Temperature-compensated intelligent prediction system for a probe station
By combining multi-source sensing and deep learning, the probe station system solves the problem of thermal deformation compensation in high and low temperature testing environments, achieving high-precision alignment and stable production efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU PANYU POLYTECHNIC
- Filing Date
- 2026-04-15
- Publication Date
- 2026-07-14
AI Technical Summary
Existing probe stations lack sufficient accuracy in thermal deformation compensation under high and low temperature testing environments, and cannot adapt to dynamically changing temperature fields and wafer batch differences, resulting in fluctuations in testing accuracy and reduced equipment capacity.
It employs a multi-source sensing module, a data preprocessing module, a central computing module, a compensation execution module, a visual verification module, and a human-computer interaction module, combined with high-precision sensors and deep learning algorithms, to achieve real-time thermal deformation prediction and adaptive compensation.
It significantly improves the alignment accuracy and test yield of the probe station under drastic temperature changes, reduces the need for frequent realignment, and enhances the robustness and production efficiency of the equipment.
Smart Images
Figure CN122387243A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of semiconductor testing technology, and in particular to a temperature compensation intelligent prediction system for a probe station. Background Technology
[0002] As integrated circuit (IC) manufacturing processes continue to evolve towards advanced process nodes, the micron-level or even sub-micron-level precision of chip pad alignment in wafer probe testing has become a core requirement. As a key piece of equipment connecting front-end manufacturing and back-end packaging, the motion positioning accuracy of the probe station under complex temperature fields directly affects test yield, test efficiency, and equipment reliability. Currently, most probe stations on the market employ scaling compensation based on static formulas or periodic mechanical correction modes to cope with thermal deformation under high and low temperature testing environments, lacking adaptive real-time predictive capabilities based on the fusion of physical mechanisms and sensor data.
[0003] Existing technologies for probe station thermal deformation alignment compensation still have significant limitations: First, static linear compensation typically relies solely on a single thermal expansion formula for global scaling, ignoring the temperature gradient differences between the wafer edge and center, as well as the nonlinear bending effects caused by multilayer stacking, resulting in limited compensation accuracy under large temperature differences. Second, periodic realignment requires periodic interruptions in the testing process for visual calibration. While this can correct accumulated deviations, it significantly increases system cycle time and reduces uptime per hour (UPH), and cannot suppress ongoing thermal drift during the interval between realignments. Third, traditional lookup table methods rely on pre-set offline measurement data, making it difficult to adapt to complex conditions such as dynamically changing temperature fields, wafer batch differences, and fixture stress in actual production, resulting in insufficient system generalization and flexibility. Finally, existing compensation schemes generally lack a quantitative prediction mechanism for thermal hysteresis effects, failing to effectively combine thermoelasticity laws with temporal characteristic data, leading to fluctuations in alignment accuracy during heating and cooling processes. These factors collectively restrict the intelligence level and overall production efficiency of probe stations in extreme temperature testing scenarios. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a temperature compensation intelligent prediction system for probe stations. Through a high-precision multi-source sensor layout, a physically guided variable structure deep learning algorithm, and a nanometer-scale piezoelectric actuator, it successfully solves the urgent need for high-precision alignment in probe testing under dynamic and drastic temperature changes. This not only significantly improves the test yield but also ensures the robustness of the equipment during long-term operation through online adaptive evolution logic.
[0005] This invention provides a temperature compensation intelligent prediction system for a probe station, the system comprising a multi-source sensing module, a data preprocessing module, a central computing module, a compensation execution module, a visual verification module, and a human-computer interaction module; The multi-source sensing module is used to perform multimodal digital reconstruction of the physical temperature field of the wafer testing environment and generate multidimensional feature vectors. The data preprocessing module is used to monitor the motion coordinates and thermoelastic mechanical parameters of the wafer stage in real time. The central computing module has a built-in deep learning model guided by physical mechanisms, which is used to predict and generate thermal deformation compensation instructions at a specific time in the future based on multi-dimensional feature vectors and thermoelastic mechanical benchmarks. The compensation execution module includes a piezoelectric ceramic fine-tuning platform and a multi-axis motion controller, which are used to dynamically adjust the spatial relative displacement of the wafer stage or probe card according to the thermal deformation compensation command. The visual verification module is used to collect the actual residual of the wafer alignment mark points during the test interval, and after identifying the deviation threshold, it sends a weight fine-tuning request to the central computing module to execute an online adaptive update strategy. The human-computer interaction module renders and displays the thermal drift compensation curve and system status in real time through a graphical interface, provides a model parameter configuration interface, and has a built-in model prediction confidence evaluation program.
[0006] Preferably, the multi-source sensing module includes a center temperature sensor, an edge sensor group, and a probe card temperature sensor; The center temperature sensor is embedded at the geometric center of the wafer chuck and is used to collect real-time temperature data at the center point of the wafer. The edge sensor group is distributed in four quadrants on the edge of the wafer chuck to monitor the temperature gradient difference in the radial direction of the wafer, thus compensating for the sensing defects of single-point sampling under non-uniform temperature fields. The probe card temperature sensor is located on the back of the probe card and uses an integrated digital sensing chip. Its surface is covered with a high thermal conductivity insulating encapsulation layer. It is used to collect local hot spot information of the test environment and distinguish different test stages such as preheating, steady state and cooling. During the feature extraction process, the central computing module receives the multi-point temperature sequence and first performs physical baseline calculation; then it calls the variable structure recurrent neural network to perform nonlinear feature extraction; then it performs spatial mapping and fusion of the historical alignment residuals to generate a multidimensional input vector.
[0007] Preferably, the data preprocessing module includes a temperature change rate extractor, a coordinate feedback encoder, a thermal expansion parameter memory, and a residual sequence buffer; The temperature change rate extractor calculates the current time in real time using differential operations. parameter; The coordinate feedback encoder monitors the current absolute coordinates of the wafer stage in real time through the grating ruler signal, which is used to determine the stroke redundancy of the compensation actuator. The thermal expansion parameter memory stores the thermal expansion coefficients and geometric dimensions of wafers made of different materials; The residual sequence buffer updates the alignment deviation value at historical moments in real time using a first-in-first-out protocol.
[0008] Preferably, during the time window determination stage of the neural network, the data preprocessing module dynamically adjusts the window size based on the temperature change rate and prediction error, and the window adjustment equation is as follows: ; In the formula, for The adaptive time window length at each moment. and For adjustment coefficients, This represents the absolute value of the current rate of temperature change. This represents the standard deviation of recent prediction errors.
[0009] Preferably, the compensation execution module includes: The piezoelectric ceramic fine-tuning platform adopts a six-axis parallel structure, and the ceramic stacked substrate is made of high-rigidity aluminum nitride material and is fastened to the wafer stage support. The multi-axis motion controller includes a DSP processing core and a high-voltage drive circuit; wherein, the high-voltage drive circuit supports 20-bit DAC output, which can realize nanometer-level step adjustment of the wafer stage in the X and Y directions; In addition, the compensation execution module is also configured with a physical compensation logic unit; the physical compensation logic unit is located inside the central computing module, and calculates static deformation variables as input to the initial bias term of the neural network.
[0010] Preferably, the central computing module executes differentiated prediction and compensation logic based on the working condition perception results as follows: Under drastic temperature changes, if the absolute value of the temperature change rate is greater than the preset temperature threshold, the central computing module actively increases the length of the adaptive time window, switches the routing weight to the fast response loop unit mode, and increases the frequency of model forward inference; if the absolute value of the temperature change rate is less than the preset temperature threshold, the routing weight is switched to the slow response loop unit mode. Under steady-state testing conditions, the data from the residual sequence buffer is read in real time; when the standard deviation of the prediction error is detected to be lower than the preset tolerance, the number of active nodes in the hidden layer is reduced and the learning rate is lowered. Under long-period drift conditions, the system receives data from the coordinate feedback encoder in real time. When the number of consecutive probe insertions exceeds a preset threshold, the system drives the piezoelectric ceramic fine-tuning platform to perform first-order linear compensation of the accumulated residual, thereby achieving real-time offsetting of thermal drift of the optical axis relative to the test trajectory. In addition, based on the predicted variance detected by the visual verification module, if it is determined that the current condition is uncertain, the predicted variance will be output as a safety redundancy limit for compensation.
[0011] Preferably, the visual verification module includes a CCD camera and a feature matching unit; The CCD camera is used to capture the actual spatial coordinates of the marker points on the wafer surface; The feature matching unit then uses a convolution operator to perform sub-pixel level edge detection on the acquired image; When the actual alignment residual is detected to exceed the preset dead zone range, a high-priority interrupt request is immediately sent to the central computing module, and an adaptive fine-tuning strategy is executed. The adaptive fine-tuning strategy is as follows: first, the actual deviation value is used as the true prediction value to calculate the loss function of the current model, and backpropagation is performed only for the currently activated learning weights; when the prediction accuracy recovers to the preset range, parameter optimization is stopped to ensure the convergence and real-time performance of the system control loop.
[0012] Preferably, the compensation execution module is surrounded by a physical protection structure, which includes a shell made of Invar alloy, and the surface of the shell is provided with a heat-reflective coating.
[0013] Preferably, the system is further configured with a power stabilization unit, which is connected to the high-voltage purified power supply of the probe station. The power stabilization unit internally employs a multi-stage precision voltage regulation circuit and a transient suppression array to absorb the reverse electromotive force generated by the piezoelectric ceramic fine-tuning platform during rapid response.
[0014] Compared with the prior art, the present invention has significant technical advantages: 1. By integrating physical mechanism guidance with deep learning dynamic structure analysis, the compensation execution module is driven to achieve full-dimensional adaptive and proactive adjustment of alignment parameters. This abandons the traditional static lookup table mode and avoids the problem of unpredictable thermal hysteresis under drastic temperature changes. At the same time, it accurately matches the thermal deformation characteristics of wafers of various materials, effectively eliminates cumulative drift, suppresses nonlinear distortion, and significantly improves the alignment accuracy and insertion success rate of wafer probe testing. 2. The visual verification module identifies prediction deviations in real time and links with the central computing module to execute online evolution and smooth update strategies for model weights. This effectively avoids model failure caused by wafer batch differences or sudden environmental changes, while ensuring the continuity of the testing process without interrupting the process for frequent realignment. 3. Low latency in real-time compensation calculation is ensured through a variable structure neural network mechanism. At the same time, the human-computer interaction module enables switching between physical parameters and data models as dual-drive modes and confidence monitoring. While ensuring submicron-level adaptive compensation capability, it is compatible with the high dynamic response requirements under complex testing conditions, thereby improving the overall production efficiency of the probe station in advanced process testing. Attached Figure Description
[0015] Figure 1 This is a system block diagram of the real-time advance compensation system for wafer thermal deformation provided by the present invention.
[0016] Figure 2 This is a flowchart illustrating the working condition identification and differential prediction compensation logic in this invention.
[0017] Figure 3 This is a schematic diagram illustrating the principle of predicting and compensating for thermal hysteresis under drastic temperature changes in this invention. Detailed Implementation
[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.
[0019] Example 1 Please see Figures 1-3 This invention provides a temperature compensation intelligent prediction system for probe stations. Its overall architecture is logically rigorous and highly integrated, aiming to provide sub-micron-level alignment compensation for wafer testing of advanced process chips under complex temperature variations. The system includes a multi-source sensing module, a data preprocessing module, a central computing module, a compensation execution module, a visual verification module, and a human-machine interaction module. Each module communicates bidirectionally with the central computing module via a high-speed industrial Ethernet bus. In this embodiment, the high-speed industrial Ethernet bus employs distributed clock synchronization technology, with a communication rate set to 100Mbps and a synchronization cycle of 500μs for key compensation commands, ensuring the real-time and deterministic transmission of thermoelastic mechanical data within the system.
[0020] The multi-source sensing module is used for multimodal digital reconstruction of the physical temperature field of the wafer testing environment. It includes a center temperature sensor, an edge sensor group, and a probe card temperature sensor. The center temperature sensor is embedded at the geometric center of the wafer chuck, employing a high-precision PT1000 platinum resistance thermometer with a temperature measurement accuracy better than ±0.01℃, used to collect real-time temperature data at the wafer center point. The edge sensor group is distributed in four quadrants along the edge of the wafer chuck, integrating a thermistor array. Its distribution radius is set to 0.9 times the wafer radius, with a sampling frequency of 100Hz, used to monitor the radial temperature gradient difference of the wafer, compensating for the sensing defects of single-point sampling in non-uniform temperature fields. The probe card temperature sensor is located on the back of the probe card, employing an integrated digital sensing chip covered with a high thermal conductivity insulating encapsulation layer. Its dynamic detection range covers -55℃ to +150℃, used to collect local hotspot information of the testing environment, distinguishing between different testing stages such as preheating, steady state, and cooling. During feature extraction, the multi-source sensing module transmits the collected multi-point temperature sequences to the central computing module via shielded cables. Upon receiving the sequences, the central computing module first performs noise reduction, linear interpolation, and baseline offset processing on the original sequences. Then, it calls the built-in variable structure recurrent neural network to perform nonlinear feature extraction, thereby accurately identifying the current thermal deformation evolution trend of the wafer (such as center bulging, edge curling, or overall expansion). The visual verification module is responsible for compensating for the perception blind spots of the physical model during test intervals. The image stream generated by the high-resolution CCD camera is processed by a sub-pixel-level edge detection algorithm to extract the actual displacement residuals of the alignment markers. The data from each sensor together provide a reference for the formulation of the system's compensation strategy.
[0021] Furthermore, the central control module spatially maps and fuses the radial gradient fed back by the edge sensor group with the time-domain curve of the central temperature sensor to establish a unified wafer thermodynamic coordinate system, and calculates the static deformation benchmark using physical mechanisms, ultimately generating a multi-dimensional feature vector containing thermal hysteresis characteristics, spatial distortion weights, and prediction variance; the multi-dimensional feature vector is updated at a frequency of 200 times per second, serving as the real-time base for generating prediction compensation instructions.
[0022] The data preprocessing module is used to monitor the motion coordinates and thermoelastic mechanical parameters of the wafer stage in real time, including a temperature change rate extractor, a coordinate feedback encoder, a thermal expansion parameter memory, and a residual sequence buffer; wherein, the temperature change rate extractor calculates the current moment's data in real time through differential operations. The parameters include: the coordinate feedback encoder connected to the grating ruler below the wafer stage, employing high-resolution electronic interpolation technology to monitor the current absolute coordinates of the wafer stage in real time, used to determine the travel redundancy of the compensation actuator; the thermal expansion parameter memory stores the coefficient of thermal expansion (CTE) and geometric dimensions of wafers of different materials (such as Si, SiC, GaN), serving as prior knowledge for physical guidance; and the residual sequence buffer updates and stores alignment deviation values within the past 50 sampling periods in real time using a first-in-first-out (FIFO) protocol, used to support the temporal prediction of the deep learning model.
[0023] To eliminate noise interference to the thermosensitive signal caused by high-frequency mechanical vibrations generated during high-speed movement or vacuum pump adsorption of the wafer stage, this invention dynamically adjusts the window size based on the temperature change rate and prediction error during the determination of the neural network's time window stage. The window adjustment equation is described as follows: ; In the formula, for The adaptive time window length at each moment. and For adjustment coefficients, This represents the absolute value of the current rate of temperature change. This represents the standard deviation of recent prediction errors. Through this dynamic windowing technique, the system can extract smooth features reflecting thermal hysteresis effects from the raw signal, which is often accompanied by noise. For example, during rapid heating, the window automatically widens to capture historical trends over a longer span, thus preventing overshoot or hysteresis in the compensation execution module.
[0024] The central computing module adopts a hybrid computing platform based on high-performance FPGA and multi-core DSP, and has a built-in deep learning model guided by physical mechanisms; the deep learning model has a built-in thermoelasticity benchmark library, which stores the theoretical displacement mapping matrix of the wafer under different temperature gradients. The central computing module also includes a data fusion algorithm and a compensation parameter adjustment algorithm. The data fusion algorithm is used to fuse and analyze the temperature field features and visual residuals transmitted by the multi-source sensing module with the motion coordinates transmitted by the data preprocessing module to determine the compensation requirements of the current test condition. The compensation parameter adjustment algorithm is used to generate thermal deformation compensation commands (including displacement increments in the X, Y, and Z directions) based on the fusion analysis results. At the same time, it combines the calculation results of the physical compensation logic unit to generate drive control signals for the piezoelectric ceramic. The command calculation delay does not exceed 2ms to ensure the real-time performance of submicron compensation.
[0025] The compensation execution module is the terminal unit for achieving high-precision spatial adjustment, including a piezoelectric ceramic fine-tuning platform and a multi-axis motion controller. The piezoelectric ceramic fine-tuning platform adopts a six-degree-of-freedom parallel mechanism, and the ceramic stacked substrate is made of high-rigidity aluminum nitride material and is fastened between the wafer stage support and the chuck. The multi-axis motion controller includes a DSP processing core and a high-voltage drive circuit. The high-voltage drive circuit supports 20-bit DAC output and can achieve nanometer-level step adjustment of the wafer stage in the X and Y directions. In addition, the system is also equipped with a physical compensation logic unit; static deformation is calculated in real time through a central computing module. This serves as the initial bias for the neural network, thereby accelerating the convergence speed of the deep learning model.
[0026] In practice, the central computing module executes a highly differentiated prediction and compensation scheme based on the operating condition identification results of the multi-source sensing module.
[0027] Under drastic temperature changes, the difference in thermal expansion coefficients between the wafer material and the chuck material easily leads to nonlinear thermal hysteresis effects; in this situation, the central computing module adjusts in real time based on the absolute value of the temperature change rate; if If the temperature fluctuation exceeds a preset threshold (e.g., 0.5℃ / s), the central computing module proactively switches the routing weight to the fast-response loop unit mode, compensating for the excess. Follow the functional relationship below: ; In the formula, The total compensation amount is the final output. The theoretical expansion value calculated by the static formula. For the current high-dimensional feature vector, This is the deep learning prediction function. By increasing the frequency of forward inference in the model, the compensation accuracy is maintained within a preset submicron range.
[0028] Under steady-state testing conditions, the temperature field tends to be in equilibrium, and the alignment accuracy is mainly affected by minute drifts. The central computing module reads data from the residual sequence buffer in real time; when the standard deviation of the prediction error is detected to be lower than the preset tolerance (e.g., 0.1 μm), the system reduces the number of active nodes in the hidden layer and automatically switches to the slow-response recurrent unit mode. This effectively reduces computational power consumption, and at the same time, by reducing the learning rate, ensures extremely smooth output of the compensation amount, avoiding probe scratches caused by model jitter.
[0029] Under long-period drift conditions, i.e., when the number of consecutive probe insertions exceeds a preset threshold (e.g., 5000 times), the central computing module is activated via a coordinate feedback encoder. When it is determined that the current optical axis has accumulated deviation due to slow deformation of the structural components caused by heat, the piezoelectric ceramic fine-tuning platform is driven to perform first-order linear compensation, achieving real-time offsetting of the optical axis's thermal drift on the test trajectory. Simultaneously, if the prediction variance detected by the visual verification module... If the safety threshold is exceeded, the system will automatically limit the maximum compensation range to prevent mechanical interference under uncertain operating conditions.
[0030] The visual verification module includes a high-resolution CCD camera and a feature matching unit; the high-resolution CCD camera is positioned above the probe station to capture the actual spatial coordinates of the marker points on the wafer surface; the feature matching unit uses a convolution operator to perform sub-pixel level edge detection on the acquired image. When the actual alignment residual is detected to exceed the preset dead zone range (e.g., 0.2μm), the visual verification module immediately sends a high-priority interrupt request to the central computing module. The central computing module executes an adaptive fine-tuning strategy: first, it uses the actual deviation value as the true prediction value to calculate the loss function of the current model, and only performs backpropagation updates for the currently activated learning weights; when the prediction accuracy recovers to the preset range, it stops parameter optimization to ensure the convergence and real-time performance of the system control loop.
[0031] The human-computer interaction module includes a graphical workstation and a model monitoring interface; the graphical workstation renders the thermal drift compensation curve, system temperature field cloud map, and the extension and retraction displacement of the piezoelectric ceramic fine-tuning platform actuator in real time; the model monitoring interface is used to set the scaling factor of the adaptive time window. and In addition, the human-machine interaction module has a built-in predictive confidence assessment program that monitors the heartbeat messages of each module on the high-speed industrial Ethernet bus in real time; once an encoder feedback abnormality or piezoelectric ceramic fine-tuning platform driver overcurrent is detected, the system immediately pops up an early warning window and executes emergency pin retraction protection logic.
[0032] At the hardware level, the system is also equipped with a power stabilization unit and a thermal isolation structure. The power stabilization unit is connected to the high-voltage purified power supply of the probe station and uses a multi-stage precision voltage regulation circuit to output an ultra-low ripple driving voltage. The unit is equipped with a transient suppression array for the high-voltage driving circuit, which can effectively absorb the reverse electromotive force generated by the piezoelectric ceramic fine-tuning platform during rapid response.
[0033] The physical protection structure fully considers the influence of the temperature field inside the test chamber and optimizes the structure as follows: the shell of the piezoelectric ceramic fine-tuning platform is made of Invar alloy and the surface is treated with a special heat-reflective coating; all sensor leads are made of polytetrafluoroethylene shielded wire; the whole machine has an anti-electromagnetic interference design to ensure that the weak internal temperature sensing signal can still be transmitted stably in a strong electrostatic test environment.
[0034] The technical effects of the present invention will be quantitatively demonstrated through specific embodiments below.
[0035] Comparison Case 1: A static linear compensation system based on the thermal expansion coefficient formula is used; the control method is offline calculation with no real-time sensor feedback; the actuator is a common ball screw drive system.
[0036] Comparison Case 2: A conventional lookup table compensation system is used; the control core is a PLC; the sensor is only a single-point thermometer; there is no deep learning prediction and visual online verification mechanism.
[0037] Experimental environment settings: A 12-inch wafer testing process was selected. The ambient temperature was rapidly increased from 25℃ to 125℃ and maintained, then cooled to -40℃. Recorded metrics included: average alignment residual (μm), first-pass success rate, system response time (ms), and UPH (upper throughput per hour) loss rate. Data comparison results are shown in Table 1 below. Table 1. Comparison of performance test data between the embodiments of the present invention and the comparative examples. Analysis of the data in Table 1 shows that this embodiment significantly outperforms the comparative case under extreme temperature differences. At a high temperature of 125℃, thanks to its deep learning-based nonlinear prediction, the alignment residual is only 0.35μm, far lower than the 2.8μm in comparative case 1. During rapid temperature changes, thanks to the adaptive time window guided by a physical mechanism, this embodiment effectively suppresses thermal hysteresis, achieving a thermal hysteresis elimination rate of 97.4%. Furthermore, by reducing frequent visual downtime for realignment, UPH is improved by 18.5% compared to the traditional solution, demonstrating extremely high production efficiency.
[0038] In practical engineering applications, system initialization calibration is equally crucial. After the system hardware installation is complete, the central computing module automatically enters self-learning calibration mode. First, it establishes the mechanical origin of the wafer stage through a coordinate feedback encoder. Then, it drives the chuck to perform stepped heating, collecting static pose changes at each temperature point to establish an initial deformation database. All weight parameters are stored in non-volatile memory, ensuring that the system maintains extremely high prediction accuracy each time it starts up.
[0039] To address the needs of high-frequency continuous shooting tests, this invention further optimizes algorithm power consumption. The central computing module monitors the FPGA core temperature in real time; when the core temperature exceeds 75°C due to frequent inference, the system initiates a load reduction strategy to dynamically reduce the feature extraction frequency of inactive neurons; simultaneously, the physical protection structure of the compensation execution module is equipped with a micro-circulation air duct to achieve rapid heat dissipation of electronic components without affecting the temperature field uniformity, ensuring the long-term reliability of the system.
[0040] In summary, the intelligent temperature compensation prediction system for probe stations provided by this invention, through high-precision multi-source sensor layout, physically guided variable structure deep learning algorithm, and nanometer-level piezoelectric actuator, successfully solves the urgent need for high-precision alignment in probe testing under dynamic and drastic temperature changes. It not only significantly improves the test yield, but also ensures the robustness of the equipment during long-term operation through online adaptive evolution logic, and has extremely high industrial application value and market promotion prospects.
[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A temperature compensation intelligent prediction system for a probe station, characterized in that, The system includes a multi-source sensing module, a data preprocessing module, a central computing module, a compensation execution module, a visual verification module, and a human-computer interaction module. The multi-source sensing module is used to perform multimodal digital reconstruction of the physical temperature field of the wafer testing environment and generate multidimensional feature vectors. The data preprocessing module is used to monitor the motion coordinates and thermoelastic mechanical parameters of the wafer stage in real time. The central computing module has a built-in deep learning model guided by physical mechanisms, which is used to predict and generate thermal deformation compensation instructions at a specific time in the future based on multi-dimensional feature vectors and thermoelastic mechanical benchmarks. The compensation execution module includes a piezoelectric ceramic fine-tuning platform and a multi-axis motion controller, which are used to dynamically adjust the spatial relative displacement of the wafer stage or probe card according to the thermal deformation compensation command. The visual verification module is used to collect the actual residual of the wafer alignment mark points during the test interval, and after identifying the deviation threshold, it sends a weight fine-tuning request to the central computing module to execute an online adaptive update strategy. The human-computer interaction module renders and displays the thermal drift compensation curve and system status in real time through a graphical interface, provides a model parameter configuration interface, and has a built-in model prediction confidence evaluation program.
2. The intelligent prediction system for temperature compensation of a probe station according to claim 1, characterized in that, The multi-source sensing module includes a center temperature sensor, an edge sensor group, and a probe card temperature sensor. The center temperature sensor is embedded at the geometric center of the wafer chuck and is used to collect real-time temperature data at the center point of the wafer. The edge sensor group is distributed in four quadrants on the edge of the wafer chuck to monitor the temperature gradient difference in the radial direction of the wafer, thus compensating for the sensing defects of single-point sampling under non-uniform temperature fields. The probe card temperature sensor is located on the back of the probe card and uses an integrated digital sensing chip. Its surface is covered with a high thermal conductivity insulating encapsulation layer. It is used to collect local hot spot information of the test environment and distinguish different test stages such as preheating, steady state and cooling. During the feature extraction process, the central computing module receives the multi-point temperature sequence and first performs physical baseline calculation; then it calls the variable structure recurrent neural network to perform nonlinear feature extraction; then it performs spatial mapping and fusion of the historical alignment residuals to generate a multidimensional input vector.
3. The intelligent prediction system for temperature compensation of a probe station according to claim 2, characterized in that, The data preprocessing module includes a temperature change rate extractor, a coordinate feedback encoder, a thermal expansion parameter memory, and a residual sequence buffer. The temperature change rate extractor calculates the current time in real time using differential operations. parameter; The coordinate feedback encoder monitors the current absolute coordinates of the wafer stage in real time through the grating ruler signal, which is used to determine the stroke redundancy of the compensation actuator. The thermal expansion parameter memory stores the thermal expansion coefficients and geometric dimensions of wafers made of different materials; The residual sequence buffer updates the alignment deviation value at historical moments in real time using a first-in-first-out protocol.
4. The intelligent prediction system for temperature compensation of a probe station according to claim 3, characterized in that, During the time window determination stage of the neural network, the data preprocessing module dynamically adjusts the window size based on the temperature change rate and prediction error. The window adjustment equation is as follows: ； In the formula, for The adaptive time window length at each moment. and For adjustment coefficients, This represents the absolute value of the current rate of temperature change. This represents the standard deviation of recent prediction errors.
5. The intelligent prediction system for temperature compensation of a probe station according to claim 4, characterized in that, The compensation execution module includes: The piezoelectric ceramic fine-tuning platform adopts a six-axis parallel structure, and the ceramic stacked substrate is made of high-rigidity aluminum nitride material and is fastened to the wafer stage support. The multi-axis motion controller includes a DSP processing core and a high-voltage drive circuit; wherein, the high-voltage drive circuit supports 20-bit DAC output, enabling nanometer-level step adjustment of the wafer stage in the X and Y directions; In addition, the compensation execution module is also configured with a physical compensation logic unit; the physical compensation logic unit is located inside the central computing module, and calculates static deformation variables as input to the initial bias term of the neural network.
6. The intelligent prediction system for temperature compensation of a probe station according to claim 5, characterized in that, The central computing module executes differentiated prediction and compensation logic based on the operating condition perception results as follows: Under drastic temperature changes, if the absolute value of the temperature change rate is greater than the preset temperature threshold, the central computing module actively increases the length of the adaptive time window, switches the routing weight to the fast response loop unit mode, and increases the frequency of model forward inference; if the absolute value of the temperature change rate is less than the preset temperature threshold, the routing weight is switched to the slow response loop unit mode. Under steady-state testing conditions, the data from the residual sequence buffer is read in real time; when the standard deviation of the prediction error is detected to be lower than the preset tolerance, the number of active nodes in the hidden layer is reduced and the learning rate is lowered. Under long-period drift conditions, the data from the coordinate feedback encoder is received in real time. When the number of consecutive probe insertions exceeds the preset threshold, the piezoelectric ceramic fine-tuning platform is driven to perform first-order linear compensation of the cumulative residual, thereby realizing real-time offsetting of thermal drift of the optical axis to the test trajectory. In addition, based on the prediction variance detected by the visual verification module, if it is determined that the current working condition is uncertain, the prediction variance will be output as the compensation range for safety redundancy limits.
7. The intelligent prediction system for temperature compensation of a probe station according to claim 6, characterized in that, The visual verification module includes a CCD camera and a feature matching unit; The CCD camera is used to capture the actual spatial coordinates of the marker points on the wafer surface; The feature matching unit then uses a convolution operator to perform sub-pixel level edge detection on the acquired image; When the actual alignment residual is detected to exceed the preset dead zone range, a high-priority interrupt request is immediately sent to the central computing module, and an adaptive fine-tuning strategy is executed. The adaptive fine-tuning strategy is as follows: first, the actual deviation value is used as the true prediction value to calculate the loss function of the current model, and backpropagation is performed only for the currently activated learning weights; when the prediction accuracy recovers to the preset range, parameter optimization is stopped to ensure the convergence and real-time performance of the system control loop.
8. The intelligent prediction system for temperature compensation of a probe station according to claim 5, characterized in that, The compensation execution module is surrounded by a physical protection structure, which includes a shell made of Invar alloy, and the surface of the shell is provided with a heat-reflective coating.
9. The intelligent prediction system for temperature compensation of a probe station according to claim 1, characterized in that, The system is also equipped with a power stabilization unit, which is connected to the high-voltage purified power supply of the probe station. The power stabilization unit uses a multi-stage precision voltage regulation circuit and transient suppression array to absorb the reverse electromotive force generated by the piezoelectric ceramic fine-tuning platform during rapid response.