A high-precision, multi-modal integrated circuit packaging material stress test system and method

Abstract

The application discloses an integrated circuit packaging material stress test system and method, and belongs to the field of integrated circuits. The system comprises a multi-axis stress applying device, an environment control unit, a deformation data acquisition unit and a data analysis module. By applying multi-axis composite stress and using high-precision DIC or FBG sensors to acquire strain data, real-time analysis and prediction are carried out in combination with CNN, LSTM and reinforcement learning algorithms, and the loading path is dynamically optimized. The application realizes complex working condition simulation, accurate capture of micro-deformation and intelligent optimization of the test process, and significantly improves the test precision and efficiency.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuits, and in particular relates to a high-precision, multi-modal integrated circuit (IC) packaging material stress testing system and method. Background Technology

[0002] In today's era of rapid digital and intelligent development, integrated circuit (IC) packaging is the core cornerstone of modern electronic devices, playing a crucial role. From consumer electronics such as smartphones and tablets to automotive electronic control systems, industrial automation devices, and even high-end medical equipment and aerospace instruments, all rely on high-performance integrated circuit packaging technology. With the continuous advancement of technology, electronic devices are making great strides towards miniaturization, high performance, and high reliability, which places increasingly stringent demands on the comprehensive performance of packaging materials. Specifically, the mechanical properties of packaging materials need to be sufficiently strong to withstand various external impacts, vibrations, and pressures; durability must be excellent, withstanding the tests of complex environments such as high temperature, high humidity, and chemical corrosion during long-term use; integrity is also crucial, ensuring stable connections between the chip and external components and preventing faults such as signal leakage and short circuits.

[0003] However, current integrated circuit packaging stress testing methods suffer from numerous drawbacks that urgently need to be addressed, severely hindering the further development of packaging technology. Firstly, most existing stress testing methods can only apply uniaxial stress. This testing approach is too one-sided and singular, failing to realistically simulate the complex load conditions faced by electronic devices in actual use scenarios. In real-world environments, electronic devices often endure multi-axial combined stresses. For example, during vehicle operation, onboard electronic devices must not only cope with vertical vibration stress caused by uneven road surfaces, but also withstand horizontal inertial stress generated during vehicle acceleration and braking, as well as lateral stress during cornering—a combination of complex stresses. Uniaxial stress testing clearly cannot comprehensively and accurately assess the performance of packaging materials under complex operating conditions, leading to significant discrepancies between test results and actual conditions.

[0004] Secondly, in terms of strain monitoring, existing technologies generally have low resolution, and their ability to capture minute changes such as microcracks or microdisplacements is significantly insufficient. In integrated circuit packaging structures, the generation and propagation of microcracks are often early signals leading to package failure. If these minute strain changes cannot be monitored in a timely and accurate manner, it is difficult to take effective preventive measures in the early stages of failure. For example, in high-precision chip packaging, even a microdisplacement of only a few micrometers can cause loosening of internal circuit connections, thereby affecting the normal operation of the equipment. However, due to resolution limitations, existing monitoring systems often fail to detect these potential hazards in the first instance, significantly increasing the risk of failure.

[0005] Furthermore, existing stress testing methods lack real-time data analysis and failure prediction capabilities. In the manufacturing, use, and maintenance of modern electronic equipment, real-time monitoring and data analysis of the stress state of packaging materials are crucial. Real-time data collection and analysis can promptly identify abnormal trends in stress changes and predict potential failures, providing a valuable window of opportunity for equipment maintenance and repair. However, most current testing methods rely on post-event analysis, only identifying the problem after significant damage to the packaging material has already occurred. This delayed detection approach not only fails to effectively prevent failures but may also lead to sudden equipment damage, causing substantial economic losses and safety hazards.

[0006] In summary, current integrated circuit packaging stress testing methods are no longer sufficient to meet the growing demand for high-performance packaging material research and application. There is an urgent need to develop new testing technologies that can apply multiaxial stress, have high-resolution strain monitoring, and possess powerful real-time data analysis and failure prediction capabilities. This will propel the integrated circuit packaging industry to new heights, provide a solid guarantee for the reliable operation and technological innovation of electronic devices, and is also the key to the continued healthy development of the entire electronics industry.

[0007] In the field of integrated circuit packaging stress testing, current mainstream testing methods exhibit significant limitations and urgently require improvement and breakthroughs. The specific manifestations are as follows: First, most stress loading methods employ uniaxial stress loading, which is insufficient for simulating real-world operating conditions. In actual electronic device operating environments, packaging materials often face complex and variable stress states, with multiaxial stress coupling being the norm. However, uniaxial stress loading cannot reproduce such complex scenarios, offering only a partial view and failing to comprehensively assess the mechanical response of packaging materials in real-world usage. Moreover, the ability to control environmental factors is limited. Environmental variables such as temperature and humidity are difficult to maintain precisely and consistently at set values ​​during testing. Even slight fluctuations in these environmental factors interact with stress loading, further exacerbating the deviation in test results. This results in test data that differs significantly from actual application performance, greatly impacting the accuracy of performance evaluation of packaging materials.

[0008] Secondly, in terms of strain measurement, traditional measurement methods such as strain gauges have significant drawbacks. At the microscale resolution, traditional strain gauges, limited by their physical structure and measurement principles, struggle to capture the minute strain changes accompanying the initial initiation of microcracks. The generation of microcracks is often a precursor to encapsulation failure, but traditional strain gauges fail to detect this, much like using a low-powered telescope to observe the microscopic world, missing crucial details. Regarding response speed, traditional strain gauges suffer from significant response delays. Faced with the rapidly changing strain state of encapsulation materials under complex loads, they cannot respond promptly with real-time data, resulting in coarse and distorted strain-time curves. This greatly diminishes their value in studying the dynamic mechanical behavior of encapsulation materials and fails to meet the current demand for refined and real-time characterization of encapsulation materials. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a high-precision, multi-modal stress testing system and method for integrated circuit packaging materials.

[0010] The present invention is implemented as follows: Firstly, it provides a stress testing system for integrated circuit packaging materials, comprising: Multiaxial stress application equipment is capable of applying multiaxial stress combinations to integrated circuit packaging materials; The environmental control unit can dynamically adjust the temperature and humidity of the test environment; The deformation data acquisition unit is used to capture strain data on the surface of integrated circuit packaging materials in real time after being subjected to compressive stress by a multiaxial stress application device; The data analysis module is used to extract the surface deformation mode and dynamic characteristics of strain over time from the strain data of the circuit package material, and to optimize the loading path of the multiaxial stress application device using the reinforcement learning submodule.

[0011] Preferably, the multiaxial stress combination is a combination of multiple stresses among tensile, compressive, shear, and thermal stresses.

[0012] Preferably, the multiaxial stress combination employs different types of stress simultaneously or sequentially.

[0013] Preferably, the deformation data acquisition unit employs a high-resolution microscale digital image correlation technology system to capture in real time the strain image of the integrated circuit packaging material surface after being subjected to compressive stress by a multiaxial stress application device.

[0014] Preferably, the data analysis module includes: The spatial feature extraction submodule is used to extract the deformation patterns of the circuit-encapsulated material surface based on the convolutional neural network (CNN). The input of the convolutional neural network (CNN) is the strain image obtained by the high-resolution microscale digital image correlation technology system, and the output is a spatial feature tensor, in which each point corresponds to the deformation pattern encoding, including the principal strain direction, shear angle, and displacement gradient. The time series modeling submodule is used to capture the dynamic features of strain changing over time based on a long short-term memory network (LSTM). The input of the long short-term memory network (LSTM) is the spatial feature tensor output by the convolutional neural network (CNN), and the output is the predicted strain tensor for the next time step, i.e., the predicted strain field. The reinforcement learning submodule is used to optimize the loading path of a multi-axis stress application device based on DQN (Deep Q-Network); the state of the DQN is formed by splicing the predicted strain field output by the Long Short-Term Memory Network (LSTM), the historical stress application action loading curve, and the ambient temperature and humidity, with each action corresponding to a stress application action. Preferably, the deformation data acquisition unit employs a fiber Bragg grating (FBG) sensor array and a signal processing unit. The fiber Bragg grating (FBG) sensor array is distributed at multiple key points on the surface of the packaging material. The reflected light wavelength signal acquired in real time by the fiber Bragg grating (FBG) sensor is demodulated by the signal processing unit, and the fiber Bragg grating wavelength sequence is extracted as strain data.

[0015] Preferably, the data analysis module includes: The spatial feature extraction submodule is used to reconstruct a high-resolution strain field based on a convolutional neural network (CNN). The input of the convolutional neural network (CNN) is a fiber Bragg grating wavelength sequence, and the output is a pseudo-strain image, i.e., a high-resolution strain field. The time series modeling submodule is used to capture the dynamic features of strain changing over time based on a long short-term memory network (LSTM). The input of the long short-term memory network (LSTM) is the pseudo-strain image output by the convolutional neural network (CNN), and the output is the predicted strain tensor for the next time step, i.e., the predicted strain field. The reinforcement learning submodule is used to optimize the loading path of a multiaxial stress application device based on DQN. The state of the DQN is formed by splicing the predicted strain field output by the Long Short-Term Memory Network (LSTM), the historical stress application action loading curve, and the ambient temperature and humidity. The action corresponds to the stress application action.

[0016] Preferably, the environmental control unit is capable of adjusting the temperature range to -40℃ to +150℃ and the humidity control range to -2%RH to +2%RH.

[0017] Secondly, a method for stress testing of integrated circuit packaging materials is provided, including the following steps: Apply a pre-defined multiaxial stress combination to the integrated circuit packaging material; Real-time capture of strain data on the surface of integrated circuit packaging materials after being subjected to compressive stress by a multiaxial stress application device; Based on strain data, the surface deformation mode and dynamic characteristics of strain change over time of the circuit package material are extracted, and the loading path of the multiaxial stress application device is optimized using a reinforcement learning submodule to optimize the testing process.

[0018] The beneficial effects of this invention are at least as follows: This invention uses a multiaxial stress application device to apply a combination of tensile, compressive, shear, and thermal stresses to the packaging material, which can be applied simultaneously or sequentially. This realistically reproduces the complex mechanical environment in which the material is located in the actual packaging structure, directly overcoming the limitations of traditional uniaxial loading. This allows the test results to comprehensively and accurately reflect the true mechanical properties and failure behavior of the material under multiaxial stress coupling.

[0019] This invention employs a high-resolution microscale digital image correlation (DIC) system or a fiber Bragg grating (FBG) sensor array as the deformation data acquisition unit. The DIC system utilizes a high-frame-rate, high-resolution camera to achieve sub-micron-level precision non-contact full-field strain measurement; the FBG array achieves multi-point, high-precision, and interference-resistant distributed strain sensing through demodulation of wavelength shift. This enables the system to capture microscale strain data on or within the material surface in real time, breaking through the bottleneck of microscale deformation monitoring and achieving highly sensitive capture of early failure signs such as microcrack initiation and propagation.

[0020] This invention innovatively integrates a Convolutional Neural Network (CNN), a Long Short-Term Memory Network (LSTM), and a reinforcement learning submodule based on DQN into the data analysis module. The CNN submodule is responsible for extracting high-dimensional spatial deformation features from strain images or FBG data; the LSTM submodule models the high-dimensional spatial deformation features, captures the dynamic law of strain evolution over time, and outputs a predicted strain field. Based on this, the reinforcement learning (DQN) submodule uses the predicted strain field, historical loading actions, and environmental parameters as states to dynamically plan and optimize the loading path of the multi-axis stress application device, achieving high-accuracy prediction of failure points and stress concentration areas. It can also adjust the testing strategy in real time, thereby actively avoiding meaningless loading, accelerating the exposure of potential defects, and greatly improving testing efficiency and scientific rigor.

[0021] In summary, this invention systematically solves the three major pain points in current stress testing of integrated circuit packaging materials—namely, distortion in working condition simulation, insufficient monitoring sensitivity, and lack of intelligent prediction and optimization capabilities—by organically combining core technologies such as multi-axis composite loading, high-precision microscale sensing, artificial intelligence analysis and prediction with closed-loop optimization, and high-precision environmental control. It provides a high-precision, intelligent, and closed-loop-optimizable testing solution. Attached Figure Description

[0022] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a system framework diagram of an integrated circuit packaging material stress testing system provided in an embodiment of the present invention.

[0024] Figure 2 This is a flowchart of a method for stress testing of integrated circuit packaging materials provided in an embodiment of the present invention.

[0025] Figure 3 This is a system layout diagram of the overall architecture of the testing system provided in this embodiment of the invention.

[0026] Figure 4 This is a flowchart of the stress application process provided by an embodiment of the present invention, which includes multiaxial stress application, environmental control, and data feedback closed loop.

[0027] Figure 5 This is an example analysis report of the test system output results provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0029] This embodiment provides a stress testing system for integrated circuit packaging materials. See attached document. Figure 1 include: Multiaxial stress application equipment is capable of applying multiaxial stress combinations to integrated circuit packaging materials; The environmental control unit can dynamically adjust the temperature and humidity of the test environment; The deformation data acquisition unit is used to capture strain data on the surface of integrated circuit packaging materials in real time after being subjected to compressive stress by a multiaxial stress application device; The data analysis module is used to extract the surface deformation mode and dynamic characteristics of strain over time from the strain data of the circuit package material, and to optimize the loading path of the multiaxial stress application device using the reinforcement learning submodule.

[0030] The stress testing method for integrated circuit packaging materials of the above system is also provided; see appendix. Figure 2 Includes the following steps: Apply a pre-defined multiaxial stress combination to the integrated circuit packaging material; Real-time capture of strain data on the surface of integrated circuit packaging materials after being subjected to compressive stress by a multiaxial stress application device; Based on strain data, the surface deformation mode and dynamic characteristics of strain change over time of the circuit package material are extracted, and the loading path of the multiaxial stress application device is optimized using a reinforcement learning submodule to optimize the testing process.

[0031] Specifically, the multiaxial stress application device, as the core component of the system, adopts an advanced modular design concept, possessing the ability to apply tensile (-500MPa to +500MPa), compressive (-500MPa to +500MPa), shear (-300MPa to +300MPa), and thermal stress (-40℃ to +150℃ cycling). This modular design strategy not only simplifies the installation and commissioning process of the equipment but also greatly enhances its flexibility and scalability, allowing for customized configuration according to different testing needs. Simultaneously, the stress loading modules within the equipment work collaboratively through a precise mechanical transmission mechanism and a high-precision sensor network, ensuring that the application of various stresses accurately reaches the set values, thereby enabling the encapsulation material to fully demonstrate its mechanical properties under complex multiaxial stress conditions during experiments.

[0032] This invention also introduces a multi-degree-of-freedom stress actuator into the multiaxial stress application device, which can realize precise multiaxial stress application and dynamic adjustment, meeting the needs of complex stress testing.

[0033] The environmental control unit plays a crucial role in the testing system, providing a high-precision environmental simulation platform for the multiaxial stress application equipment. The temperature control module employs an advanced PID control algorithm, using high-precision heating and cooling elements to achieve precise temperature control within an error range of -0.5℃ to +0.5℃. The humidity control module utilizes a highly sensitive humidity sensor and efficient humidification and dehumidification devices to achieve humidity control accuracy within a range of -2% RH to +2% RH. Furthermore, the dynamic adjustment function of the environmental control unit allows it to quickly respond to and adjust environmental parameters based on real-time monitoring and feedback information from the stress loading module, thereby providing a stable and controllable testing environment for the packaging materials and minimizing the interference of environmental factors on the test results. This invention, through the environmental control unit, can dynamically adjust the temperature (e.g., -40℃ to +150℃) and humidity (e.g., control accuracy ±2% RH) of the testing environment (claim 8). This technology provides a precise, stable, and controllable environmental simulation platform for stress testing, effectively isolating noise introduced by fluctuations in environmental variables, ensuring the consistency, comparability, and high repeatability of test results in different batches or under different conditions, and providing a reliable basis for the performance evaluation of materials.

[0034] The deformation data acquisition unit is a key technical means to achieve accurate strain monitoring. It is used to capture the strain data of the surface of integrated circuit packaging materials after being subjected to compressive stress by a multiaxial stress application device in real time.

[0035] The data analysis module, driven by artificial intelligence (AI), is the intelligent core of this invention, endowing the testing system with powerful data processing and predictive capabilities. This system employs various advanced AI algorithms, including Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Depth-Quantity Reinforcement (DQN).

[0036] In one embodiment, the deformation data acquisition unit employs a high-resolution microscale digital image correlation (DIC) system. This system is equipped with four synchronous high-speed industrial cameras, each with a frame rate of up to 5000 fps and a resolution of 2048×2048. It can capture real-time compressive deformation images of the encapsulation material under multiaxial stress with extremely high temporal and spatial resolution, reflecting minute deformations and displacements. The synchronization control accuracy is within 10 μs, ensuring high synchronization of images acquired by multiple cameras and providing a reliable data foundation for subsequent image processing and data analysis. To further improve image quality, the high-resolution microscale DIC system uses wavelet transform and adaptive histogram equalization algorithms to filter image noise after acquiring the compressive deformation images before inputting them into the data analysis module. Wavelet transform effectively decomposes the image signal, separates and suppresses noise components, while preserving important details in the image. The adaptive histogram equalization algorithm automatically adjusts the image contrast according to the local characteristics of the image, enhancing the visual effect and making minute deformation features more clearly discernible, providing high-quality image data for accurate strain measurement.

[0037] For example, in the deformation data acquisition unit, the CNN is mainly used to extract the spatial features of material surface deformation from the deformation images. Through learning and training on a large number of deformation images, the CNN can automatically identify various feature patterns related to material deformation in the images, such as strain distribution and displacement field, and transform these spatial features into quantifiable data information. LSTM focuses on modeling the strain time series of materials, capturing the dynamic laws and trends of strain changes over time, thereby revealing the aging behavior and potential failure signs of materials under long-term stress. The DQN algorithm optimizes the stress application strategy through reinforcement learning. Based on the real-time monitored material strain data and the predicted failure risk, it dynamically adjusts the order and magnitude of stress loading to achieve efficient testing and accurate evaluation of material performance. Based on the synergistic effect of these AI algorithms, the system can achieve a failure prediction accuracy of ≥90% and a material remaining life prediction error of ≤±5%, providing a solid technical guarantee for the reliability and life assessment of integrated circuit packaging materials.

[0038] This invention achieves precise capture of micron-level deformation and microcracks in encapsulation materials by combining a multiaxial stress application device with microscale DIC technology. The multiaxial stress application system can simulate complex stress states in real-world application scenarios, ensuring a comprehensive evaluation of material performance under the combined action of multiple directions and types of stress. Meanwhile, the strain data acquisition unit, employing a microscale DIC system with its high-resolution industrial camera and advanced image processing algorithms, acts like a high-precision "microscope" for detection, clearly observing micron-level changes on the material surface and precisely capturing the initiation and propagation of microcracks. This provides crucial data support for early failure warning and effectively improves the sensitivity and accuracy of the test.

[0039] The data analysis module incorporates artificial intelligence (AI), enabling real-time intelligent optimization of the testing process. This module can analyze image data acquired by the DIC system in real time, using deep learning algorithms to quickly extract material deformation characteristics and accurately predict failure points and stress concentration areas. Based on the prediction results, the system dynamically adjusts the stress application strategy through reinforcement learning algorithms, automatically optimizing the sequence and magnitude of stress loading to achieve closed-loop control. This real-time intelligent optimization not only improves testing efficiency and reduces ineffective or excessive loading but also significantly enhances the scientific rigor and rationality of the testing, providing a more accurate basis for material performance evaluation.

[0040] For example, the data analysis module in this embodiment includes a spatial feature extraction submodule, a time series modeling submodule, and a reinforcement learning submodule; A spatial feature extraction submodule is used to extract surface deformation patterns of circuit packaging materials based on a convolutional neural network (CNN); the CNN is a lightweight U-Net architecture, comprising: (1) Input layer: Receives grayscale strain images (size 1024×1024 pixel, 16-bit) acquired by the high-resolution microscale DIC system at 500 Hz. The images have been preprocessed with zero mean normalization and CLAHE enhancement. (2) 3-layer downsampling path: Each layer consists of 2 3×3 convolutions (ReLU) + 1 2×2 max pooling, with the number of channels being 32-64-128 respectively; a spatial attention module (SAM) is introduced at the end of the downsampling path to generate a spatial weight map M∈R^(128×128×128), which is used to strengthen local high gradient regions such as shear bands and microcracks; (3) 3-layer upsampling path: Each layer consists of 2 3×3 convolutions + 1 2×2 transposed convolution, with the number of channels reversed from 128 to 64 to 32; skip connections fuse downsampling features, and finally output spatial feature tensor F with a size of 1024×1024×32, where the 32-dimensional vector of each pixel is the deformation mode encoding at that position (including principal strain direction, shear angle, displacement gradient, etc.). The time series modeling submodule is used to capture the dynamic features of strain changing over time based on a Long Short-Term Memory (LSTM) network; the LSTM is a two-layer bidirectional LSTM, with the following specific configuration: (1) Input: Divide the spatial feature tensor F output by CNN into T=20 frames according to time step t=1…T. Expand each frame into a 1×1048576 vector along the spatial dimension H×W, and then reduce the dimension to 1×256 by PCA to form the spatiotemporal feature sequence X. (2) Network structure: 128 hidden units per layer, dropout=0.2; the hidden state h is output after concatenating the forward and backward hidden states; (3) Output head: Fully connected layer + ReLU, outputs the predicted strain tensor Ŷ for the next time step, with a size of 1024×1024×1.

[0041] The LSTM loss function is the sum of mean squared error and gradient difference error: β=0.1 is used to penalize strain gradient prediction bias.

[0042] The reinforcement learning submodule is used to optimize the loading path of a multi-axis stress application device based on DQN. It is deployed on an STM32H7+FPGA edge computing card; the inference latency is <5 ms, and the optimal loading action is output in real time, controlled by a servo motor via EtherCAT bus. Experimental results show that compared with the traditional PID loading strategy, the failure point advance time is extended from 1.2 s to 1.4 s, and the stress concentration factor is reduced by 9.3%. The DQN adopts a dual-network structure (eval-net & target-net), and the state space s, action space a, and reward r are defined as follows: (1) State s: It is composed of the current strain field output by LSTM, historical loading curve, and ambient temperature and humidity. Dimension = 1024×1024+200+2; (2) Action a: Discretize 9 loading modes (axial tension, biaxial tension, pure shear, tension-shear combination, etc.) × 5 speed levels (0.1-1 mm / s), for a total of 45 discrete actions; (3) Reward r: Where ε is the uniform elongation of the material, ε_target is the target uniform elongation, t is the loading time, and λ=0.01 is used to penalize the time cost; (4) Training: An ε-greedy strategy was adopted, with an experience replay pool size of 10,000, a batch size of 64, and a learning rate of 1×10⁻⁶. -3 The target-net is updated every 50 steps. After training, DQN outputs the optimal action sequence, which delays the occurrence of failure points by ≥15% and reduces the stress concentration factor by ≥8%.

[0043] In another embodiment, the strain data acquisition unit uses a fiber Bragg grating (FBG) sensor array and a signal processing unit to acquire material strain data. An FBG sensor is a sensor based on fiber Bragg grating technology, capable of high-precision and high-sensitivity measurement of physical quantities such as strain and temperature of the measured object. Fabricating the FBG sensors in an array allows for distributed strain monitoring of multiple key points on the surface of the encapsulation material. Each FBG sensor reflects the magnitude and direction of local strain through changes in the wavelength of its reflected light, offering advantages such as non-contact operation, resistance to electromagnetic interference, high precision, and high resolution. Furthermore, it can be directly attached to or embedded within the encapsulation material, enabling non-destructive monitoring of internal stress.

[0044] During monitoring, the FBG sensor array acquires real-time strain data of the encapsulation material under multiaxial stress and converts it into optical wavelength signals, which are then transmitted to the signal processing unit via optical fiber. The signal processing unit demodulates the received optical wavelength signals and extracts precise strain information. This strain data is then input into the data analysis module.

[0045] The data analysis module utilizes deep learning algorithms, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs), to extract features and dynamically model multi-dimensional strain data. CNNs analyze the spatial distribution characteristics of strain and identify potential stress concentration areas, while LSTMs model time-series data, capturing the trend of strain changes over time to predict the formation and propagation paths of microcracks. Through reinforcement learning (DQN) algorithms, the data analysis module can perform real-time feedback control of the multi-axis stress application device, dynamically adjusting the stress loading strategy to optimize the testing process and improve testing efficiency and accuracy. Compared to DIC systems, the FBG sensor array solution has unique advantages in certain special application scenarios, such as light-sensitive packaging materials or harsh environments such as high temperature and high humidity, providing more stable and reliable monitoring data and strong support for the performance evaluation and reliability verification of IC packaging materials.

[0046] For example, the data analysis module in this embodiment includes a spatial feature extraction submodule, a time series modeling submodule, and a reinforcement learning submodule; A spatial feature extraction submodule is used to process FBG wavelength sequences based on 1D-CNN, thereby reconstructing a high-resolution strain field; the 1D-CNN sequentially includes: (1) Input layer: Receive FBG reflection wavelength offset sequence Δλ∈R^n×1 (n is the number of sensors, typically ≥36), which has been baseline removed and Savitzky-Golay smoothed; (2) 3-layer one-dimensional convolution: kernel length 7-5-3, number of channels 16-32-64, each layer is followed by BatchNorm+ReLU+MaxPool (pooling size=2). (3) Fully connected layer: outputs the reconstructed strain field vector ε∈R m (m=1024, corresponding to the number of DIC mesh nodes), through the pre-calibrated mapping matrix P∈R (m×n) Dense the sparse data from the sensor; (4) Output: ε is reshaped into a 32×32 grid and then upsampled to 1024×1024 to obtain the pseudo strain image I (i.e., the strain field image reconstructed by FBG), which is used as the input for subsequent LSTM. The time series modeling submodule is used to capture the dynamic features of strain changing over time based on a Long Short-Term Memory (LSTM) network; the LSTM is a two-layer bidirectional LSTM, with the following specific configuration: (1) Input: The pseudo-strain image I output by CNN is divided into T frames according to time step t=1…T. Each frame is expanded into a vector along the spatial dimension and then reduced in dimensionality by PCA to form a spatiotemporal feature sequence X; (2) Network structure: 128 hidden units per layer, dropout=0.2; the hidden state h is output after concatenating the forward and backward hidden states; (3) Output head: Fully connected layer + ReLU, output the predicted strain tensor Ŷ for the next time step.

[0047] The reinforcement learning submodule is the same as in the previous embodiment.

[0048] This invention provides strong quantitative evidence for the selection and reliability verification of IC packaging materials, effectively supporting process optimization. By accurately evaluating the performance of different materials under complex stress conditions, the system can provide reliable data support for IC packaging designers, helping them select higher-quality packaging materials more suitable for specific application scenarios. Simultaneously, the system's accurate prediction and analysis of material failure processes contributes to a deeper understanding of the failure mechanisms of packaging materials, providing crucial guidance for optimizing packaging process parameters, extending package lifespan, and improving overall product reliability. This, in turn, drives the continuous advancement of IC packaging technology to meet the ever-increasing high-performance demands of electronic devices.

[0049] Appendix Figure 3The system layout diagram is a visual representation of the overall architecture of the testing system of this invention, clearly showing the layout of each functional unit and their connections. In the diagram, the multiaxial stress application device is located at the center. Its modular design allows for the tight integration of tensile, compressive, shear, and thermal stress application modules, ensuring that various stresses are precisely applied to the packaged sample. Environmental control units are distributed around the device, with heating, cooling, humidification, and dehumidification devices surrounding the stress application device, forming a comprehensive environmental control system that dynamically provides stable temperature and humidity conditions for testing. Four high-speed industrial cameras of the high-resolution microscale DIC system are aimed at the packaged sample from different angles, ensuring comprehensive capture of deformation images of the sample. They are connected to the central processing unit via high-speed data transmission lines, transmitting the acquired image data to the AI-driven data analysis and prediction system in real time. The entire layout diagram not only reflects the spatial distribution of the system's various parts but also highlights the collaborative working relationships between the units, providing users with a clear visual guide to understand the overall system architecture and workflow.

[0050] Appendix Figure 4 The stress application flowchart details the process of multiaxial stress application, environmental control, and data feedback closed-loop control. The process begins with the initial state detection of the packaged sample. The AI ​​system formulates an initial stress application plan based on the sample's initial characteristics. Subsequently, the multiaxial stress application device applies tensile, compressive, shear, and thermal stresses to the sample according to the set plan. During stress application, the environmental control unit monitors and adjusts environmental parameters in real time to ensure that temperature and humidity are maintained within the set range. Simultaneously, the microscale DIC system continuously acquires deformation images of the sample and transmits the data to the data analysis module in real time. The data analysis module uses convolutional neural networks (CNN) and long short-term memory networks (LSTM) to analyze the strain data, extracting spatial and temporal features, and predicting the material's failure points and stress concentration areas. If the prediction results indicate that the sample is close to failure or the environmental parameters deviate from the set range, the reinforcement learning (DQN) algorithm optimizes the stress application strategy, dynamically adjusting the order and magnitude of stress loading, forming a closed-loop feedback control process. This flowchart comprehensively demonstrates the system's automation, intelligence, and dynamic adjustment capabilities, enabling users to clearly understand the collaborative working mechanism of stress application and environmental control.

[0051] Appendix Figure 5The example analysis report is a concrete presentation of the output results of the testing system of this invention, containing rich data analysis and visualization content. The core of the report is the stress-strain curves, which clearly demonstrate the mechanical response of the encapsulation material under multiaxial stress. The horizontal axis represents the applied stress, and the vertical axis represents the corresponding strain. The shape and slope of the curves reflect key mechanical properties of the material, such as elastic modulus, yield strength, and ductility. The 3D deformation visualization section visually presents the overall deformation of the material under complex stress through three-dimensional images, clearly observing minute displacements and deformation distributions, helping users quickly identify potential stress concentration areas. The failure point prediction heatmap is one of the highlights of the report. It uses color intensity to indicate the level of failure risk, with red areas representing high-risk failure points, providing users with an intuitive warning of failure risk. Finally, the performance index data section lists various performance indicators of the material, such as maximum stress, strain energy density, and fatigue life, and compares them with standard values ​​or historical data to quantitatively evaluate the material's performance. The final output includes a comprehensive analysis report containing stress-strain curves, 3D deformation fields, failure prediction heatmaps, and quantitative performance indicators, providing researchers and engineers with a powerful tool for in-depth understanding of packaging material performance and optimization design. This provides direct performance comparison data for packaging material selection and clear mechanism visualization tools for failure analysis. More importantly, AI-based prediction and optimization capabilities offer forward-looking data guidance and simulation verification methods for optimizing packaging structure design and process parameters (such as molding compound selection, curing curves, and wiring schemes), thereby shortening the R&D cycle, reducing trial-and-error costs, and fundamentally improving the long-term reliability and lifespan of integrated circuit packaging.

[0052] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A stress testing system for integrated circuit packaging materials, characterized in that, The system includes: Multiaxial stress application equipment is capable of applying multiaxial stress combinations to integrated circuit packaging materials; The environmental control unit can dynamically adjust the temperature and humidity of the test environment; The deformation data acquisition unit is used to capture strain data on the surface of integrated circuit packaging materials in real time after being subjected to compressive stress by a multiaxial stress application device; The data analysis module is used to extract the surface deformation mode and dynamic characteristics of strain over time from the strain data of the circuit package material, and to optimize the loading path of the multiaxial stress application device using the reinforcement learning submodule.

2. The system according to claim 1, characterized in that, The multiaxial stress combination is a combination of various stresses including tensile, compressive, shear, and thermal stresses.

3. The system according to claim 2, characterized in that, The multiaxial stress combination employs different types of stress simultaneously or sequentially.

4. The system according to claim 1, characterized in that, The strain data acquisition unit employs a high-resolution microscale digital image correlation technology system to capture strain images of the surface of integrated circuit packaging materials after being subjected to compressive stress by a multiaxial stress application device in real time.

5. The system according to claim 4, characterized in that, The data analysis module includes: The spatial feature extraction submodule is used to extract the deformation patterns of the material surface of the circuit-encapsulated material based on the convolutional neural network. The input of the convolutional neural network is the strain image obtained by the high-resolution microscale digital image correlation technology system, and the output is a spatial feature tensor, in which each point corresponds to the deformation pattern encoding, including the principal strain direction, shear angle, and displacement gradient. The time series modeling submodule is used to capture the dynamic features of strain changing over time based on a long short-term memory network. The input of the long short-term memory network is the spatial feature tensor output by the convolutional neural network, and the output is the predicted strain tensor for the next time step, i.e., the predicted strain field. The reinforcement learning submodule is used to optimize the loading path of a multiaxial stress application device based on DQN. The state of the DQN is formed by splicing the predicted strain field output by the long short-term memory network, the historical stress application action loading curve, and the ambient temperature and humidity, with each action corresponding to a stress application action.

6. The system according to claim 1, characterized in that, The strain data acquisition unit employs a fiber Bragg grating sensor array and a signal processing unit. The fiber Bragg grating sensor array is distributed at multiple key points on the surface of the packaging material. The reflected light wavelength signal acquired in real time by the fiber Bragg grating sensor is demodulated by the signal processing unit, and the fiber Bragg grating wavelength sequence is extracted as strain data.

7. The system according to claim 6, characterized in that, The data analysis module includes: The spatial feature extraction submodule is used to reconstruct a high-resolution strain field based on a convolutional neural network. The input of the convolutional neural network is a fiber Bragg grating wavelength sequence, and the output is a pseudo-strain image, i.e., a high-resolution strain field. The time series modeling submodule is used to capture the dynamic features of strain changing over time based on a long short-term memory network. The input of the long short-term memory network is the pseudo-strain image output by the convolutional neural network, and the output is the predicted strain tensor for the next time step, i.e., the predicted strain field. The reinforcement learning submodule is used to optimize the loading path of a multiaxial stress application device based on DQN. The state of the DQN is formed by splicing the predicted strain field output by the long short-term memory network, the historical stress application action loading curve, and the ambient temperature and humidity, with each action corresponding to a stress application action.

8. The system according to claim 1, characterized in that, The environmental control unit can adjust the temperature range from -40℃ to +150℃ and the humidity control range from -2%RH to +2%RH.

9. A method for stress testing of integrated circuit packaging materials based on the system described in any one of claims 1-8, characterized in that, The method includes the following steps: A preset multiaxial stress combination is applied to the integrated circuit packaging material; Real-time capture of strain data on the surface of integrated circuit packaging materials after being subjected to compressive stress by a multiaxial stress application device; Based on strain data, the surface deformation mode and dynamic characteristics of strain change over time of the circuit package material are extracted, and the loading path of the multiaxial stress application device is optimized using a reinforcement learning submodule to optimize the testing process.

Images