Automated alignment coupling method and curing apparatus

Abstract

The application provides an automatic alignment coupling method and curing equipment, relates to the field of optoelectronic technology, and the method comprises the following steps: acquiring a multimodal visual image representing a coupling interface between a fiber array end face and a silicon optical chip, performing spatial registration based on the image and generating registration data; controlling a multi-degree-of-freedom micro-displacement platform to adjust the pose of the fiber array until a preset alignment condition is met; in response to the alignment condition being met, driving a crimping mechanism to perform crimping and acquiring online detection data in real time; judging the coupling quality according to the online detection data, and if the coupling quality meets the requirements, emitting ultraviolet light to form a connecting body, otherwise releasing the fiber array and re-registering; the method improves the precision and stability of automatic alignment and improves the production yield.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optoelectronic technology, in particular, to an automatic alignment coupling method and a curing device. BACKGROUND

[0002] Silicon optical chips have become an important research direction in the field of modern optoelectronics due to their advantages in high-speed data transmission, integrated density and cost control. However, high-precision automatic alignment coupling between silicon optical chips and fiber arrays is still one of the important process problems that need to be solved in current mass production. In the existing technology of automatic alignment coupling between silicon optical chips and fiber arrays, a single waveband visual imaging system is usually used to identify the pose of the fiber end face and the chip surface. However, this method can only obtain surface or shallow structure features during imaging, and it is difficult to effectively obtain multi-dimensional information of buried waveguides inside the chip and surface structures at the same time, thereby limiting the spatial pose registration accuracy and making it difficult to meet the increasing requirements of optical communication coupling precision.

[0003] In addition, the existing multi-degree-of-freedom displacement control platform usually faces problems such as mechanical coupling between platform axes, limited control instruction resolution, and error accumulation caused by environmental factors when performing high-precision adjustment tasks, which can easily lead to misjudgment or inaccurate convergence of the control loop, reducing the long-term stability and reliability in batch automatic production.

[0004] Therefore, the existing automatic coupling alignment technology of silicon optical chips needs to be further improved to solve the above technical problems, so as to improve the coupling alignment precision, reliability and production efficiency under batch production conditions. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides an automatic alignment coupling method and a curing device.

[0006] In a first aspect, the present application provides an automatic alignment coupling method, comprising:

[0007] Obtaining a multi-modal visual image for characterizing the coupling interface between the end face of the fiber array and the silicon optical chip;

[0008] Based on the multi-modal visual image, performing spatial registration to generate first registration data;

[0009] According to the first registration data, controlling a multi-degree-of-freedom micro-displacement platform to adjust the pose of the fiber array until a pose difference value is less than a first preset threshold, and outputting an alignment completion signal;

[0010] In response to receiving the alignment completion signal, a crimping mechanism is driven to crimp the fiber array to a coupling interface of the silicon optical chip at a preset crimping speed, and online detection data in a crimped state is acquired, the online detection data including an alignment error;

[0011] In response to the alignment error being less than a second preset threshold, a UV curing module is synchronously controlled to emit UV light to the coupling interface to cure the pre-positioned optical glue and form a cured connector, and a coupling completion instruction is output.

[0012] In response to the alignment error being greater than or equal to the second preset threshold, the crimping mechanism is controlled to release the fiber array and return to perform the spatial registration.

[0013] As an optional implementation, the generating first registration data includes:

[0014] Obtaining layout information of a metal interconnection layer of the silicon optical chip, and generating a mask corresponding to the metal interconnection layer;

[0015] Edge filtering is performed on the multi-modal visual image using the mask to remove edge information located in a metal interconnection layer region;

[0016] Based on the multi-modal visual image after edge filtering, a coupling reference feature point is extracted, and a coordinate transformation relationship between an end face of the fiber array and the coupling interface of the silicon optical chip is solved to generate the first registration data.

[0017] As an optional implementation, the removing edge information located in the metal interconnection layer region includes:

[0018] According to a near-infrared channel of the multi-modal visual image, a first edge set is extracted;

[0019] According to a visible light channel of the multi-modal visual image, a second edge set is extracted;

[0020] Consistency matching is performed on the first edge set and the second edge set to generate a common edge set;

[0021] Edge pixels not belonging to the common edge set are deleted within a metal interconnection layer region defined by the mask, and a multi-modal visual image after edge filtering is output.

[0022] As an optional implementation, the generating the first registration data includes:

[0023] At a first acquisition time, a first reference feature point set is extracted from the multi-modal visual image, and an initial coordinate transformation relationship is calculated;

[0024] At a second acquisition time that is separated from the first acquisition time by a preset synchronization delay, the second set of reference feature points is extracted again from the multimodal visual image, and the global tilt offset is calculated.

[0025] By integrating the initial coordinate transformation relationship with the global tilt offset, a time-compensated coordinate transformation relationship is generated, and the first registration data is output.

[0026] As an optional implementation, the calculation of the global tilt offset includes:

[0027] The two end feature points with channel number 1 and channel number M are determined sequentially according to the channel number, where M is the total number of channels in the fiber array.

[0028] Obtain the coordinate data of the two end feature points in a preset three-dimensional Cartesian coordinate system, and calculate the first pitch angle of the fiber array relative to the XZ plane based on the coordinate data using a least squares linear fitting algorithm.

[0029] In the second set of reference feature points, two corresponding end feature points with the same channel number are determined as channel number 1 and channel number M. The coordinate data of the corresponding end feature points in the three-dimensional Cartesian coordinate system are obtained, and the second pitch angle of the fiber array relative to the XZ plane is calculated by the least squares straight line fitting algorithm.

[0030] Calculate the absolute difference between the first pitch angle and the second pitch angle, and set this absolute difference as the global tilt offset.

[0031] As an optional implementation, before fusing the initial coordinate transformation relationship with the global tilt offset, the method further includes:

[0032] At least two temperature sensors are placed at opposite corners of the silicon photonics chip substrate, and the real-time temperature data output by the temperature sensors is periodically collected.

[0033] The real-time temperature data is input into a preset thermal expansion model to calculate the thermal expansion prediction vector corresponding to the geometric center of the chip. The thermal expansion model is established based on the linear expansion coefficient of the chip packaging material and the chip-fixture assembly dimensions.

[0034] The thermal expansion prediction vector is vector-superimposed with the translation component of the initial coordinate transformation relationship and the rotation component of the global tilt offset to generate a thermally compensated coordinate transformation relationship.

[0035] As an optional implementation, the output alignment completion signal includes:

[0036] Using the first registration data as input, a first pose adjustment command is generated according to a preset gain coefficient, and the first pose adjustment command is sent to the multi-degree-of-freedom micro-displacement platform.

[0037] The feedback displacement of each drive axis of the multi-degree-of-freedom micro-displacement platform is monitored in real time. If the amplitude of the coupled displacement between adjacent drive axes exceeds the preset inter-axis coupling threshold, the gain coefficient is reduced and a second pose adjustment command is generated.

[0038] After executing the second pose adjustment command, the pose difference is monitored again; if the pose difference falls within the preset threshold window and the pose difference of all channels meets the channel consistency judgment condition, the alignment completion signal is output.

[0039] As an optional implementation, the output alignment completion signal further includes:

[0040] After the pose difference falls into the preset threshold window, a periodic micro-amplitude jitter command with an amplitude not greater than the minimum quantization step of the capacitive displacement sensor is injected into the multi-degree-of-freedom micro-displacement platform, and the average value of the feedback displacement of the capacitive displacement sensor is calculated within N consecutive sampling periods.

[0041] If the mean calculation result still falls within the preset threshold window, the alignment completion signal is maintained; if the mean calculation result exceeds the preset threshold window, the gain coefficient of the multi-degree-of-freedom micro-displacement platform is iteratively adjusted according to the second pose adjustment command.

[0042] As an optional implementation, performing consistency matching on the first edge set and the second edge set to generate a common edge set includes:

[0043] Based on the periodic channel arrangement of the fiber array in the V-groove fixture, V-groove edge reference marks corresponding to multiple adjacent channels are extracted.

[0044] Using the reference markers and pre-stored channel spacing information, the local distortion center offset of the image region where each channel is located is determined;

[0045] For each channel region, the corresponding local distortion correction is applied to map the pixel coordinates of the first edge set and the second edge set to a unified correction coordinate system;

[0046] Under the unified calibration coordinate system, the first edge set and the second edge set are matched one-to-one according to the spatial nearest neighbor or feature descriptor similarity to generate the common edge set.

[0047] Secondly, this application provides a curing device, comprising:

[0048] The acquisition module acquires multimodal visual images to characterize the coupling interface between the fiber array end face and the silicon photonic chip;

[0049] The registration module performs spatial registration based on the multimodal visual image to generate first registration data;

[0050] The first control module controls the multi-degree-of-freedom micro-displacement platform to adjust the pose of the fiber array according to the first registration data until the pose difference is detected to be less than the first preset threshold, and outputs an alignment completion signal.

[0051] The second control module, in response to receiving the alignment completion signal, drives the crimping mechanism to crimp the fiber array to the coupling interface of the silicon photonic chip at a preset crimping speed, and acquires online detection data under the crimping state, the online detection data including alignment error;

[0052] The detection module, in response to the alignment error being less than the second preset threshold, synchronously controls the ultraviolet curing module to emit ultraviolet light toward the coupling interface to cure the pre-placed optical adhesive and form a cured connector, and outputs a coupling completion command; in response to the alignment error being greater than or equal to the second preset threshold, it controls the crimping mechanism to release the fiber array and return to perform the spatial registration.

[0053] Compared with existing technologies, the automatic alignment coupling method of this application, by acquiring multimodal visual images and performing spatial registration based on multispectral information fusion, can effectively overcome the problem of insufficient information in traditional single-band imaging, thereby helping to improve the accuracy and reliability of alignment between silicon photonic chips and fiber arrays. Furthermore, this application utilizes a closed-loop feedback control mechanism of a multi-degree-of-freedom micro-displacement platform to ensure timely convergence of the pose adjustment process, reducing the adverse effects of error accumulation and inter-axis mechanical coupling, effectively improving the accuracy and stability of the alignment process. Moreover, this application employs a real-time online detection method during the pressing process, which can dynamically judge the coupling quality and promptly detect and correct alignment deviations, avoiding scrap problems caused by irreversible curing after pressing, thereby effectively reducing production risks and improving the yield and process stability of mass production. Attached Figure Description

[0054] Figure 1 A flowchart of the automatic alignment coupling method provided in this application;

[0055] Figure 2 A flowchart of a method for generating first registration data provided in this application;

[0056] Figure 3 A flowchart of a method for removing edge information located in the region of the metal interconnect layer provided in this application;

[0057] Figure 4 This is a schematic diagram of the curing equipment provided in this application. Detailed Implementation

[0058] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0059] Reference Figure 1 The diagram shows a flowchart of the automatic alignment coupling method provided in this application. The method includes steps S101 to S105, wherein:

[0060] S101: Acquire multimodal visual images to characterize the coupling interface between the fiber array end face and the silicon photonic chip;

[0061] S102: Based on the multimodal visual image, perform spatial registration to generate first registration data;

[0062] S103: Based on the first registration data, control the multi-degree-of-freedom micro-displacement platform to adjust the pose of the fiber array until the pose difference is detected to be less than the first preset threshold, and output the alignment completion signal.

[0063] S104: In response to receiving the alignment completion signal, drive the crimping mechanism to crimp the fiber array to the coupling interface of the silicon photonic chip at a preset crimping speed, and acquire online detection data under the crimping state, the online detection data including alignment error;

[0064] S105: In response to the alignment error being less than the second preset threshold, the UV curing module is synchronously controlled to emit UV light to the coupling interface to cure the pre-set optical adhesive and form a cured connector, and a coupling completion command is output; in response to the alignment error being greater than or equal to the second preset threshold, the crimping mechanism is controlled to release the fiber array and return to perform the spatial registration.

[0065] The system for performing the method described in this application may include silicon photonics (SiP), fiber array (FA), multi-degree-of-freedom micro-displacement platform, crimping mechanism, multimodal vision imaging system, ultraviolet curing module, and central controller.

[0066] For example, the multimodal vision imaging system is typically located above the coupling interface to acquire alignment images; the fiber array can be fixed on a multi-degree-of-freedom micro-displacement platform, which is typically used to drive the fiber array to achieve precise alignment relative to the silicon photonic chip; the crimping mechanism and the UV curing module can be arranged coaxially or parallel to each other to perform crimping and curing operations after alignment is completed.

[0067] The multi-degree-of-freedom micro-displacement platform can be based on a six-degree-of-freedom parallel mechanism (Hexapod Platform) driven by piezoelectric ceramics, with a displacement resolution of, for example, up to 1 nanometer; the multimodal vision imaging system can, for example, include a visible light camera and a near-infrared camera arranged coaxially, and share an optical path through a beam splitter; the pressing mechanism can, exemplarily, employ a voice coil motor, with a precision typically reaching the sub-micrometer level; the ultraviolet curing module can, for example, employ an ultraviolet light emitting diode (UV-LED) light source or a laser, with a typical wavelength of 365 nm.

[0068] Regarding the above S101:

[0069] Multimodal visual images refer to a set of complementary images acquired simultaneously at the same coupling interface under different spectral channels. The visible light channel highlights the metal markings on the surface of the silicon photonic chip and the geometric contours of the fiber end face, while the near-infrared channel can penetrate the silicon substrate to show the position of the waveguide and fiber core inside the chip, avoiding the limitation of only being able to select one and discard the others when using traditional single-band imaging.

[0070] In practical implementation, silicon materials exhibit significant light reflection in the visible light region, resulting in extremely high contrast between the surface metallic structure and the substrate. However, in the near-infrared region with wavelengths greater than approximately 1100 nm, silicon's light absorption coefficient decreases significantly, allowing imaging light to penetrate silicon substrates hundreds of micrometers thick. This enables the visualization of the embedded waveguide and the core outline of the mating optical fiber. By spatially synchronizing and registering the two-channel images, the algorithm can simultaneously locate surface feature points and the internal optical centerline in a unified coordinate system, thereby establishing a complete six-degree-of-freedom relative pose model of the "chip-fiber".

[0071] For example, the silicon photonics chip is first fixed on a vacuum adsorption base, and the fiber array is mounted on a piezoelectrically driven six-degree-of-freedom platform. A multimodal imaging device is positioned directly above the interface and simultaneously distributes reflected light to both cameras via a coaxial beam-splitting prism. The visible light channel uses a ring-shaped white LED for diffuse illumination, with exposure time and aperture coefficient adaptively adjusted according to ambient brightness to ensure the metal markings are not overexposed. The near-infrared channel uses a 1310nm line light source for coaxial illumination, with exposure time automatically searching for a region with high signal-to-noise ratio based on the silicon substrate thickness and waveguide reflection intensity. The system can simultaneously acquire two images with a single trigger, then map them to a unified three-dimensional reference frame using a calibration matrix and output a feature point coordinate file, providing input for the next registration algorithm.

[0072] Furthermore, if the production line needs to adjust parameters due to changes in chip specifications, light source power, or ambient light, it can be quickly recalibrated using preset automatic exposure-automatic gain routines without altering the imaging hardware architecture.

[0073] In this way, the acquired multimodal visual images provide a reference benchmark that includes both surface and internal information for subsequent spatial registration, which helps to reduce alignment drift caused by the lack of single-band features and improves the robustness of pose solving. At the same time, synchronous acquisition avoids mechanical errors introduced by multiple workpiece movements, which makes a positive contribution to improving overall coupling accuracy and shortening the cycle time of the workstation.

[0074] Regarding S102 above:

[0075] After acquiring multimodal visual images, the system enters the spatial registration stage. In this application, spatial registration refers to mapping the surface feature points and internal optical feature points extracted from the visible light and near-infrared two-channel images to the same three-dimensional reference frame, and further calculating the relative pose relationship between the fiber array end face and the silicon photonic chip coupling interface. This pose relationship is output in the form of rotation and translation, and is denoted as the first registration data.

[0076] The principle can be understood as conformal matching of two images with different information dimensions: the visible light channel provides a planar positioning reference, and the near-infrared channel provides a depth or refractive layer positioning reference. After the two are transformed by the calibration matrix, they are superimposed in the computational space, thereby obtaining the relative pose of six degrees of freedom.

[0077] In practice, the registration process can be divided into two stages, from coarse to fine. In the coarse registration stage, the system prioritizes identifying the large metal cross mark and the fiber end face ring in the visible light image, while simultaneously extracting the brightness peak distribution center of the waveguide region in the near-infrared image. At this point, a fast matching strategy based on cross-correlation or normalized gray-level difference can initially align the reference coordinates of the chip and the fiber array in the camera's field of view, reducing their pose difference to the micrometer level and providing convergent initial values ​​for subsequent high-precision algorithms.

[0078] In the fine registration stage, based on the coarse registration results, subpixel-level centroid fitting is performed on each matched feature, and the rotation and translation are optimized simultaneously using the least mean square criterion. The rigid body transformation of the feature point coordinates can be solved and iteratively converged until the reprojection error of all features is less than one-tenth of the visual resolution.

[0079] It is worth emphasizing that the number of iterations mentioned above is not a fixed value, but is dynamically adjusted by the algorithm in real time to monitor the error convergence speed. When the image signal-to-noise ratio is low or the sample shows local reflection, the iteration termination condition can be automatically relaxed and the registration results can be cross-validated to prevent getting trapped in local optima.

[0080] For example, firstly, a pre-calibrated bi-target positioning matrix is ​​loaded, projecting the two-channel images onto a unified photographic base coordinate system. Then, a feature detector based on differential Gaussian and fast linear classifiers is run in each channel to filter out key points such as metal markers, waveguide bright lines, and fiber cores. Next, nearest neighbor search is performed on cross-channel features, and mismatches are eliminated using a random consistency check. Finally, the remaining features are fed into a gradient descent-based rigid body pose solver, which outputs the three-dimensional rotation and translation vectors of the fiber array relative to the silicon photonic chip, and records the confidence level of each feature point as a quality evaluation index. To balance processing speed and stability, the system can use an industrial control CPU or FPGA supporting parallel SIMD instruction sets to pipeline and accelerate the feature extraction and matching processes. In high-reflectivity or low-contrast scenarios, the controller automatically switches to a multi-scale template strategy, using template matching instead of point feature matching to ensure successful registration.

[0081] The generated first registration data retains both overall pose information and feature confidence levels and reprojection residuals, providing a quantifiable adjustment basis for the closed-loop control stage. Using this data, the multi-degree-of-freedom micro-displacement platform can place the coupling interface at the center of the capture window within a single closed loop, avoiding multiple backtracking operations due to error accumulation. Simultaneously, the residual distribution of the registration output can be used as a process health indicator in post-process analysis, helping to predict optical path contamination or device defects.

[0082] In summary, the above spatial registration steps not only improve the accuracy and robustness of pose solving, but also lay a repeatable, high-quality alignment foundation for the subsequent pressing and curing processes.

[0083] Regarding the above S103:

[0084] After obtaining the first registration data, the multi-degree-of-freedom micro-displacement platform is driven in real time to complete fine pose adjustment based on the rotation and translation vectors provided. The pose difference refers to the residual error in six degrees of freedom between the fiber array reference coordinate system and the registered target pose, taking into account both translation and rotation.

[0085] The platform control follows the approach of "coarse entry - fine scanning - convergence and locking": initially, the platform moves rapidly according to the full compensation value given by the registration, compressing the residual error to the sub-micron level; then, it enters the fine scanning stage, where the platform injects small sinusoidal jitter with an amplitude lower than the field resolution into each degree of freedom, evaluates the error gradient by real-time monitoring of optical path coupling power and feedback from capacitive displacement sensors, and continuously iterates to reduce the jitter amplitude; when the pose difference of all degrees of freedom falls into the pre-set first threshold window, the central controller determines that the alignment is complete and issues a locking command.

[0086] For example, the six-degree-of-freedom platform uses piezoelectric ceramic drive, and the minimum step size of a single axis is preferably controlled in the tens of nanometers range. The central controller reads the registration residual and sensor feedback every 2ms and runs parallel PID loops for the three translational and three rotational degrees of freedom respectively; in order to address possible inter-axis coupling, the proportional coefficient can be dynamically adjusted during operation to keep the coupled displacement of adjacent drive axes within the platform's resolvable step size.

[0087] Furthermore, to prevent the platform from jittering and locking up when approaching the target pose, the system will temporarily reduce the integral gain when the error converges to near the upper limit of the threshold, and at the same time enable the power maximization judgment. If the optical power curve maintains a stable peak value in multiple consecutive sampling periods, the position will be directly identified as the optimal coupling point.

[0088] Understandably, the first preset threshold is not a fixed constant, but rather a comprehensive setting considering the chip end-face roughness, the fiber array angle error budget, and the final coupling loss target. For example, in a common 4×4 channel array, the threshold can be set to an in-plane translation of no more than 100nm, an out-of-plane height difference of no more than 200nm, and pitch and yaw angles of less than 0.01°. In higher-density arrays or scenarios more sensitive to coupling loss, the same algorithm can be reused simply by adjusting the threshold. At the end of the process, the central controller records the locking command and the current pose in a log for verification in subsequent pressing and curing stages.

[0089] This closed-loop adjustment strategy achieves a balance between the platform's maximum travel and minimum step size, avoiding both prolonged fine-grained searches and overshoot due to large-scale compensation at once. Compared with traditional pure power scanning or single-camera closed-loop methods, it utilizes spatial registration residuals as an "absolute" reference and power changes as a "relative" optimization, enabling rapid determination and adaptive convergence of pose errors. This positively contributes to improving coupling success rate and shortening equipment cycle time.

[0090] Regarding S104 above:

[0091] Upon detecting the alignment completion signal, the central controller triggers the crimping mechanism to begin uniform downward pressing. The crimping mechanism can employ a programmable voice coil motor, coaxially mounted with a small-range force sensor. The platform continuously outputs a force-displacement curve during the pressing process. The crimping speed can be adaptively set according to the chip passivation layer thickness and the viscosity of the optical adhesive, typically increasing gradually within the range of a few micrometers per second to tens of micrometers per second to ensure that the optical adhesive spreads sufficiently without excessive overflow during pressure. Simultaneously, the system continuously acquires online detection data to determine whether new pose drift or coupling failure occurs during the crimping process.

[0092] For example, the controller injects a stable power probe light from one end of the fiber optic array, and a multi-channel photodiode array is configured at the chip output to acquire the coupling power curve in real time. During the crimping process, the optical power exhibits a typical "bell-shaped" change with axial displacement. The controller performs a moving average on the power curve and calculates its derivative. When the first derivative stably approaches zero within a preset window, and the power value enters and remains within the peak fluctuation range, the current position is recorded as the coupling peak point, and the corresponding residual pose difference is defined as the alignment error. If the error is consistently below the second threshold, the system considers the crimping successful.

[0093] Furthermore, to cover potential optical path obstruction or power drift scenarios, the system can also activate the lateral visual monitoring channel in parallel. The lateral camera observes the chip and array sidewalls through a long-working-distance microscope lens, attaching a pair of high-contrast reflective markers to the outer edges of each. During the pressing process, the software tracks the relative distance and angle changes of these markers in real time, mapping the results to out-of-plane tilt and height changes. If there is a significant deviation between the lateral measurement and optical power assessment, the controller will briefly stop pressing, triggering platform fine-tuning before resuming uniform pressing. The two measurement links are independent yet cross-verifiable, giving the system strong fault tolerance.

[0094] Within the normal process window, the crimping force-displacement curve eventually reaches a plateau where the force rises rapidly while the optical power remains high. At this point, the central controller determines that the alignment error meets the set parameters, immediately locks the voice coil motor position, and initiates the UV curing process. If a sharp drop in optical power or lateral pose exceeding the threshold is observed at any stage, the controller releases the voice coil motor to a safe height and calls the registration-fine-tuning subroutine to re-establish coupling. Through this online detection and dynamic backoff mechanism, the device can continuously achieve stable crimping quality without relying on operator subjective judgment and significantly reduce yield losses caused by instantaneous drift.

[0095] Regarding the above S105:

[0096] In this embodiment, the second preset threshold is used to determine whether the coupling quality after crimping is sufficient to enter the irreversible curing stage. It belongs to the same evaluation system as the aforementioned first preset threshold, but at a different level of function. The first preset threshold focuses on the geometric residuals during the platform's closed-loop adjustment process, emphasizing whether it has entered the operable range; while the second preset threshold comprehensively considers indicators such as geometric error, coupling optical power, and changes in crimping force, emphasizing whether it has met the final delivery requirements. The first is broader than the second, ensuring both adjustment efficiency and finished product quality.

[0097] In practice, the central controller continuously collects optical power, displacement, and force curves during the crimping process. When the alignment error calculated in real time stably falls within the second threshold window, the coupling state is determined to have reached the curing condition. The threshold itself is not a fixed constant and can be dynamically set with reference to the flatness of the chip end face, the flowability of the optical adhesive, and the target insertion loss budget. For example, in a four-channel array scenario, the in-plane residual displacement can be limited to the tens of nanometers level, and the coupling power fluctuation can be maintained within a small range. If the array density or optical power requirements are higher, the same control logic can be reused simply by tightening the window accordingly.

[0098] In practice, after the pressing reaches a steady state, the system first locks the position of the multi-degree-of-freedom platform and maintains a constant pressing force output by the voice coil motor. The UV curing module is then triggered by the central controller, and the UV beam irradiates the coupling interface along the coaxial path of the array-chip. The irradiation intensity and duration are automatically matched according to the optical adhesive curing curve. During the curing process, the optical power and displacement curves are still recorded in real time. If the curves remain stable, the curing process automatically outputs a coupling completion command after the preset time, and simultaneously submits the process log to the upper-level database. If the error is detected to rise to the upper limit of the second threshold at any time, the controller immediately turns off the UV light, releases the pressing force, separates the fiber array from the chip, and returns to the spatial registration subprocess to re-establish alignment.

[0099] By introducing stricter secondary thresholds and adopting a real-time rollback mechanism before curing, the system can avoid irreversible fixing on unqualified poses while ensuring efficiency. This helps to improve the yield and stability of mass production and reduce the cost and downtime risk caused by rework or scrap.

[0100] As an optional implementation method, see [link to implementation details]. Figure 2 The flowchart of a method for generating first registration data provided in this application includes steps S201 to S203, wherein:

[0101] S201: Obtain the layout information of the metal interconnect layer of the silicon photonics chip, and generate a mask corresponding to the metal interconnect layer;

[0102] S202: Use the mask to perform edge filtering on the multimodal visual image to remove edge information located in the metal interconnect layer region;

[0103] S203: Based on the multimodal visual image after edge filtering, extract the coupling reference feature points, solve the coordinate transformation relationship between the fiber array end face and the silicon photonic chip coupling interface, and generate the first registration data.

[0104] In practical applications, for silicon photonic chips with complex metal interconnect structures, images acquired in the visible light channel often contain numerous regular and intensely bright edges. These edges mainly originate from the reflected signals of the metal wiring layers, possessing good structural continuity and contrast. They are easily misidentified as valid features by traditional edge detection algorithms, leading to subsequent problems such as feature point mismatch, coordinate fitting drift, or unstable registration results. Especially under conditions where waveguides and fiber arrays have weak geometry and low signal-to-noise ratios, this interference may cause the registration results to deviate completely from the coupling center, thereby affecting the stability of the entire coupling link.

[0105] To mitigate the problem of spurious feature extraction caused by such metal interference, this application introduces a layout mask strategy. This involves introducing masking information consistent with the spatial distribution of the chip's metal interconnect layers into the visual processing chain. This information is used to filter edge responses in corresponding areas of the image, eliminating potential interference at its source. A layout mask refers to a binary mask image generated by spatially mapping and rasterizing the metal wiring layers exported from the chip's process design file (GDSII or OASIS format), creating an image that corresponds one-to-one with the image's field of view. In the mask image, the areas marked as masked will not participate in the subsequent image feature extraction process.

[0106] The principle is based on the spatial mapping between the layout and image coordinates. In specific implementation, firstly, during the system calibration stage, a two-dimensional affine transformation model between the chip design coordinate system and the visual image coordinate system is established using standard Fiducial markings (three-point method) on three silicon photonics chips. Taking the two metal layers "Metal3" and "Metal4" in the chip GDS file as examples, their polygonal boundaries can be exported as vector graphics using layout processing tools such as KLayout. This vector graphic is then transformed using the aforementioned affine matrix, and finally rasterized into a bitmap mask using the cv2.fillPoly() function in the OpenCV library. The bitmap resolution should be consistent with the actual acquired image, such as 2048×2048, to ensure pixel-by-pixel alignment.

[0107] In the visual image processing stage, the mask serves as a weighted filtering matrix before edge detection in the image convolution process. Taking Canny edge detection as an example, before calculating the image gradient, pixel-wise multiplication is performed using the mask to forcibly set the grayscale value corresponding to the shielded area in the mask to zero, or to set the gradient response of that area to infinitesimal. Then, the Sobel operator is used to calculate the image gradient, and finally, a double-threshold connection algorithm is used to generate the edge map. In this way, even if the metal interconnect layer has strong reflections, its edges will no longer participate in the feature point extraction and matching process in the visual image.

[0108] In a complete implementation process, after the system loads the mask image, it first performs masking on the multimodal image, and then runs predefined edge detection and feature extraction algorithms in the visible light and near-infrared images respectively. In this embodiment, a feature point detection algorithm based on LoG (Laplacian of Gaussian) weighted gradient can be used, and its implementation includes:

[0109] The image is first subjected to Gaussian filtering to remove noise, and then a second-order Laplacian operator is executed to extract edge centers, obtaining sub-pixel-level feature points. The feature point set then enters the matching module for spatial registration, calculates the transformation relationship between the fiber array and the chip, and outputs the first registration data.

[0110] In this way, interference from metal edges can be effectively avoided on chips with dense metal interconnect structures, improving the stability of feature extraction, especially in coupled structures with a limited number or uneven distribution of edges, thus maintaining registration accuracy. Furthermore, since the masking strategy is logically decoupled from the image processing module and does not depend on specific image content, it can be extended to various chip layouts or product scenarios with significant variations in metal layers. This helps reduce algorithm parameter tuning costs, shorten registration time, and improve the overall environmental adaptability and yield performance of the system.

[0111] As an optional implementation method, see [link to implementation details]. Figure 3 The flowchart of a method for removing edge information located in the metal interconnect layer region provided in this application includes steps S301 to S304, wherein:

[0112] S301: Extract the first edge set based on the near-infrared channel of the multimodal visual image;

[0113] S302: Extract the second edge set based on the visible light channel of the multimodal visual image;

[0114] S303: Perform consistency matching on the first edge set and the second edge set to generate a common edge set;

[0115] S304: Delete edge pixels that do not belong to the common edge set within the metal interconnect layer area defined by the mask, and output the multimodal visual image after edge filtering.

[0116] In real-world production environments, the metal interconnect layer layout design of silicon photonics chips often deviates slightly from the actual manufacturing process. A purely static masking scheme can only provide theoretical shielding based on the layout data, failing to address issues such as layout distortion, lighting variations, or metal layer shifts caused by changes in shooting angle in actual images. Therefore, relying solely on masks for edge shielding may miss some metal edges or even mistakenly delete real coupled edges. To further improve the accuracy of edge filtering, this application introduces a dynamic edge filtering method based on consistency matching between multimodal image channels.

[0117] Specifically, the consistency matching described in this application refers to independently extracting the edge sets from the multimodal visual images of the near-infrared and visible light channels, and then comparing the features such as the position, length, contour shape, and direction of the edges between the two channels to filter out the real edges that appear consistently in both channels and have the same spatial position, i.e., the common edge set. Edges that appear only in a single channel are considered to be false edges and are discarded.

[0118] The principle behind this method is that the metal interconnect layers of silicon photonics chips typically reflect light significantly only in the visible light band, while transmitting or reflecting light weakly in the near-infrared band. However, the waveguide structures or fiber end faces actually used for optical coupling alignment can form clear and positionally stable edge features in both the visible and near-infrared bands. Therefore, through cross-channel consistency matching, the differences between the image and the layout can be dynamically adapted during actual production, enhancing the robustness of edge filtering.

[0119] In the specific implementation, firstly, multimodal visual images of the coupling interface are acquired synchronously, specifically including one visible light channel image and one near-infrared channel image, ensuring that the two channel images have completed pixel-level registration. Then, edge detection is performed independently on each channel image to generate a corresponding edge set. In this embodiment, the exemplary edge detection algorithm is the Canny algorithm, which can be implemented using the cv2.Canny() function in the OpenCV library. Typical parameter settings include a low threshold of 50, a high threshold of 150, and edges with a connection pixel length greater than 5 pixels are retained to filter out noise and isolated points. Thus, a first edge set for the near-infrared channel and a second edge set for the visible light channel are obtained.

[0120] Then, consistency matching is performed on the first edge set and the second edge set to determine the common edge set that matches the spatial position between the two sets.

[0121] For example, step one: traverse each edge curve in the visible light channel and calculate its center point coordinates, length, circumscribed rectangle size, and principal direction angle;

[0122] Step 2: Perform the same feature calculations as in Step 1 for each edge curve in the near-infrared channel;

[0123] Step 3: Match the edge curve features of the visible light channel with the edge curve features of the near-infrared channel one by one. The specific matching criteria include, but are not limited to: the Euclidean distance between the center points of the two curves is less than a preset threshold, such as 5 pixels; the length ratio is between 0.8 and 1.2; the intersection-over-union (IoU) ratio of the circumscribed rectangle is greater than 0.6; and the difference in the principal direction angle is less than 10 degrees.

[0124] Step 4: For edge pairs that meet the above matching criteria, further call the shape similarity comparison function for precise confirmation, such as cv2.matchShapes(). If the return value is less than 0.2, the shape is considered similar, and the confirmed edge pairs are included in the common edge set.

[0125] Subsequently, based on the generated common edge set, the system performs edge filtering on the initial visible light and near-infrared channel edge images, that is, deleting pixels located within the metal interconnect layer region defined by the mask that do not belong to the common edge set.

[0126] For example, first, a mask image generated from the layout data is loaded to define the spatial region for edge filtering. Each edge pixel within the mask region is checked individually. If the edge pixel does not belong to the common edge set, it is removed from the original image by median filtering or by directly setting it to zero. Taking median filtering as an example, the cv2.medianBlur() function of the OpenCV library can be used for filtering; if zeroing is used, the image grayscale value at the edge pixel position is directly assigned to the background grayscale value.

[0127] For example, in a visible light image of a chip, a horizontal edge about 50 pixels long is detected, but no obvious edge or only 10 pixels long is detected at the corresponding position in the near-infrared image. After applying the above matching criteria, the two edges cannot meet the consistency condition. Therefore, it can be determined that the edge is a pseudo edge caused by the metal layer in the visible light channel and should be deleted.

[0128] On the other hand, if an edge appears in both channels simultaneously, with a positional deviation of no more than 3 pixels, a length ratio of approximately 0.95, and a shape matching score of less than 0.1, it is considered to be a true coupled structure edge and should be retained in the common edge set.

[0129] Ultimately, the image after consistency filtering contains fewer false feature edges and allows for more reliable extraction of true feature points for subsequent coupling reference. Based on this image, the system further performs feature extraction and spatial registration algorithms to generate the first registration data after dynamic edge filtering. Feature extraction can include, for example, sub-pixel center localization and edge contour fitting, while the spatial registration algorithm can be, for example, a rigid body transformation fitting algorithm based on minimum mean square error.

[0130] In this way, by using dynamic cross-channel consistency matching, false edge features under a single channel can be effectively identified and eliminated, making up for the problem of missed filtering or accidental deletion caused by the deviation between the layout and the actual image in static masks. In production scenarios with different chip batches and large fluctuations in layout accuracy, it can adaptively improve the stability of image processing and the reliability of spatial registration, and ultimately help improve the overall accuracy of chip-fiber array coupling and production line yield.

[0131] As an optional implementation, generating the first registration data includes:

[0132] At the first acquisition time, a first set of reference feature points is extracted from the multimodal visual image, and the initial coordinate transformation relationship is calculated;

[0133] At a second acquisition time that is separated from the first acquisition time by a preset synchronization delay, the second set of reference feature points is extracted again from the multimodal visual image, and the global tilt offset is calculated.

[0134] By integrating the initial coordinate transformation relationship with the global tilt offset, a time-compensated coordinate transformation relationship is generated, and the first registration data is output.

[0135] In actual manufacturing processes, the spatial alignment accuracy of silicon photonic chips and fiber arrays is often easily affected by environmental vibrations, temperature drift, or instantaneous equipment jitter in the production line. This means that the coordinate transformation relationship obtained at a certain acquisition moment may only represent the state at that instant, but cannot accurately characterize the continuous coupling pose change trend throughout the entire crimping process. To further reduce the impact of such temporal disturbances on the coupling process, this application proposes a coordinate transformation fusion method based on temporal compensation.

[0136] Specifically, the timing compensation described in this application involves extracting reference feature point sets at two adjacent acquisition times with a preset delay, and analyzing the displacement changes between these feature points to deduce the global tilt offset trend caused by environmental disturbances. This allows for real-time correction of the coordinate transformation relationship. Using this method, the system can dynamically predict and compensate for timing-related errors, enabling the generated first registration data to more accurately represent the continuous true alignment state, rather than being limited to a single static moment.

[0137] In practice, at the first acquisition moment, the system synchronously acquires multimodal visual images of the coupling interface and performs feature extraction algorithms on the images to obtain a stable set of first reference feature points.

[0138] For example, sub-pixel feature detection can be performed using the LoG (Laplacian of Gaussian) operator to clearly mark the cross-shaped metal markings on the chip surface, the circular features on the fiber array end face, and the brightness center feature point of the internal waveguide. Subsequently, the system uses the coordinates of these feature points to solve the initial coordinate transformation relationship between the chip and the fiber array, typically using a rigid body transformation fitting algorithm based on the least squares method, and outputs the initial coordinate transformation matrix.

[0139] At a second acquisition time, separated from the first acquisition time by a preset synchronization delay (e.g., 50 to 200 milliseconds) to match typical environmental disturbance cycles, the system again acquires multimodal visual images of the coupled interface and re-executes the same feature extraction process to generate a second set of reference feature points. To ensure the reliability of the timing compensation, the images at the second acquisition time should use the same optical system, camera settings, and exposure parameters as those at the first acquisition time.

[0140] Subsequently, based on the difference between the two sets of reference feature points, the system calculates the global tilt offset generated by the coupling interface within the delay interval.

[0141] For example, the system can perform accurate calculations using the following methods:

[0142] Step 1: Establish 3D spatial coordinate lists for the feature point sets acquired at the two time points, and use the ICP (Iterative Closest Point) algorithm for matching. In specific implementation, the pcl::IterativeClosestPoint function in the open-source point cloud library PCL can be used to perform rigid registration of the coordinate point cloud to extract the overall rotation and translation trend caused by environmental disturbances.

[0143] Step 2: During the execution of the ICP algorithm, the rigid transformation matrix between two point sets is solved iteratively until the sum of squared residuals reaches the convergence condition, for example, a convergence accuracy of... The output rigid transformation matrix at this point is the global tilt offset within that time interval. To ensure computational stability, the system can be set to a maximum of 100 iterations, and outlier filtering can be performed in each iteration, such as a filtering method based on the mean and standard deviation, to improve the robustness of the fit.

[0144] Step 3: The offset matrix determined by the ICP algorithm is specifically represented by a rotation matrix and a translation vector in three-dimensional space. The displacement and rotation trends it expresses are the global drift caused by the environment, which can be used to predict the trend of coupled pose changes in the short term.

[0145] Next, the system integrates the initial coordinate transformation relationship with the global tilt offset to generate a time-compensated coordinate transformation relationship. The specific implementation process is as follows:

[0146] First, the pose represented by the initial coordinate transformation matrix is ​​defined as the base state. Then, the offset matrix output by the ICP algorithm is used to compensate and fuse this base state, for example, by performing matrix multiplication and superposition, to predict the coupled pose trajectory in the near future. The system can further perform interpolation or filtering on the fused matrix, such as exponential smoothing filtering or Kalman filtering, to avoid excessive amplification of instantaneous errors and to smoothly output the predicted future pose state.

[0147] For example, at the first acquisition moment, the initial rotation angle between the chip and the fiber array is detected to be 0.005°, and the translational deviation is 80 nanometers. At the second moment, ICP matching indicates that after the delay time, the global rotation change is an additional 0.002°, and the translational change is 10 nanometers. Therefore, the fused, time-compensated coordinate transformation relationship is a rotation of 0.007° and a translation of 90 nanometers. The fused data is used to drive the pre-compensation action of the multi-degree-of-freedom micro-displacement platform, which can significantly improve the initial alignment efficiency and success rate in the subsequent coupling stage.

[0148] Thus, the coordinate transformation relationship after time-compensated processing is output as the first registration data in this embodiment. Compared with static measurement at a single moment, this fusion method provides a more robust prediction of environmental disturbances, helping the platform to prevent instantaneous alignment deviations caused by equipment vibration or thermal drift in advance, significantly reducing the accumulation of dynamic errors caused by environmental factors, thereby improving the overall coupling stability and accuracy of the production line, shortening the closed-loop time of a single alignment, and improving the overall yield during mass production.

[0149] As an optional implementation, the calculation of the global tilt offset includes:

[0150] The two end feature points with channel number 1 and channel number M are determined sequentially according to the channel number, where M is the total number of channels in the fiber array.

[0151] Obtain the coordinate data of the two end feature points in a preset three-dimensional Cartesian coordinate system, and calculate the first pitch angle of the fiber array relative to the XZ plane based on the coordinate data using a least squares linear fitting algorithm.

[0152] In the second set of reference feature points, two corresponding end feature points with the same channel number are determined as channel number 1 and channel number M. The coordinate data of the corresponding end feature points in the three-dimensional Cartesian coordinate system are obtained, and the second pitch angle of the fiber array relative to the XZ plane is calculated by the least squares straight line fitting algorithm.

[0153] Calculate the absolute difference between the first pitch angle and the second pitch angle, and set this absolute difference as the global tilt offset.

[0154] This application proposes a more efficient, faster, and more targeted implementation detail for the aforementioned timing compensation method. Although the full six-degree-of-freedom (6-DOF) calculation based on the iterative nearest-point algorithm can comprehensively and accurately obtain the rotation and translation amounts, in actual production scenarios, changes in the pitch angle usually have the greatest impact on the coupling performance between the silicon photonics chip and the fiber array, and are the main source of error causing coupling loss. In addition, the ICP algorithm requires complex iterative calculations for all feature points, which consumes a lot of computational resources and may be difficult to achieve real-time response on high-paced production lines. Therefore, this embodiment proposes a fast calculation method specifically for the pitch angle component to prioritize and accurately compensate for pitch angle changes.

[0155] The specific implementation steps are as follows:

[0156] First, at the initial acquisition time, feature points at both ends of the fiber array are accurately extracted using multimodal visual images, specifically the end feature points with channel number 1 (leftmost) and channel number M (rightmost). Subpixel-level precision LoG or Sobel operators are recommended for feature point extraction to ensure coordinate extraction accuracy.

[0157] Secondly, the obtained two-dimensional image coordinates of the end feature points are mapped to a preset three-dimensional Cartesian coordinate system through the camera calibration model to determine their three-dimensional spatial coordinate data. Subsequently, the least squares line fitting algorithm is used to calculate the spatial line passing through the two end feature points, and the initial pitch angle of the line relative to the XZ plane in the device coordinate system is calculated.

[0158] Next, at the second acquisition time, which is separated from the first acquisition time by a preset synchronization delay, the end feature points of channel number 1 and channel number M are extracted again in the same way, and the second pitch angle is calculated by the least squares linear fitting algorithm. The typical setting of the synchronization delay is 50 milliseconds to 200 milliseconds.

[0159] Next, by comparing the difference between the first and second pitch angles, a precise numerical value of the pitch angle change is obtained. This value directly reflects the overall tilt change of the fiber array caused by environmental disturbances or equipment vibrations during the delay period.

[0160] Finally, the system integrates the calculated pitch angle change value as a key component in the overall timing compensation process. Specifically, the calculated pitch angle change value is converted into a rotation transformation matrix around a specific axis (e.g., the X-axis), and then fused with the initial coordinate transformation matrix to efficiently and quickly compensate for the most critical pitch angle error.

[0161] In this way, by accurately calculating the pitch angle change of the fiber array, a lightweight and efficient implementation method is provided, which effectively reduces the computational load. It can be used as a quick replacement or preferred supplement to the complete ICP algorithm. It is particularly suitable for practical application scenarios that are extremely sensitive to pitch angle changes and require high-speed production, thereby improving overall production efficiency and coupling accuracy, and reducing the risk of coupling errors caused by environmental disturbances.

[0162] As an optional implementation, before fusing the initial coordinate transformation relationship with the global tilt offset, the method further includes:

[0163] At least two temperature sensors are placed at opposite corners of the silicon photonics chip substrate, and the real-time temperature data output by the temperature sensors is periodically collected.

[0164] The real-time temperature data is input into a preset thermal expansion model to calculate the thermal expansion prediction vector corresponding to the geometric center of the chip. The thermal expansion model is established based on the linear expansion coefficient of the chip packaging material and the chip-fixture assembly dimensions.

[0165] The thermal expansion prediction vector is vector-superimposed with the translation component of the initial coordinate transformation relationship and the rotation component of the global tilt offset to generate a thermally compensated coordinate transformation relationship.

[0166] Furthermore, this application introduces a mechanism for monitoring chip substrate temperature and predicting thermal expansion to effectively reduce coupling errors caused by changes in ambient temperature. In actual production environments, the difference in thermal expansion between the silicon photonics chip and the fixture material often causes minute displacements in the chip position. Especially under continuous production conditions, this thermal effect can gradually accumulate into a non-negligible source of error. Therefore, this application provides a specific implementation method for predicting and compensating for thermal expansion to further improve alignment and coupling accuracy.

[0167] In practice, firstly, at least two high-sensitivity temperature sensors are installed diagonally on the silicon photonics chip substrate. Thermistors or thermocouples are recommended for these sensors to ensure real-time and accurate temperature data acquisition. The specific installation locations can be the upper left and lower right corners of the chip substrate. The system periodically acquires the real-time temperature data output by the temperature sensors at a certain sampling period, for example, once per second.

[0168] Secondly, based on a pre-established thermal expansion model, the system processes the collected temperature data in real time and accurately calculates the predicted thermal expansion vector at the chip's geometric center. Specifically, the thermal expansion model is a mathematical model predetermined based on the linear expansion coefficients of the chip and fixture materials, as well as the actual chip-fixture assembly dimensions. Using this model, the potential thermal expansion displacement trends of the chip in the X, Y, and Z directions can be calculated based on the temperature changes at various measuring points on the base.

[0169] For example, the steps for constructing this thermal expansion model include: pre-calibrating the linear expansion coefficient of the chip fixture material and obtaining specific material property values ​​through experiments or data provided by the manufacturer. Subsequently, a spatial displacement function relationship is constructed based on the geometric dimensions of the chip fixture, that is, the predicted displacement vector of the chip center point is directly derived through the temperature difference.

[0170] Next, after obtaining the thermal expansion prediction vector, the system further fuses it with the previously calculated initial coordinate transformation relationship and global tilt offset. The specific fusion process includes two independent and explicit mathematical steps:

[0171] Step 1: Compensate for translation components. Add the translation components of the thermal expansion prediction vector to the initial coordinate transformation relationship to compensate for the translational drift of the chip's physical position caused by thermal effects.

[0172] Step 2: Compensate for the rotation component. Multiply the rotation component represented by the global tilt offset (represented as a rotation matrix or a matrix generated by a specific rotation angle) with the rotation component of the initial coordinate transformation relationship to compensate for the rotational drift of the device attitude.

[0173] For example, assuming the translation components of the initial coordinate transformation relationship are 10.0 nm (X-axis), 20.0 nm (Y-axis), and 5000.0 nm (Z-axis), and the rotation components are determined by the rotation matrix... This indicates that if real-time temperature monitoring shows a 2°C increase in the chip substrate temperature, based on the linear expansion coefficient of the fixture material and the chip size parameters, a predicted upward thermal expansion displacement vector of approximately 100 nanometers is calculated at the chip center along the Z-axis. Simultaneously, the timing compensation module calculates the rotation matrix. This represents the global tilt offset.

[0174] Subsequently, the system executes the specific fusion compensation process:

[0175] New translation component = (10.0, 20.0, 5000.0) + (0.0, 0.0, 100.0) = (10.0, 20.0, 5100.0) nanometers.

[0176] New rotation component = .

[0177] The translation vector and rotation matrix obtained through the above operations together constitute the final coordinate transformation relationship after thermal compensation and timing compensation, which is used to control the platform to achieve precise pre-compensation actions.

[0178] In this way, by introducing temperature monitoring and thermal expansion prediction models, spatial displacement errors caused by changes in ambient temperature are effectively compensated. Compared with traditional methods that rely solely on static or short-term timescale predictions, this application provides a dynamic, real-time thermal compensation strategy, which significantly improves alignment accuracy and stability under long-term operating conditions. It is particularly suitable for automated coupling production lines of silicon photonic chips and fiber arrays that are sensitive to thermal effects and have high precision requirements.

[0179] As an optional implementation, the output alignment completion signal includes:

[0180] Using the first registration data as input, a first pose adjustment command is generated according to a preset gain coefficient, and the first pose adjustment command is sent to the multi-degree-of-freedom micro-displacement platform.

[0181] The feedback displacement of each drive axis of the multi-degree-of-freedom micro-displacement platform is monitored in real time. If the amplitude of the coupled displacement between adjacent drive axes exceeds the preset inter-axis coupling threshold, the gain coefficient is reduced and a second pose adjustment command is generated.

[0182] After executing the second pose adjustment command, the pose difference is monitored again; if the pose difference falls within the preset threshold window and the pose difference of all channels meets the channel consistency judgment condition, the alignment completion signal is output.

[0183] Furthermore, this application proposes a multi-degree-of-freedom micro-displacement platform control strategy to effectively avoid the reduction in alignment accuracy or instability caused by excessive coupling displacement between platform drive shafts.

[0184] In practice, the system first generates initial pose adjustment commands based on the previously obtained registration data using a closed-loop PID control algorithm. Specifically, the gain coefficients represent the proportional gain, integral gain, and derivative gain parameters of the PID controller. These parameters are initially set high to ensure a rapid response and drive the fiber optic array to make initial adjustments towards the target position.

[0185] Secondly, during the execution of the initial pose adjustment command by the multi-degree-of-freedom micro-displacement platform, the system monitors the feedback displacement of each drive axis of the platform in real time. Inter-axis coupling displacement specifically refers to the unexpected displacement that occurs in other non-target axis directions due to the platform's structural characteristics when the system sends a displacement command to a certain drive axis. The system quantitatively calculates the inter-axis coupling ratio by comparing the difference between the commanded displacement and the actual feedback displacement. For example, when sending a displacement command of a certain amplitude to the X-axis, if an unexpected displacement change occurs in the Y-axis and its amplitude exceeds a preset threshold (e.g., 5%), it is determined that the inter-axis coupling exceeds the limit.

[0186] Subsequently, once an inter-axis coupling over-limit is detected, the system reduces the gain coefficient of the currently used PID controller, for example, by a fixed percentage, such as 20%. Then, the pose adjustment command is regenerated based on the updated gain parameters and sent again to the multi-degree-of-freedom micro-displacement platform to achieve more stable pose adjustment.

[0187] After the multi-degree-of-freedom micro-displacement platform executes the gain-adjusted pose command, the system again monitors the pose difference of the platform in real time, i.e., the residual translation and rotation errors between the fiber array and the target alignment position. In addition, the system simultaneously monitors the optical coupling power loss values ​​of all channels. By calculating the range of these loss values—that is, the difference between the maximum and minimum loss values—the system determines the consistency of the channels. For example, when the range is less than a preset power threshold (e.g., 0.5 dB), it is determined that the error distribution among the channels is uniform and no obvious anomalies have occurred.

[0188] Once the pose difference successfully falls within the preset threshold window and simultaneously meets the channel consistency condition, the central controller officially outputs an alignment completion signal to lock the current pose state for use in subsequent pressing and curing processes.

[0189] Through the above-mentioned multi-axis displacement monitoring and gain adaptive adjustment mechanism, this application effectively avoids mutual interference between platform drive axes caused by excessively high gain, and improves the stability and accuracy of automatic alignment adjustment. It is especially suitable for high-precision, low-fault-tolerance silicon photonic chip and fiber array automatic coupling application scenarios.

[0190] As an optional implementation, the output alignment completion signal further includes:

[0191] After the pose difference falls into the preset threshold window, a periodic micro-amplitude jitter command with an amplitude not greater than the minimum quantization step of the capacitive displacement sensor is injected into the multi-degree-of-freedom micro-displacement platform, and the average value of the feedback displacement of the capacitive displacement sensor is calculated within N consecutive sampling periods.

[0192] If the mean calculation result still falls within the preset threshold window, the alignment completion signal is maintained; if the mean calculation result exceeds the preset threshold window, the gain coefficient of the multi-degree-of-freedom micro-displacement platform is iteratively adjusted according to the second pose adjustment command.

[0193] Furthermore, this application proposes a precise and effective closed-loop stability verification strategy to ensure that the actual coupling position after pose adjustment can be stably maintained within the expected target range.

[0194] In practice, once the pose difference of the multi-degree-of-freedom micro-displacement platform successfully falls within a preset threshold window, the system continuously injects a set of micro-amplitude periodic jitter commands into the platform. The amplitude of these jitter commands is very small, typically set slightly below the minimum quantization step of the capacitive displacement sensor, with a typical value of around 0.8 nanometers.

[0195] Although the amplitude of a single jitter is smaller than the single measurement accuracy of the sensor, by continuously sampling over multiple cycles, the system can utilize the inherent random measurement noise of the sensor to achieve higher effective resolution, accurately detect the platform's minute response under jitter excitation, and thus verify the true stability of the pose.

[0196] Secondly, the system acquires displacement data from the capacitive displacement sensors on each drive shaft in real time over multiple consecutive sampling periods, such as N sampling periods, typically 10 sampling periods. This displacement data is affected by both measurement noise and minute responses. By averaging the data over multiple periods, the system can effectively improve the signal-to-noise ratio of the measurement results, achieving superior resolution.

[0197] The aforementioned technique for improving resolution by utilizing sensor noise is essentially oversampling and averaging. The principle of this method is that the random noise generated by the sensor in each measurement tends to cancel out during the averaging process of multiple samplings, while the regular response signal generated by the actual micro-motion of the platform can be revealed through mean calculation, thereby enabling accurate measurement of micro-displacements below the quantization step of a single measurement.

[0198] The purpose of implementing micro-amplitude periodic jitter on the platform is to proactively overcome potential issues such as static friction, controller integral saturation, and sensor quantization dead zones. These factors may cause the platform to appear stable but harbor hidden instability. Once the platform is subjected to external disturbances during actual operation, this hidden instability will lead to unexpected changes in pose. By applying periodic jitter commands, the system can proactively stimulate minute responses from the platform, exposing and eliminating these potential stability risks in advance.

[0199] For example, after the system injects a jitter command with an amplitude of 0.8 nanometers and a frequency of several hertz to tens of hertz into the platform, it continuously collects feedback displacement data for 10 cycles with a sampling period of 2 milliseconds, and calculates the average value of these data.

[0200] The system then evaluates the calculated mean. If the result remains stable within the preset pose difference threshold window, it indicates that the platform has good closed-loop stability and no obvious pose drift occurs, and the system continues to output the alignment completion signal.

[0201] Conversely, if the calculated mean exceeds the threshold window, it indicates that the platform's actual pose has a significant drift or stability issue. In this case, the system cancels the current alignment completion signal and re-activates the gain coefficient corresponding to the second pose adjustment command for iterative adjustment until the platform pose stabilizes again.

[0202] For example, when the mean displacement exceeds the allowable range, such as more than 5 nanometers, the system immediately performs a fine-tuning of the pose to ensure that the platform eventually stabilizes at the expected target position.

[0203] By clearly defining the super-resolution mean calculation principle and the purpose of jitter excitation, this application provides a fully disclosed closed-loop stability verification mechanism, which effectively improves the long-term stability of the automatic alignment coupling system. It is especially suitable for automatic coupling applications of silicon photonic chips and fiber arrays with strict requirements for long-term stability.

[0204] As an optional implementation, performing consistency matching on the first edge set and the second edge set to generate a common edge set includes:

[0205] Based on the periodic channel arrangement of the fiber array in the V-groove fixture, extract the V-groove edge reference marks corresponding to several adjacent channels;

[0206] Using the reference markers and pre-stored channel spacing information, the local distortion center offset of the image region where each channel is located is determined;

[0207] For each channel region, the corresponding local distortion correction is applied to map the pixel coordinates of the first edge set and the second edge set to a unified correction coordinate system;

[0208] Under the unified calibration coordinate system, the first edge set and the second edge set are matched one-to-one according to the spatial nearest neighbor or feature descriptor similarity to generate the common edge set.

[0209] Furthermore, this application also provides a multi-channel consistency matching method based on local distortion correction to ensure that the edge information of the visible light and near-infrared channel images can be accurately matched, thereby improving the accuracy of the final registration data.

[0210] In practical implementation, considering the periodic channel arrangement of the fiber array within the V-groove fixture, the system extracts several V-groove edge reference markers corresponding to adjacent channels from the acquired multimodal visual images. These V-groove edge reference markers specifically refer to the groove edge feature lines, such as the left and right edge lines of each groove. The system can use differential Gaussian filtering (DoG) or the Sobel operator to detect these groove edge lines with high precision, in order to accurately determine the position and periodic structure of the V-groove.

[0211] Secondly, the system uses the extracted V-groove reference marks to measure the spacing between adjacent V-grooves in the actual image. Specifically, the system accurately measures the actual distance between the edge lines of each groove in the image coordinate system and compares it one by one with the ideal channel spacing pre-stored in the chip design. The difference between the two is the spacing deviation, which indicates the degree of local imaging distortion.

[0212] Subsequently, the system calculates the local distortion center offset of each channel region based on the measured spacing deviation.

[0213] For example, the specific calculation method is as follows:

[0214] First, for the spacing deviation values ​​of multiple consecutive measurement points, the distortion offset trend of the entire channel region is derived by using a linear or quadratic polynomial fitting model; then, the distortion center point in each channel image region is accurately located based on this trend function, serving as the reference point for correction transformation.

[0215] Next, based on the calculated distortion center offset, the system establishes a local distortion correction model for each channel region. Commonly used models include local affine transformation or quadratic polynomial distortion correction. In specific implementation, the system uses the distortion center offset as a reference and solves for the affine transformation parameters between matching point pairs within the local image region using the least squares method, or directly determines the polynomial coefficients by substituting the offset into a preset polynomial model to accurately compensate for local image geometric distortion. The corrected image coordinates are uniformly mapped to the same correction coordinate system through this transformation model.

[0216] Finally, under a unified calibration coordinate system, the system performs precise consistency matching on the calibrated first and second edge sets. The matching method can employ either a spatial nearest neighbor algorithm or feature descriptor similarity determination to ensure a one-to-one correspondence between edge points in the two channels. For example, the system sets a spatial distance threshold of 3 pixels; when the distance between a pair of edge points from different channels is less than the threshold, they are considered a matching edge pair and included in the common edge set.

[0217] Through the aforementioned local distortion correction and precise matching steps, the system ultimately achieves accurate multi-channel consistency matching, effectively eliminating image mismatch problems caused by local distortion. This method is particularly suitable for automated alignment and coupling applications of high-density, high-precision silicon photonic chips and fiber arrays.

[0218] For example, after extracting several consecutive V-groove edge reference markers and measuring the spacing deviation, the system derives the distortion center distribution trend of each channel region through quadratic polynomial fitting, and then establishes a local distortion correction model for each channel. When these correction models are applied, the local geometric distortions that previously existed in the two-channel images are significantly reduced, and the edge overlap of each channel region is significantly improved. After matching verification under a unified coordinate system, the corresponding edges of the visible light and near-infrared channels remain closely consistent in space, laying a stable foundation for the construction of the common edge set and providing more reliable reference information for subsequent pose adjustment.

[0219] Based on the same inventive concept, this application also provides a curing device corresponding to the automatic alignment coupling method. Since the principle of the device in this application is similar to the automatic alignment coupling method described above in this application, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0220] Reference Figure 4 The diagram shown is a schematic of the curing equipment provided in this application, including:

[0221] Acquisition module 10 acquires multimodal visual images to characterize the coupling interface between the fiber array end face and the silicon photonic chip;

[0222] Registration module 20 performs spatial registration based on the multimodal visual image to generate first registration data;

[0223] The first control module 30 controls the multi-degree-of-freedom micro-displacement platform to adjust the pose of the fiber array according to the first registration data until the pose difference is detected to be less than the first preset threshold, and outputs an alignment completion signal.

[0224] The second control module 40, in response to receiving the alignment completion signal, drives the crimping mechanism to crimp the fiber array to the coupling interface of the silicon photonic chip at a preset crimping speed, and acquires online detection data under the crimping state, the online detection data including alignment error;

[0225] The detection module 50, in response to the alignment error being less than the second preset threshold, synchronously controls the ultraviolet curing module to emit ultraviolet light to the coupling interface to cure the pre-placed optical adhesive and form a cured connector, and outputs a coupling completion command; in response to the alignment error being greater than or equal to the second preset threshold, controls the crimping mechanism to release the fiber array and return to perform the spatial registration.

[0226] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

Claims

1. An automatic alignment and coupling method, characterized in that, include: Acquire multimodal visual images to characterize the coupling interface between the fiber array end face and the silicon photonic chip; Based on the multimodal visual image, spatial registration is performed to generate first registration data; Based on the first registration data, the multi-degree-of-freedom micro-displacement platform is controlled to adjust the pose of the fiber array until the pose difference is detected to be less than the first preset threshold, and an alignment completion signal is output. In response to receiving the alignment completion signal, the crimping mechanism is driven to crimp the fiber array to the coupling interface of the silicon photonic chip at a preset crimping speed, and online detection data under the crimping state is acquired, the online detection data including alignment error; In response to the alignment error being less than the second preset threshold, the UV curing module is synchronously controlled to emit UV light to the coupling interface to cure the pre-placed optical adhesive and form a cured connector, and a coupling completion command is output; in response to the alignment error being greater than or equal to the second preset threshold, the crimping mechanism is controlled to release the fiber array and return to perform the spatial registration.

2. The automatic alignment and coupling method according to claim 1, characterized in that, The generation of the first registration data includes: Obtain the layout information of the metal interconnect layer of the silicon photonics chip, and generate a mask corresponding to the metal interconnect layer; The mask is used to perform edge filtering on the multimodal visual image to remove edge information located in the metal interconnect layer region; Based on the edge-filtered multimodal visual image, coupling reference feature points are extracted, and the coordinate transformation relationship between the fiber array end face and the silicon photonic chip coupling interface is solved to generate the first registration data.

3. The automatic alignment and coupling method according to claim 2, characterized in that, The removal of edge information located in the metal interconnect layer region includes: Based on the near-infrared channel of the multimodal visual image, extract the first edge set; Extract the second edge set based on the visible light channel of the multimodal visual image; Perform consistency matching on the first edge set and the second edge set to generate a common edge set; Edge pixels that do not belong to the common edge set are deleted within the metal interconnect layer area defined by the mask, and the multimodal visual image after edge filtering is output.

4. The automatic alignment and coupling method according to claim 2, characterized in that, The generation of the first registration data includes: At the first acquisition time, a first set of reference feature points is extracted from the multimodal visual image, and the initial coordinate transformation relationship is calculated; At a second acquisition time that is separated from the first acquisition time by a preset synchronization delay, the second set of reference feature points is extracted again from the multimodal visual image, and the global tilt offset is calculated. By integrating the initial coordinate transformation relationship with the global tilt offset, a time-compensated coordinate transformation relationship is generated, and the first registration data is output.

5. The automatic alignment and coupling method according to claim 4, characterized in that, The calculation of the global tilt offset includes: The two end feature points with channel number 1 and channel number M are determined sequentially according to the channel number, where M is the total number of channels in the fiber array. Obtain the coordinate data of the two end feature points in a preset three-dimensional Cartesian coordinate system, and calculate the first pitch angle of the fiber array relative to the XZ plane based on the coordinate data using a least squares linear fitting algorithm. In the second set of reference feature points, two corresponding end feature points with the same channel number are determined as channel number 1 and channel number M. The coordinate data of the corresponding end feature points in the three-dimensional Cartesian coordinate system are obtained, and the second pitch angle of the fiber array relative to the XZ plane is calculated by the least squares straight line fitting algorithm. Calculate the absolute difference between the first pitch angle and the second pitch angle, and set this absolute difference as the global tilt offset.

6. The automatic alignment and coupling method according to claim 4 or 5, characterized in that, Before fusing the initial coordinate transformation relationship with the global tilt offset, the method further includes: At least two temperature sensors are placed at opposite corners of the silicon photonics chip substrate, and the real-time temperature data output by the temperature sensors is periodically collected. The real-time temperature data is input into a preset thermal expansion model to calculate the thermal expansion prediction vector corresponding to the geometric center of the chip. The thermal expansion model is established based on the linear expansion coefficient of the chip packaging material and the chip-fixture assembly dimensions. The thermal expansion prediction vector is vector-superimposed with the translation component of the initial coordinate transformation relationship and the rotation component of the global tilt offset to generate a thermally compensated coordinate transformation relationship.

7. The automatic alignment and coupling method according to claim 6, characterized in that, The output alignment completion signal includes: Using the first registration data as input, a first pose adjustment command is generated according to a preset gain coefficient, and the first pose adjustment command is sent to the multi-degree-of-freedom micro-displacement platform. The feedback displacement of each drive axis of the multi-degree-of-freedom micro-displacement platform is monitored in real time. If the amplitude of the coupled displacement between adjacent drive axes exceeds the preset inter-axis coupling threshold, the gain coefficient is reduced and a second pose adjustment command is generated. After executing the second pose adjustment command, the pose difference is monitored again; if the pose difference falls within the preset threshold window and the pose difference of all channels meets the channel consistency judgment condition, the alignment completion signal is output.

8. The automatic alignment and coupling method according to claim 7, characterized in that, The output alignment completion signal also includes: After the pose difference falls into the preset threshold window, a periodic micro-amplitude jitter command with an amplitude not greater than the minimum quantization step of the capacitive displacement sensor is injected into the multi-degree-of-freedom micro-displacement platform, and the average value of the feedback displacement of the capacitive displacement sensor is calculated within N consecutive sampling periods. If the mean calculation result still falls within the preset threshold window, the alignment completion signal is maintained; if the mean calculation result exceeds the preset threshold window, the gain coefficient of the multi-degree-of-freedom micro-displacement platform is iteratively adjusted according to the second pose adjustment command.

9. The automatic alignment and coupling method according to claim 3, characterized in that, The step of performing consistency matching on the first edge set and the second edge set to generate a common edge set includes: Based on the periodic channel arrangement of the fiber array in the V-groove fixture, V-groove edge reference marks corresponding to multiple adjacent channels are extracted. Using the reference markers and pre-stored channel spacing information, the local distortion center offset of the image region where each channel is located is determined; For each channel region, the corresponding local distortion correction is applied to map the pixel coordinates of the first edge set and the second edge set to a unified correction coordinate system; Under the unified calibration coordinate system, the first edge set and the second edge set are matched one-to-one according to the spatial nearest neighbor or feature descriptor similarity to generate the common edge set.

10. A curing device, characterized in that, include: The acquisition module acquires multimodal visual images to characterize the coupling interface between the fiber array end face and the silicon photonic chip; The registration module performs spatial registration based on the multimodal visual image to generate first registration data; The first control module controls the multi-degree-of-freedom micro-displacement platform to adjust the pose of the fiber array according to the first registration data until the pose difference is detected to be less than the first preset threshold, and outputs an alignment completion signal. The second control module, in response to receiving the alignment completion signal, drives the crimping mechanism to crimp the fiber array to the coupling interface of the silicon photonic chip at a preset crimping speed, and acquires online detection data under the crimping state, the online detection data including alignment error; The detection module, in response to the alignment error being less than the second preset threshold, synchronously controls the ultraviolet curing module to emit ultraviolet light toward the coupling interface to cure the pre-placed optical adhesive and form a cured connector, and outputs a coupling completion command; in response to the alignment error being greater than or equal to the second preset threshold, it controls the crimping mechanism to release the fiber array and return to perform the spatial registration.

Images