4. Initial topology skeleton: A Vietoris-Rips complex containing 120 topological rings was constructed (with an initial radius r = 10 μm).

[0133] Step 2: Dynamic topology reconstruction and denoising 1. Entropy analysis: The entropy value of the noise area is H=2.3 (superthreshold), and H=1.2 after gradually increasing r to 18μm.

[0134] 2. False ring filtering: 87 short survival rings (survival time < 60 μm) were eliminated and 33 real defect rings were retained.

[0135] 3. Optimized skeleton: Twelve long survival rings (survival time 180-300 μm) were retained in the interface delamination region.

[0136] Step 3: Hodge decomposition and defect quantification 1. Component decomposition: Spiral gradient field: Two spiral rings (120 μm in diameter) were detected, corresponding to the overall TSV warpage (warpage 3.5 μm).

[0137] Vortex circulation component: 3 closed chains (length 280 μm), precisely locating the interface delamination path.

[0138] Divergence anomaly: 1 thermal stress zone (volume 0.015 mm³), corresponding to the nanopore cluster.

[0139] 2. Geometric invariance verification: When scanning the same defect multiple times, the standard deviation of the curvature radius is <0.08μm, meeting the repeatability requirements.

[0140] Step 3.4: Geometric invariance optimization and feature component generation The geometric invariance of each subspace component is optimized. The evolution path of the curvature tensor is constrained by the Seiberg-Witten equation to generate a structural deformation topological ring, an interface fracture chain, and a thermal stress anomaly surface with differential homeomorphism invariance. Three characteristic components are obtained. The specific implementation is as follows: 1. Geometric invariance optimization goal Goal: Ensure that feature components remain stable under different coordinate system transformations (such as rotation, translation, and scaling) to avoid feature distortion caused by detection angle or device errors.

[0141] Diffeomorphism invariance: By constraining the evolution path of the curvature tensor, the geometric properties of the characteristic components remain unchanged under smooth coordinate transformations.

[0142] Application of the Seiberg-Witten equation: This equation is used to describe the spin structure on a manifold. Here it is used to control the canonical conditions of the curvature tensor and eliminate redundant degrees of freedom.

[0143] 2. Optimize processing flow Curvature tensor extraction: Structural deformation topological ring: Calculate the curvature tensor components (such as Gaussian curvature and mean curvature) of the spiral gradient field to characterize the degree of TSV warping.

[0144] Interface fracture chain: Extract the torsion tensor of the vortex circulation component to describe the torsional characteristics of the crack edge.

[0145] Thermal stress anomaly surface: Calculates the principal curvature direction of the divergence anomaly component, reflecting the deformation pattern caused by thermal expansion.

[0146] Seiberg-Witten equation constraints: Gauge condition setting: The Lorenz gauge (▽·A=0) is applied to each subspace component, where A is the gauge field of the curvature tensor to ensure the uniqueness of the solution to the equation.

[0147] Curvature evolution path optimization: By iteratively adjusting the covariant derivative of the curvature tensor to satisfy the equation constraints, artifacts introduced by the choice of coordinate system are eliminated.

[0148] Invariance verification: Coordinate system perturbation test: Random rotation (±10°) and translation (±50μm) are applied to the optimized characteristic components to verify the fluctuation range of curvature properties (such as Gaussian curvature change <0.5%).

[0149] Example: The torsion of a fracture chain at an interface remains 2.1 μm after coordinate rotation. -1 , standard deviation <0.02μm -1 .

[0150] Feature component generation: Structural deformation topological ring: After optimization, the standard deviation of the curvature radius of the spiral ring is reduced from 5μm to 0.3μm, indicating a stable warping mode.

[0151] Interface fracture chain: The torsion distribution uniformity of the closed vortex chain is improved (variance reduced by 60%), and the crack path is accurately located.

[0152] Abnormal thermal stress surface: The consistency between the main curvature direction and the thermal load direction is >95%, and the source of thermal stress is clear.

[0153] Output: Three characteristic components with differential homeomorphism invariance, which can be directly used for defect comparison analysis across devices and batches.

[0154] S203, performing cross-domain fusion processing of the three feature components using a dynamic heterogeneous neural network, and generating a defect topology fingerprint map through a dual-path attention gating mechanism, wherein the cross-domain fusion processing adopts a hybrid architecture of a spatiotemporal causal convolutional network and a graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy; The three feature components are subjected to cross-domain fusion processing using a dynamic heterogeneous neural network, and a defect topology fingerprint is generated through a dual-path attention gating mechanism. This cross-domain fusion process utilizes a hybrid architecture combining a spatiotemporal causal convolutional network and a graph neural network. This hybrid neural network architecture enables deep fusion of multi-source defect features, while the dual-path attention mechanism highlights key defect characteristics. This method effectively correlates the cross-scale characteristics of nanopore defects, improving the accuracy of defect characterization.

[0155] Specifically, the structural deformation topological ring can be input into the spatiotemporal causal convolutional network, and the deformation evolution characteristics in the time dimension can be extracted through the dilated convolution kernel with memory gating to generate a feature tensor containing multi-scale spatiotemporal correlations; the spatiotemporal convolutional network effectively captures the dynamic evolution characteristics of defects, and the memory gating mechanism enhances the modeling ability of long-term dependencies.

[0156] The interface fracture chain is input into a graph neural network, a dynamic adjacency matrix is ​​constructed based on the nanoscale spatial coordinates of the fracture nodes, and an edge weight propagation algorithm is used to generate graph features that characterize the topological connectivity of pore defects. The graph neural network accurately models the spatial correlation characteristics of nanopores, and the dynamic adjacency matrix adapts to defect structures of different scales.

[0157] The thermal stress anomaly surface is subjected to three-dimensional curvature parameterization processing, and the surface gradient distribution is matched by a deformable convolution kernel to extract the characteristic vector of the thermal stress anomaly area; the curvature parameterization enhances the expressive ability of the surface features, and the deformable convolution improves the modeling accuracy of complex stress distribution.

[0158] An asymmetric weight sharing strategy is implemented to perform cross-network feature fusion and generate enhanced features that integrate cross-physical field characteristics; asymmetric sharing enables efficient feature fusion, and cross-physical field integration provides a more comprehensive defect characterization.

[0159] These enhanced features are fed into a dual-path attention gating mechanism. The first path uses channel-weighted attention to filter key frequency band features, while the second path uses a spatial topological focusing mechanism to enhance the response to defect regions. The final output is a defect topological fingerprint with cross-scale correlation. This dual-path attention mechanism highlights key defect features, and the topological fingerprint achieves precise defect characterization.

[0160] Step 4.1: Spatiotemporal Causal Convolutional Network Processing Structural Deformation Topological Ring 1. Network architecture design: Dilated Causal Convolution: A convolution kernel (size 3×3) with a dilation rate of d=2 is used to ensure causality in the time dimension (relying only on historical data).

[0161] Memory gating mechanism: LSTM units are introduced as gates to dynamically adjust convolution kernel weights and capture long-term dependencies of deformations.

[0162] Example: Input a structural deformation topological ring (time series length 100 frames, spatial resolution 10μm), and output a feature tensor (size 50×50×256).

[0163] 2. Multi-scale feature extraction: First layer: expansion rate d=1, extracting short-term deformation fluctuations (such as microsecond vibrations).

[0164] Second layer: expansion rate d=4, capturing minute-level trends (such as slow warping caused by thermal stress accumulation).

[0165] The third layer: expansion rate d=8, which perceives hourly evolution (such as the package aging process).

[0166] 3. Feature Fusion Feature maps of different scales are fused through skip connections.

[0167] Use 1×1 convolution to compress the number of channels and generate the final feature tensor (size 50×50×128).

[0168] Step 4.2: Graph Neural Network Processing Interface Broken Chain 1. Graph structure construction: Node definition: Each fracture endpoint is used as a graph node (coordinate accuracy ±1nm), and its attributes include curvature, torsion, and stress concentration factor.

[0169] Edge connection rule: If the distance between two nodes is less than 50 nm and the stress field is continuous, an edge connection is established, and the initial weight is the inverse of the distance.

[0170] 2. Graph Convolutional Network (GCN) Design: Message passing mechanism: Node feature update formula: : Features of node i in layer l N(i): The neighbor set of node i d_i : degree of node i (number of connected edges) : Learnable weight matrix σ : ReLU activation function Example: Input 500 nodes and generate 128-dimensional graph features through 3 layers of GCN.

[0171] 3. Dynamic adjacency matrix update: After each iteration, new potential connections are added based on the similarity of node features (cosine similarity > 0.8).

[0172] Weak connections with weight < 0.1 are removed to ensure the sparsity and physical meaning of the graph structure.

[0173] Step 4.3: Thermal stress anomaly surface feature extraction 1. Curvature parameterization: Principal curvature calculation: Calculate the maximum / minimum curvature of each point through surface fitting algorithm .

[0174] Shape index encoding: in, , -1 represents a bowl-shaped depression, and +1 represents a peak protrusion.

[0175] 2. Deformable convolution operation: The initial position of the convolution kernel is dynamically offset according to the curvature direction (the offset Δx, Δy, Δz is determined by the curvature gradient).

[0176] Example: A 3×3×3 deformable convolution kernel automatically focuses on the area with the largest deformation in a region of concentrated thermal stress (high curvature).

[0177] 3. Feature Dimensionality Reduction: Principal component analysis (PCA) is used to compress the 512-dimensional surface features to 64 dimensions, retaining 95% variance.

[0178] Step 4.4: Asymmetric weight sharing fusion 1. Cross-domain connection layer design: The spatiotemporal convolutional network (128 dimensions) and the graph network (128 dimensions) share the first 64 dimensions of the weight matrix.

[0179] Asymmetric activation function: Convolutional network side: Swish function (enhanced nonlinearity) Network side: LeakyReLU (α=0.2, to prevent gradient disappearance) 2. Geometric similarity matrix calculation: S=ξ(F_t^T·W_2·F_g) in, : spatiotemporal feature tensor; : Graph features; : Shared weight matrix; ξ is the Sigmoid function, outputting a similarity score [0,1].

[0180] 3. Feature Fusion Perform Hadamard product (element-wise multiplication) on the similarity matrix S and the thermal stress eigenvector to filter cross-domain related features.

[0181] The fully connected layers are mapped to a 256-dimensional enhanced feature space.

[0182] Step 4.5: Generate fingerprints using dual-path attention gating 1. Channel Attention: Calculate the global average pooling value of each feature channel and generate weights through two layers of MLP.

[0183] Example: Filter out 32 key channels (weight > 0.7) from 256-dimensional features.

[0184] 2. Spatial Attention: Construct a heat map of defect areas and apply a 3x weight to high-response areas (heat value > threshold).

[0185] The dynamic mask is multiplied with the feature map point by point to highlight key areas such as cracks and pores.

[0186] 3. Fingerprint generation: The weighted feature map is reconstructed in three dimensions (resolution 1 μm) to mark cross-scale correlation regions.

[0187] The output is a 256×256×256 topological fingerprint map, which contains the defect type, location, and correlation strength encoding.

[0188] Specifically, implementing an asymmetric weight sharing strategy to fuse cross-network features and generate enhanced features that fuse cross-physical field characteristics can include: A cross-domain connection layer is established between the spatiotemporal causal convolutional network and the graph neural network. The high-dimensional feature channels of the convolutional network and the node embedding layer of the graph neural network share a weight matrix, each using different nonlinear activation functions to achieve asymmetric mapping. A cross-domain connection layer is established between the spatiotemporal causal convolutional network and the graph neural network to achieve asymmetric weight sharing. This cross-domain connection promotes feature interaction between different networks, while the asymmetric design preserves the unique advantages of each network.

[0189] The feature tensor is concatenated with the graph features, and the geometric similarity matrix S of the nanopore defects is calculated using a cross-scale association module. Here, S = ξ(F_t^T·W_2·F_g), where F_t is the feature tensor, T is the matrix transpose, F_g is the graph feature, ξ is the sigmoid function, and W_2 is the learnable shared weight. The feature tensor is concatenated with the graph features, and the geometric similarity matrix is ​​calculated using a cross-scale association module. This tensor concatenation achieves deep feature fusion, and the similarity matrix quantifies cross-scale associations.

[0190] Performing a Hadamard product on the geometric similarity matrix and the thermal stress eigenvectors generates enhanced features that incorporate cross-physics characteristics. Performing a Hadamard product on the geometric similarity matrix and the thermal stress eigenvectors generates enhanced features. The Hadamard product enables selective feature enhancement, increasing the salience of key defect features.

[0191] Step 6.1: Cross-domain connection layer construction and asymmetric mapping 1. Cross-domain connection layer design: Weight matrix sharing mechanism: define shared weight matrix , where d_c=256 is the output feature dimension of the spatiotemporal convolutional network, and d_g=128 is the node embedding dimension of the graph neural network.

[0192] Asymmetric activation function: LeakyReLU (negative slope 0.2) is used on the spatiotemporal convolution side: it retains negative features and captures the inhibitory effect in deformation evolution.

[0193] The Swish function is used on the graph network side to enhance nonlinear expression capabilities and adapt to the complex correlation of pore topology.

[0194] Parameter initialization: The weight matrix W_2 is initialized using the Xavier normal distribution and the bias term is set to 0.

[0195] 2. Feature channel alignment: Dimension matching: Spatiotemporal convolution features are transformed through the fully connected layer Map to , and graph features Dimension alignment (N=1000, M=500).

[0196] For example, in a TSV detection task, the spatiotemporal features contain 1000 nodes (N=1000), and the graph features contain 500 pore nodes (M=500). After alignment by the fully connected layer, both are mapped to 128 dimensions.

[0197] 3. Asymmetric mapping verification: Activation distribution analysis: After LeakyReLU, the negative value retention rate of spatiotemporal features is 15%; after Swish, the nonlinear response strength of image features is increased by 30%.

[0198] Gradient flow monitoring: Gradient clipping (threshold 1.0) is used to prevent gradient explosion during cross-domain training.

[0199] Step 6.2: Geometric similarity matrix calculation 1. Tensor splicing and interactive computing: Splicing method: Splice F_t and F_g along the feature dimension to generate a joint feature matrix .

[0200] Similarity measure: Calculate the geometric correlation between F_t and F_g. The formula is expanded as follows:

[0201] Among them, i∈[1,N], j∈[1,M], represents the association strength between spatiotemporal nodes and graph nodes.

[0202] 2. Learnable parameter optimization: Loss function design: Using contrastive loss to maximize the similarity of similar defects ( Sij →1), minimizing heterogeneous similarity ( Sij →0).

[0203] Backpropagation strategy: use Adam optimizer (learning rate 1e-4), batch size 32, and 1000 iterations.

[0204] 3. Matrix normalization and screening: Threshold processing: setting similarity threshold θ =0.7, retain Sij ≥ θ ,filter out noise interference.

[0205] For example, in the detection of a nanopore cluster, a 500×500 similarity matrix was generated. After threshold screening, the number of valid correlation pairs was reduced from 120,000 to 15,000, and the signal-to-noise ratio was improved by 8 times.

[0206] Step 6.3: Cross-physics feature fusion 1. Thermal stress feature extraction: Curvature parameterization: Calculating the principal curvatures of thermally anomaly surfaces κ 1, κ 2. Generate feature vectors .

[0207] Normalization: Min-Max normalization is used to map the curvature value to the [0,1] interval.

[0208] 2. Hadamard Implementation: Element-wise multiplication: For each associated pair (i, j), calculate the enhanced features ,in is the thermal stress curvature corresponding to the jth graph node.

[0209] Dimension expansion: Reshape into , input to the subsequent network.

[0210] 3. Fusion feature optimization: Feature Screening: Preserving Curvature Products Eij Strong correlation features with a value of ≥ 0.5 were selected, and weak responses were eliminated.

[0211] For example, in an interface crack detection case, the number of fused features decreased from 15,000 to 8,000, but the proportion of effective features increased from 40% to 85%.

[0212] S204: Input the defect topology fingerprint map into the multi-scale entropy evaluation module to simulate the defect evolution, solve the nonlinear mapping relationship between the defect extension path and the electrothermal performance degradation based on the simulated annealing algorithm, and output a quantitative detection report including defect geometric parameters, failure probability and reliability threshold. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy model, and generates a feedback control signal to optimize the scanning parameter configuration.

[0213] The defect topology fingerprint map is fed into the multi-scale entropy evaluation module to simulate defect evolution. Using a simulated annealing algorithm, the nonlinear mapping relationship between defect propagation paths and electrothermal performance degradation is solved, and a quantitative inspection report is generated. Multi-scale entropy evaluation enables accurate prediction of defect evolution, while the simulated annealing algorithm optimizes the solution of the nonlinear mapping relationship. This quantitative inspection report provides a scientific basis for TSV package reliability assessment.

[0214] Specifically, the defect topology fingerprint map can be input into a multi-scale entropy evaluation module, and an atomic-mesoscopic-macroscopic three-level entropy calculation model can be constructed based on the defect geometric parameters. At the atomic scale, the atomic displacement entropy of the interface defect area is calculated according to the molecular dynamics simulation results; at the mesoscopic scale, the morphological entropy change of the crack propagation path is simulated based on the phase field method; at the macroscopic scale, a joint entropy function of electromagnetic conductivity, thermal resistivity and mechanical strength degradation is established; multi-scale modeling realizes a comprehensive evaluation of defect evolution, and the three-level entropy calculation covers failure mechanisms at different scales.

[0215] Based on the three-level entropy calculation model, the Hamiltonian of the simulated annealing algorithm is constructed, the defect expansion path is mapped into a spin state combination of the Ising model, and the nonlinear mapping relationship between the degradation of electrothermal performance and the change in entropy value is solved through the quantum tunneling effect; the physics-inspired algorithm design improves the solution efficiency, and the Ising model mapping realizes the effective representation of complex relationships.

[0216] Based on the Monte Carlo method, a set of defect evolution paths under different service conditions is generated. The weight parameters of the entropy calculation model are dynamically corrected by comparing the KL divergence of the measured entropy change value with the simulated value. The Monte Carlo simulation provides a reliable statistical evaluation, and the dynamic correction ensures the continuous optimization of the model.

[0217] The modified entropy calculation model generates an optimized parameter combination for the polarization field intensity ratio and the thermal load application period, which is then used as a feedback control signal to return the scanning parameter configuration. This closed-loop optimization mechanism enables the detection system to self-improve and improve the accuracy of long-term detection.

[0218] Step 5.1: Multi-scale entropy model construction 1. Calculation of atomic-scale entropy: Molecular dynamics simulation: LAMMPS software was used to simulate the atomic motion in the defect region with a time step of 1 femtosecond (fs) and a simulation duration of 100 picoseconds (ps).

[0219] Atomic displacement entropy: The entropy value S_atom = k_B ln(Ω) is calculated by the mean square displacement (MSD), where Ω is the number of possible microstates of the atomic position and k_B is the Boltzmann constant.

[0220] Example: For a nanopore area (2nm diameter), S_atom=3.2×10 -23 J / K, significantly higher than the normal area (1.8×10 -23 J / K).

[0221] 2. Mesoscopic morphological entropy change: Phase field modeling: define the phase field variable φ (0≤φ≤1, φ=1 indicates intact material, φ=0 indicates crack), the control equation φ / t = -M(δF / δφ), where F is the free energy functional and M is the mobility.

[0222] Morphological entropy calculation: S_meso = -∫[φ ln φ + (1-φ) ln (1-φ)] dx, where the integration area is the crack propagation path.

[0223] Example: The S_meso of a crack growth path increases from 0.5 to 1.2, indicating an increase in disorder.

[0224] 3. Macro-joint entropy function: Electromagnetic conductivity entropy: S_elec = -σlnσ, where σ is the attenuation ratio of conductivity relative to the nominal value (e.g., σ=0.85 means a 15% attenuation).

[0225] Thermal resistance entropy: S_therm =κln(T_grad / T_ref), κ is the thermal resistance coefficient, T_grad is the measured temperature gradient, and T_ref is the reference value.

[0226] Mechanical strength entropy: S_mech = E ln(ε / ε_yield), where E is the elastic modulus, ε is the strain, and ε_yield is the yield strain.

[0227] Joint entropy: S_total =αS_elec +βS_therm +γS_mech, weight α=0.4, β=0.3, γ=0.3.

[0228] Step 5.2: Simulated annealing algorithm solution 1. Ising model mapping: Spin state definition: Each lattice point corresponds to the defect expansion direction (↑: expansion, ↓: stagnation), and the lattice size matches the TSV structure (e.g., 100×100 grid).

[0229] Hamiltonian construction: H = -JΣσ_iσ_j - hΣσ_i, where J is the coupling constant (reflecting the toughness of the material) and h is the external field (reflecting the thermal stress).

[0230] 2. Quantum tunneling optimization: Temperature scheduling: initial temperature T=1000K, exponential cooling T_{n+1}=0.95T_n, minimum temperature T_min=1K.

[0231] Tunneling probability: P_tunnel = exp(-ΔE / (k_B T)), where ΔE is the energy difference, allowing crossing of local minima.

[0232] 3. Path solving: 1000 candidate paths are generated iteratively, and the path with the smallest entropy increase (ΔS_total < 5%) and the lowest energy is selected as the optimal solution.

[0233] Example: The entropy of the optimal path of a crack increases by 2.3%, corresponding to a 12% decrease in electrical conductivity and an 8% increase in thermal resistance.

[0234] Step 5.3: Monte Carlo evolution and error correction 1. Service Condition Modeling: Temperature cycle: -40°C 125°C, 1000 cycles.

[0235] Mechanical vibration: frequency 20-2000Hz, acceleration 5g.

[0236] Humidity shock: 85°C / 85%RH for 500 hours.

[0237] 2. KL divergence calculation: Definition: D_{KL}(P||Q) = ΣP(i) ln(P(i) / Q(i)), where P is the measured entropy change distribution and Q is the simulated distribution.

[0238] Threshold setting: Parameter correction is triggered when D_{KL}>0.1.

[0239] 3. Dynamic weight adjustment: Use gradient descent to update weights: α_{new} = α - η( D_{KL} / α), η is the learning rate (0.01).

[0240] Example: Initial α=0.4 → After correction α=0.37, D_{KL} is reduced from 0.15 to 0.07.

[0241] Step 5.4: Feedback Control Signal Generation The optimized parameter combination of polarization field intensity ratio and heat load application period is generated based on the modified entropy value calculation model, and the feedback control signal is formed to return the scanning parameter configuration. The specific implementation is as follows: 1. Parameter optimization modeling: Objective function: Min(ΔS_total) + λ(scan time), where λ is the trade-off coefficient (λ=0.5).

[0242] Constraints: polarization field intensity ratio 1.2≤E / H≤2.0, thermal load period 10ms≤T≤50ms.

[0243] 2. Multi-objective optimization: The Pareto front was generated using the NSGA-II algorithm, and solutions with ΔS_total < 3% and scan time < 15 min were selected.

[0244] Example: The optimal parameter combination is E / H=1.8, T=25ms.

[0245] 3. Signal output: Generate JSON format control instructions, for example: json { "polarization_ratio": 1.8, "thermal_cycle": 25, "scan_mode": "high_resolution" }.

[0246] It can be seen that the eddy current pulse thermal imaging device is used to perform multi-dimensional excitation scanning on the TSV packaging structure to generate a three-dimensional eddy current field distribution feature matrix; the three-dimensional eddy current field distribution feature matrix is ​​input into the topological manifold decomposition module for defect feature separation, and three feature components of structural deformation topological ring, interface fracture chain and thermal stress abnormal surface are constructed based on the continuous homology analysis method; the three feature components are cross-domain fused by dynamic heterogeneous neural network, and a defect topology fingerprint map is generated through a dual-path attention gating mechanism; the defect topology fingerprint map is input into the multi-scale entropy value evaluation module for defect evolution simulation, and a quantitative detection report containing defect geometric parameters, failure probability and reliability threshold is output, so that a comprehensive defect detection system can be built to improve the accuracy and efficiency of defect identification.

[0247] Another embodiment of the present invention provides a TSV package defect detection system. Figure 3 , the system may include: Scanning module 301 is used to perform multi-dimensional excitation scanning on the TSV packaging structure using an eddy current pulse thermal imaging device, perform multi-physics field coupling demodulation processing on the original electromagnetic echo signal based on an adaptive phase encoding algorithm, and generate a three-dimensional eddy current field distribution feature matrix, wherein the multi-dimensional excitation scanning achieves cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads; A construction module 302 is configured to input the three-dimensional eddy current field distribution characteristic matrix into a topological manifold decomposition module for defect feature separation, and construct three characteristic components: a structural deformation topological ring, an interface fracture chain, and a thermal stress anomaly surface based on a persistent homology analysis method. The topological manifold decomposition module uses a dynamic topological reconstruction technique to eliminate multiple scattering noise and extract geometric invariance features of submicron defects. Fusion module 303 is used to perform cross-domain fusion processing of the three feature components using a dynamic heterogeneous neural network, and generate a defect topology fingerprint through a dual-path attention gating mechanism. The cross-domain fusion processing adopts a hybrid architecture of a spatiotemporal causal convolutional network and a graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy; The output module 304 is used to input the defect topology fingerprint map into the multi-scale entropy value evaluation module to simulate the defect evolution, solve the nonlinear mapping relationship between the defect extension path and the electrothermal performance degradation based on the simulated annealing algorithm, and output a quantitative detection report including defect geometric parameters, failure probability and reliability threshold. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy value model and generates a feedback control signal to optimize the scanning parameter configuration.

[0248] It can be seen that the eddy current pulse thermal imaging device is used to perform multi-dimensional excitation scanning on the TSV packaging structure to generate a three-dimensional eddy current field distribution feature matrix; the three-dimensional eddy current field distribution feature matrix is ​​input into the topological manifold decomposition module for defect feature separation, and three feature components of structural deformation topological ring, interface fracture chain and thermal stress abnormal surface are constructed based on the continuous homology analysis method; the three feature components are cross-domain fused by dynamic heterogeneous neural network, and a defect topology fingerprint map is generated through a dual-path attention gating mechanism; the defect topology fingerprint map is input into the multi-scale entropy value evaluation module for defect evolution simulation, and a quantitative detection report containing defect geometric parameters, failure probability and reliability threshold is output, so that a comprehensive defect detection system can be built to improve the accuracy and efficiency of defect identification.

[0249] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0250] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps: S201, performing multi-dimensional excitation scanning on the TSV packaging structure using an eddy current pulse thermal imaging device, performing multi-physics field coupling demodulation processing on the original electromagnetic echo signal based on an adaptive phase encoding algorithm, and generating a three-dimensional eddy current field distribution feature matrix, wherein the multi-dimensional excitation scanning achieves cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads; S202: Inputting the three-dimensional eddy current field distribution feature matrix into a topological manifold decomposition module to separate defect features, and constructing three feature components: structural deformation topological rings, interface fracture chains, and thermal stress anomaly surfaces based on a continuous homology analysis method. The topological manifold decomposition module uses dynamic topological reconstruction technology to eliminate multiple scattering noise and extract geometric invariance features of submicron defects. S203, performing cross-domain fusion processing of the three feature components using a dynamic heterogeneous neural network, and generating a defect topology fingerprint map through a dual-path attention gating mechanism, wherein the cross-domain fusion processing adopts a hybrid architecture of a spatiotemporal causal convolutional network and a graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy; S204: Input the defect topology fingerprint map into the multi-scale entropy evaluation module to simulate the defect evolution, solve the nonlinear mapping relationship between the defect extension path and the electrothermal performance degradation based on the simulated annealing algorithm, and output a quantitative detection report including defect geometric parameters, failure probability and reliability threshold. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy model, and generates a feedback control signal to optimize the scanning parameter configuration.

[0251] It can be seen that the eddy current pulse thermal imaging device is used to perform multi-dimensional excitation scanning on the TSV packaging structure to generate a three-dimensional eddy current field distribution feature matrix; the three-dimensional eddy current field distribution feature matrix is ​​input into the topological manifold decomposition module for defect feature separation, and three feature components of structural deformation topological ring, interface fracture chain and thermal stress abnormal surface are constructed based on the continuous homology analysis method; the three feature components are cross-domain fused by dynamic heterogeneous neural network, and a defect topology fingerprint map is generated through a dual-path attention gating mechanism; the defect topology fingerprint map is input into the multi-scale entropy value evaluation module for defect evolution simulation, and a quantitative detection report containing defect geometric parameters, failure probability and reliability threshold is output, so that a comprehensive defect detection system can be built to improve the accuracy and efficiency of defect identification.

[0252] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0253] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0254] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program: S201, performing multi-dimensional excitation scanning on the TSV packaging structure using an eddy current pulse thermal imaging device, performing multi-physics field coupling demodulation processing on the original electromagnetic echo signal based on an adaptive phase encoding algorithm, and generating a three-dimensional eddy current field distribution feature matrix, wherein the multi-dimensional excitation scanning achieves cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads; S202: Inputting the three-dimensional eddy current field distribution feature matrix into a topological manifold decomposition module to separate defect features, and constructing three feature components: structural deformation topological rings, interface fracture chains, and thermal stress anomaly surfaces based on a continuous homology analysis method. The topological manifold decomposition module uses dynamic topological reconstruction technology to eliminate multiple scattering noise and extract geometric invariance features of submicron defects. S203, performing cross-domain fusion processing of the three feature components using a dynamic heterogeneous neural network, and generating a defect topology fingerprint map through a dual-path attention gating mechanism, wherein the cross-domain fusion processing adopts a hybrid architecture of a spatiotemporal causal convolutional network and a graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy; S204: Input the defect topology fingerprint map into the multi-scale entropy evaluation module to simulate the defect evolution, solve the nonlinear mapping relationship between the defect extension path and the electrothermal performance degradation based on the simulated annealing algorithm, and output a quantitative detection report including defect geometric parameters, failure probability and reliability threshold. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy model, and generates a feedback control signal to optimize the scanning parameter configuration.

[0255] It can be seen that the eddy current pulse thermal imaging device is used to perform multi-dimensional excitation scanning on the TSV packaging structure to generate a three-dimensional eddy current field distribution feature matrix; the three-dimensional eddy current field distribution feature matrix is ​​input into the topological manifold decomposition module for defect feature separation, and three feature components of structural deformation topological ring, interface fracture chain and thermal stress abnormal surface are constructed based on the continuous homology analysis method; the three feature components are cross-domain fused by dynamic heterogeneous neural network, and a defect topology fingerprint map is generated through a dual-path attention gating mechanism; the defect topology fingerprint map is input into the multi-scale entropy value evaluation module for defect evolution simulation, and a quantitative detection report containing defect geometric parameters, failure probability and reliability threshold is output, so that a comprehensive defect detection system can be built to improve the accuracy and efficiency of defect identification.

[0256] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for detecting TSV packaging defects, characterized in that: The method comprises: An eddy current pulse thermal imaging device is used to perform multi-dimensional excitation scanning on the TSV packaging structure. Based on an adaptive phase encoding algorithm, the original electromagnetic echo signal is subjected to multi-physics field coupling demodulation processing to generate a three-dimensional eddy current field distribution feature matrix. The multi-dimensional excitation scanning achieves cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads. The three-dimensional eddy current field distribution feature matrix is ​​input into the topological manifold decomposition module for defect feature separation. Based on the continuous homology analysis method, three feature components are constructed: structural deformation topological ring, interface fracture chain, and thermal stress abnormal surface. The topological manifold decomposition module uses dynamic topological reconstruction technology to eliminate multiple scattering noise and extract the geometric invariance characteristics of submicron defects. The three feature components are subjected to cross-domain fusion processing using a dynamic heterogeneous neural network, and a defect topology fingerprint is generated through a dual-path attention gating mechanism. The cross-domain fusion processing adopts a hybrid architecture of spatiotemporal causal convolutional network and graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy; The defect topology fingerprint map is input into a multi-scale entropy evaluation module to simulate defect evolution. The nonlinear mapping relationship between the defect extension path and the degradation of electrothermal performance is solved based on the simulated annealing algorithm, and a quantitative detection report containing defect geometric parameters, failure probability and reliability threshold is output. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy model and generates a feedback control signal to optimize the scanning parameter configuration.

2. The method according to claim 1, characterized in that The method uses an eddy current pulse thermal imaging device to perform multi-dimensional excitation scanning on the TSV packaging structure, performs multi-physics field coupling demodulation processing on the original electromagnetic echo signal based on an adaptive phase encoding algorithm, and generates a three-dimensional eddy current field distribution feature matrix. The multi-dimensional excitation scanning realizes cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads, including: Based on the dielectric constant distribution of the TSV packaging structure, a frequency combination sequence of orthogonal polarization electromagnetic fields is dynamically generated. The transverse electric field component and the longitudinal magnetic field component are alternately applied through the electromagnetic field emission array to stimulate the coupled response of surface eddy currents and deep thermal effects, and obtain the original electromagnetic echo signal. Performing dual-domain synchronous sampling processing on the original electromagnetic echo signal, using nonlinear pulse compression technology in the time domain to eliminate multipath interference, and extracting phase mutation characteristics of non-stationary signals through fractional Fourier transform in the frequency domain to generate a time-frequency joint distribution map; Based on the time-frequency joint distribution map, a multi-physics field coupling equation is constructed, and the alternating direction multiplier method ADMM is used to jointly solve the electromagnetic-heat conduction equation to obtain the vector distribution data of the three-dimensional eddy current density field and the temperature gradient field; The vector distribution data is input into a phase encoding matrix generator, and a chaotic sequence is used to dynamically encrypt and map the spatial coordinates to generate a three-dimensional eddy current field distribution characteristic matrix with spatial continuity and anti-aliasing characteristics.

3. The method according to claim 2, characterized in that The three-dimensional eddy current field distribution feature matrix is ​​input into the topological manifold decomposition module to separate defect features, and three feature components, namely, structural deformation topological ring, interface fracture chain, and thermal stress abnormal surface, are constructed based on the continuous homology analysis method. The topological manifold decomposition module uses dynamic topological reconstruction technology to eliminate multiple scattering noise and extract the geometric invariance characteristics of submicron defects, including: Performing super-pixel region division on the three-dimensional eddy current field distribution feature matrix, constructing a critical point topological skeleton of the three-dimensional gradient flow using discrete Morse theory, and generating an initial persistent homology feature map; Dynamic topology reconstruction technology is used to filter noise from the persistent homology feature graph. By calculating the lifetime entropy of each topological ring, the connection radius threshold of the Vietoris-Rips complex is dynamically adjusted to eliminate pseudo-topological structures caused by multiple scattering. Based on the Hodge decomposition theorem, the filtered topological skeleton is decomposed into three orthogonal subspaces: the spiral gradient field component caused by structural deformation, the vortex circulation component caused by interface fracture, and the divergence anomaly component formed by thermal stress anomaly. The geometric invariance of each subspace component is optimized, and the evolution path of the curvature tensor is constrained by the Seiberg-Witten equation to generate structural deformation topological rings, interface fracture chains and thermal stress anomaly surfaces with differential homeomorphism invariance, and three characteristic components are obtained.

4. The method according to claim 3, characterized in that The three feature components are subjected to cross-domain fusion processing of a dynamic heterogeneous neural network, and a defect topology fingerprint is generated through a dual-path attention gating mechanism. The cross-domain fusion processing adopts a hybrid architecture of a spatiotemporal causal convolutional network and a graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy, including: The structural deformation topological ring is input into the spatiotemporal causal convolutional network, and the deformation evolution characteristics in the time dimension are extracted through the dilated convolution kernel with memory gating to generate a feature tensor containing multi-scale spatiotemporal correlations; The interface fracture chain is input into a graph neural network, a dynamic adjacency matrix is ​​constructed based on the nanoscale spatial coordinates of the fracture nodes, and a graph feature representing the topological connectivity of the pore defects is generated using an edge weight propagation algorithm; Performing three-dimensional curvature parameterization processing on the thermal stress anomaly surface, matching the surface gradient distribution through a deformable convolution kernel, and extracting the feature vector of the thermal stress anomaly area; Implement an asymmetric weight sharing strategy for cross-network feature fusion to generate enhanced features that integrate cross-physics field characteristics; The enhanced features are input into a dual-path attention gating mechanism. The first path uses channel-weighted attention to screen key frequency band features, and the second path strengthens the defect area response through a spatial topological focusing mechanism, and finally outputs a defect topology fingerprint map with cross-scale correlation.

5. The method according to claim 4, characterized in that The defect topology fingerprint map is input into the multi-scale entropy value evaluation module to perform defect evolution simulation, and the nonlinear mapping relationship between the defect extension path and the electrothermal performance degradation is solved based on the simulated annealing algorithm. A quantitative detection report including defect geometric parameters, failure probability and reliability threshold is output. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy value model and generates a feedback control signal to optimize the scanning parameter configuration, including: The defect topological fingerprint map is input into a multi-scale entropy value evaluation module, and an atomic-mesoscopic-macroscopic three-level entropy value calculation model is constructed based on the defect geometric parameters. At the atomic scale, the atomic displacement entropy of the interface defect area is calculated based on the molecular dynamics simulation results; at the mesoscale, the morphological entropy change of the crack propagation path is simulated based on the phase field method; at the macroscale, a joint entropy function of electromagnetic conductivity, thermal resistivity and mechanical strength degradation is established; The Hamiltonian of the simulated annealing algorithm is constructed based on the three-level entropy calculation model, the defect expansion path is mapped to the spin state combination of the Ising model, and the nonlinear mapping relationship between electrothermal performance degradation and entropy value change is solved through the quantum tunneling effect; Based on the Monte Carlo method, a set of defect evolution paths under different service conditions is generated. The weight parameters of the entropy calculation model are dynamically modified by comparing the KL divergence of the measured entropy change value with the simulated value. The optimized parameter combination of polarization field intensity ratio and heat load application period is generated according to the modified entropy value calculation model, and a feedback control signal is formed to return the scanning parameter configuration.

6. The method according to claim 4, characterized in that The asymmetric weight sharing strategy is implemented to perform cross-network feature fusion to generate enhanced features that fuse cross-physical field characteristics, including: A cross-domain connection layer is established between the spatiotemporal causal convolutional network and the graph neural network. The high-dimensional feature channels of the convolutional network and the node embedding layer of the graph neural network share a weight matrix, and asymmetric mapping is achieved through different nonlinear activation functions. The feature tensor is concatenated with the graph features, and the geometric similarity matrix S of the nanopore defects is calculated through the cross-scale association module, where S = ξ(F_t^T·W_2·F_g), where F_t is the feature tensor, T is the matrix transpose, F_g is the graph feature, ξ is the Sigmoid function, and W_2 is the learnable shared weight; A Hadamard product is performed on the geometric similarity matrix and the thermal stress eigenvectors to generate enhanced features that incorporate cross-physics properties.

7. A TSV package defect detection system, characterized in that: The system comprises: A scanning module is used to perform multi-dimensional excitation scanning on the TSV packaging structure using an eddy current pulse thermal imaging device, perform multi-physics field coupling demodulation processing on the original electromagnetic echo signal based on an adaptive phase encoding algorithm, and generate a three-dimensional eddy current field distribution feature matrix, wherein the multi-dimensional excitation scanning achieves cross-modal excitation of interface defects by alternately applying orthogonal polarization electromagnetic fields and gradient thermal loads; A construction module is used to input the three-dimensional eddy current field distribution feature matrix into a topological manifold decomposition module to separate defect features, and construct three feature components: structural deformation topological rings, interface fracture chains, and thermal stress anomaly surfaces based on a continuous homology analysis method. The topological manifold decomposition module uses dynamic topological reconstruction technology to eliminate multiple scattering noise and extract geometric invariance features of submicron defects; A fusion module for performing cross-domain fusion processing of the three feature components using a dynamic heterogeneous neural network to generate a defect topology fingerprint through a dual-path attention gating mechanism. The cross-domain fusion processing adopts a hybrid architecture of a spatiotemporal causal convolutional network and a graph neural network, and realizes cross-scale feature association of nanopore defects through an asymmetric weight sharing strategy; The output module is used to input the defect topology fingerprint map into the multi-scale entropy value evaluation module to simulate the defect evolution, solve the nonlinear mapping relationship between the defect extension path and the electrothermal performance degradation based on the simulated annealing algorithm, and output a quantitative detection report containing defect geometric parameters, failure probability and reliability threshold. The evaluation module dynamically corrects the detection error by constructing an electromagnetic-thermal-mechanical entropy value model and generates a feedback control signal to optimize the scanning parameter configuration.

8. The system according to claim 7, characterized in that The scanning module is specifically used for: Based on the dielectric constant distribution of the TSV packaging structure, a frequency combination sequence of orthogonal polarization electromagnetic fields is dynamically generated. The transverse electric field component and the longitudinal magnetic field component are alternately applied through the electromagnetic field emission array to stimulate the coupled response of surface eddy currents and deep thermal effects, and obtain the original electromagnetic echo signal. Performing dual-domain synchronous sampling processing on the original electromagnetic echo signal, using nonlinear pulse compression technology in the time domain to eliminate multipath interference, and extracting phase mutation characteristics of non-stationary signals through fractional Fourier transform in the frequency domain to generate a time-frequency joint distribution map; Based on the time-frequency joint distribution map, a multi-physics field coupling equation is constructed, and the alternating direction multiplier method ADMM is used to jointly solve the electromagnetic-heat conduction equation to obtain the vector distribution data of the three-dimensional eddy current density field and the temperature gradient field; The vector distribution data is input into a phase encoding matrix generator, and a chaotic sequence is used to dynamically encrypt and map the spatial coordinates to generate a three-dimensional eddy current field distribution characteristic matrix with spatial continuity and anti-aliasing characteristics.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.

Images