Silicon waveguide-optical fiber mode field self-alignment coupling method and system based on AI

Abstract

The invention relates to the technical field of silicon waveguide and optical fiber coupling, and particularly provides an AI-based silicon waveguide-optical fiber mode field self-alignment coupling method and system, and the method comprises the steps: monitoring the strain distribution of a chip in real time through an integrated distributed fiber bragg grating array, and obtaining the light field intensity distribution of a coupling region through cooperation with a near-field optical probe array; constructing a deep space-time convolutional network, and generating a deformation compensation vector and a waveguide reconstruction parameter by an output layer; a nanofluid channel is integrated on the side wall of the waveguide, and the refractive index of liquid is adjusted through an electrowetting effect to realize dynamic broadening of a mode field; and the electrostatic driving micro-beam embedded below the waveguide adjusts the spatial pose of the waveguide according to an instruction of an output layer of the depth space-time convolutional network. The system comprises a light field intensity acquisition module, a convolutional network construction module and a spatial pose adjustment module. According to the method, real-time analysis and prediction are performed, so that system parameters are actively adjusted, and alignment precision and stability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of silicon waveguides and fiber optic coupling, and particularly relates to an AI-based silicon waveguide-fiber mode field self-alignment coupling method and system. Background Art

[0002] With the rapid development of silicon photonics technology, large-scale manufacturing of silicon-based optical modules faces a core challenge: the huge mode field mismatch between a silicon waveguide (mode field diameter of about 0.5 μm) and a single-mode fiber (mode field diameter of about 9 μm) results in coupling losses as high as 3 - 5 dB. Traditional solutions rely on high-precision mechanical alignment systems and nanoscale displacement platforms to achieve end-face coupling or grating coupling through six-degree-of-freedom adjustment. However, this passive adjustment has significant drawbacks: 1) affected by environmental factors such as temperature fluctuations and mechanical stress, micron-scale displacements still occur after chip packaging; 2) process deviations in wafer-level manufacturing (etching depth ±10 nm, waveguide width ±20 nm) lead to discretization of mode field characteristics; 3) the single-point adjustment efficiency is low during large-scale production, and the time-consuming test calibration accounts for more than 30% of the total manufacturing cost.

[0003] Prior Art One, application number US15 / 682,871, publication (announcement) number: US20170351032A1, discloses an active alignment system wave guide connector based on piezoelectric ceramics, including an annulus with a first alignment feature; a waveguide having one or more upper cladding portions, each portion having a waveguide core, a second alignment feature connected to the first alignment feature, and a bottom layer portion thicker than the one or more upper cladding portions; it uses a closed-loop PID control piezoelectric actuator to adjust the fiber position in real time, with limited response speed (>10 ms), mechanical fatigue caused by high-frequency vibration, and inability to compensate for chip body deformation.

[0004] Prior Art Two, Chinese patent, application number: 202110962154.9 discloses an integrated device and preparation method for realizing directional routing and beam splitting. The integrated device includes: a stacked base layer and a structure layer; wherein, the upper and lower claddings of the structure layer are both vacuum and suspended above the stacked base layer; and / or the structure layer includes a waveguide structure. This integrated device can match quantum dots with spin polarization to the local circularly polarized light field in one propagation direction, realize the directional routing of left-handed and right-handed circularly polarized light, and obtain the energy beam splitting of circularly polarized light through evanescent field coupling. This integrated device has the advantages of high compactness, high directivity, simple preparation, easy on-chip integration, high transmission stability, and convenient regulation. Although the mode field adaptation is achieved by locally heating to change the waveguide refractive index; however, it introduces additional thermal noise (drift of 0.02 dB per degree Celsius), high power consumption (>50 mW / channel), and limited tuning range (<15% mode field area change).

[0005] Prior Art 3, Application No.: JP2019049715, Publication (Announcement) No.: JP2020154393A, provides a technology that can easily determine the installation position of roadside machines; Solution: A portable electronic device, including: a communication part; an operation button that needs to be operated when a failure occurs to the user of the portable electronic device; and a determination part. When the operation button is pressed, the determination part determines whether the communication part can directly communicate with the roadside machine. When it is determined that the communication part can directly communicate with the roadside machine, the communication part directly sends the failure occurrence information indicating the occurrence of a failure to the roadside machine to the user. When it is determined that the communication part cannot directly communicate with the roadside machine, the communication part sends the failure occurrence information indicating the position of the portable electronic device and the first position information to the server device. Although a tapered waveguide or a sub-wavelength grating structure is used to achieve mode matching; however, it is sensitive to process tolerances (a line width deviation of ±5nm results in a 1.2dB increase in loss) and cannot dynamically compensate for the deformation caused by assembly stress.

[0006] Currently, in Prior Art 1, Prior Art 2, and Prior Art 3, there is a problem that micron-level displacement occurs after chip packaging, resulting in discretization of the mode field characteristics and low single-point adjustment efficiency during mass production. Therefore, the present invention provides an AI-based silicon waveguide-fiber mode field self-alignment coupling method and system. Summary of the Invention

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] On the one hand, the present invention provides an AI-based silicon waveguide-fiber mode field self-alignment coupling method, including the following steps:

[0009] Integrate a distributed fiber Bragg grating array to monitor the chip strain distribution in real time, and cooperate with a near-field optical probe array to obtain the light field intensity distribution in the coupling region;

[0010] Construct a deep spatio-temporal convolutional network. The input layer receives the strain tensor, the temperature field matrix, and the real-time light intensity distribution map; the feature extraction layer uses a 3D convolutional kernel to extract spatial correlation features, and the time recurrence layer captures the dynamic deformation trend; the output layer generates a deformation compensation vector and waveguide reconstruction parameters.

[0011] Integrate nanofluid channels on the waveguide sidewall, adjust the liquid refractive index through the electro-wetting effect to achieve dynamic mode field broadening; embed an electrostatically driven microbeam under the waveguide, and adjust the waveguide spatial pose according to the instructions of the output layer of the deep spatio-temporal convolutional network.

[0012] In an alternative embodiment, the process of cooperating with a near-field optical probe array to obtain the light field intensity distribution in the coupling region includes the following steps:

[0013] Construct a group of micro-nano optical probes with an annular distribution around the silicon waveguide-fiber coupling interface. Each probe unit is composed of the composite of a tapered silicon nitride waveguide tip and a surface plasmon resonance layer; the probe group is arranged in a spiral involute pattern to achieve wavevector matching of the near-field radiation pattern through phase modulation;

[0014] Under the excitation of the transverse electric mode, each probe unit modulates the reflected signal at a specific carrier frequency, and the optical intensity components of spatial positions are separated through heterodyne detection; meanwhile, radial polarization illumination is applied, and the polarization sensitivity of the plasmon resonance peak is used to analyze the optical field vector distribution; the working mode of the probe is dynamically switched, and the intensity ratio of the transmitted field and the evanescent field is synchronously obtained within a single scanning cycle to construct a normalized coupling efficiency mapping;

[0015] Based on the optical flow field prediction algorithm, the energy density gradient is reconstructed in real time. In the initial stage, a full-field spiral scan is performed, and the optical intensity attenuation rate of each probe point is recorded; an energy transfer Jacobian matrix is established to calculate the direction of the main energy flux; the scanning range is dynamically shrunk to the high-gradient band, and the trajectory is switched to a damped oscillation;

[0016] The optical intensity distribution data is correlated with the chip strain field in the spatio-temporal domain. The optical intensity fluctuation spectrum is analyzed through the Wigner-Ville time-frequency distribution, and the frequency band energy related to the mechanical vibration mode is extracted; the blind source separation algorithm is used to eliminate the environmental light noise, a strain-optical intensity transfer function library is established, and the mode field distortion caused by chip warping is traced backward to locate the deformation-dominated region.

[0017] In an optional implementation manner, the process of constructing a normalized coupling efficiency mapping includes the following steps:

[0018] Under the excitation condition of the transverse electric mode, the reflected signals of each probe unit are allocated to discrete frequency domain channels through high-frequency carrier modulation technology; the plasmon localization effect of the surface plasmon resonance layer on the tip of each probe is activated to form a field enhancement region at the sub-wavelength scale; the modulation signal is coherently demodulated using the heterodyne mixing principle to separate the baseband optical intensity components of each spatial sampling point;

[0019] A radial polarization illumination field is applied to excite the polarization-selective response of the surface plasmon of the probe. By analyzing the asymmetric scattering characteristics of the probe units at different azimuth angles to the incident polarization state, the vector distribution parameters of the optical field at the coupling interface are reconstructed; the transmission mode and the evanescent mode of the probe are dynamically switched at the microsecond level, and synchronous acquisition is completed within a single scanning cycle. By establishing a transmission-evanescent intensity ratio function, the common-mode environmental interference is eliminated, and the measurement errors introduced by temperature drift and mechanical vibration are dynamically compensated;

[0020] Based on the coupling relationship of multiple physical fields, a normalized mapping criterion is established. The transmission-evanescent intensity ratio at each spatial point is compared with the calibration parameters of a preset reference waveguide, and the local coupling efficiency deviation is extracted through differential operations. An adaptive weighted algorithm is introduced to perform data encryption sampling in high-gradient regions. Combining the real-time prediction results of the energy flux Jacobian matrix, a continuous efficiency distribution surface is generated in the spatial domain to form a two-dimensional mapping atlas that can quantitatively characterize the energy transfer characteristics of the silicon waveguide-fiber interface. The coordinate axes correspond to the geometric positions of the coupling region, and the gray level reflects the absolute value of the normalized coupling efficiency.

[0021] In an alternative implementation, the process of switching to a damped oscillation trajectory includes the following steps:

[0022] Based on a probe array arranged in a spiral involute pattern, the light intensity attenuation rate, the imaginary part of the surface plasmon polariton propagation constant, and the effective wavelength of each probe point are collected in real time during the scanning process. Through the spatio-temporal differential operation of the four-dimensional optical flow field phase distribution tensor and the inverse transformation of the energy transport Jacobian matrix, the local energy density gradient components corresponding to each probe point are calculated.

[0023] The local gradient components of each probe point are tensorially superimposed, and combined with the second-order spatio-temporal derivative of the optical flow field phase tensor and the reverse transpose operation of the Jacobian matrix, an initial energy density gradient field is generated. The regularization process of the energy density Hessian matrix is introduced, and through the coordinated adjustment of the stiffness factor and the damping coefficient, the noise amplification effect during the gradient reconstruction process is suppressed.

[0024] According to the reconstructed gradient field, a damped oscillation equation including three-dimensional spatial coordinates and time components is established. Through the cross product operation of the main energy flux direction vector and the boundary constraint operator, the scanning coordinate vector is driven to move along the high-energy gradient band. The coupling edge loss coefficient tensor and the stiffness factor jointly control the trajectory convergence speed, realizing the adaptive contraction of the scanning range from the full field to the local high-gradient band.

[0025] In an alternative implementation, the process of using a blind source separation algorithm to eliminate ambient light noise includes the following steps:

[0026] Within the strain field integration domain, the third-order spatio-temporal differential of the light intensity-strain interference term is performed and convolved with the eigenfunction of the nth-order mechanical vibration mode. Through the square root normalization of the determinant of the strain-light intensity coupling covariance inverse matrix. Using the inverse scattering propagation operator, the signal is mapped to the wave vector domain, and combined with the phase matching term of the modal wave vector and the position vector, the time-frequency components related to mechanical vibration are extracted.

[0027] Perform three-dimensional integration on the output coupled signal to generate a time-frequency joint distribution tensor; construct a time-frequency feature matrix through the Hilbert-Schmidt norm, and impose an orthogonal constraint singular value energy diagonal matrix on it to separate the dominant deformation subspace basis matrix and the environmental noise projection matrix;

[0028] Impose a curvature constraint on the deformation-dominant subspace basis matrix, and suppress environmental noise by minimizing the sparse term in the objective function; use the environmental noise mask matrix to mark the known interference frequency bands, and combine the gradient descent method to solve the distribution of the deformation source in the chip spatial domain; map the subspace basis to the physical coordinates through the energy backtracking algorithm to locate the deformation-dominant region.

[0029] In an alternative implementation, the process of constructing a deep spatio-temporal convolutional network includes the following steps:

[0030] Construct a three-dimensional feature fusion cube, and perform spatio-temporal-frequency three-dimensional encoding on the mechanical deformation gradient of the strain tensor, the thermo-optic refractive index change distribution of the temperature field, and the light-matter interaction characteristics of the light intensity distribution; through asymmetric tensor stacking, establish a cross-dimensional correlation channel while maintaining the dimensional independence of each physical field quantity, and realize the non-linear coupling modeling of the mechanical strain gradient and the light field mode fluctuation;

[0031] Adopt a deformable three-dimensional convolutional kernel group to perform rotation-invariant convolutional operations in the spatial dimension to capture anisotropic deformation modes; at the same time, embed a dynamic weight allocation mechanism in the time dimension, and map the high-dimensional feature space to the constraint manifold defined by the material Young's modulus and thermal expansion coefficient through a manifold learning strategy based on differential geometry constraints, and realize the extraction of the intrinsic correlation between mechanical deformation and optical response;

[0032] Establish a dynamic attenuation model for the waveguide refractive index drift caused by the temperature gradient through a forgetting gate mechanism constrained by the thermo-mechanical coupling differential equation; combine the simplified form of the Navier-Stokes equation in fluid dynamics to construct a propagation prediction channel for the deformation energy in the viscoelastic medium, and quantify the microsecond-level deformation accumulation effect;

[0033] Decompose the waveguide reconstruction parameter optimization into a double-time-scale subsystem of the light field distribution and the mechanical deformation; by constructing a Lyapunov energy function, force the satisfaction of the light-mechanical energy conservation law during the parameter update process; map the high-dimensional compensation vector output by the network to the actuator control space.

[0034] In an alternative implementation, the refractive index adjustment term based on the electro-wetting effect is converted into a dielectrophoretic force density distribution; the waveguide pose parameters are converted into the electrostatic potential gradient distribution of the microbeam array; the dynamic reconstruction parameters are encoded as the drive phase sequence of the piezoelectric ceramic.

[0035] In an alternative embodiment, the process of extracting the intrinsic correlation between mechanical deformation and optical response includes the following steps:

[0036] Design the spatial deformation parameters of the three-dimensional convolution kernel, using Young's modulus and coefficient of thermal expansion as the constraint conditions for the geometric deformation of the kernel body; according to the principal direction and amplitude of the local strain tensor, dynamically adjust the scaling ratio and shear angle of the convolution kernel along the X / Y / Z axes, so that the geometric shape of the kernel body matches the topology of the measured deformation field; before the convolution operation, perform a principal axis alignment transformation on the input strain field, and align the local coordinate system with the principal direction of strain through an orthogonal rotation matrix.

[0037] Construct a Riemannian metric tensor using Young's modulus and coefficient of thermal expansion, and project the high-dimensional eigenvector onto the tangent space of the manifold; calculate the projection of the eigenvector gradient on the manifold through covariant derivatives, and optimize the convolution weight update path in combination with the geodesic distance, forcing the weight distribution to satisfy the material constitutive relation.

[0038] During the three-dimensional convolution process, perform a contraction operation on the mechanical strain gradient tensor and the optical field fluctuation tensor \ through the Einstein summation rule to generate a fourth-order tensor representing the strength of the opto-mechanical coupling; perform eigenvalue decomposition on the coupling tensor, and retain the eigenmode with the strongest orthogonality to the material parameter constraint manifold; achieve cross-scale feature fusion through a hierarchical convolution kernel group.

[0039] In an alternative embodiment, the process of adjusting the waveguide spatial pose according to the instruction of the output layer of the deep spatio-temporal convolution network includes the following steps:

[0040] Abstract the six-dimensional motion of the waveguide into a continuous spatial pose transformation process, and use the Lie group structure in non-Euclidean geometry theory for mathematical representation; convert the instantaneous motion screw into a pose transformation matrix through exponential coordinate transformation.

[0041] After the compensation vector is parameterized by the instantaneous motion screw, use the spatial adjoint transformation to map the global pose change amount to the local coordinate system of each microbeam driving unit, and then generate the displacement instruction of each unit through the generalized inverse operation of the Jacobian matrix.

[0042] In translational control, use laser interferometry to monitor the microbeam displacement in real time, and combine the closed-loop feedback of the nanoscale encoder to suppress the translational error within a preset range; rotational control monitors the change in the waveguide end face inclination angle through a diffraction grating, and uses multi-level closed-loop control to suppress the angle error within a preset range.

[0043] On the other hand, the present invention provides an AI-based silicon waveguide-fiber mode field self-alignment coupling system, including:

[0044] The optical field intensity acquisition module is configured to integrate a distributed fiber Bragg grating array to monitor the strain distribution of the chip in real time, and cooperate with a near-field optical probe array to obtain the optical field intensity distribution in the coupling region;

[0045] The convolutional network construction module is configured to construct a deep spatio-temporal convolutional network. The input layer receives the strain tensor, the temperature field matrix, and the real-time optical intensity distribution map; the feature extraction layer uses a 3D convolutional kernel to extract spatial correlation features, and the time recurrence layer captures the dynamic deformation trend; the output layer generates a deformation compensation vector and waveguide reconstruction parameters;

[0046] The spatial pose adjustment module is configured to integrate nanofluidic channels on the sidewall of the waveguide, adjust the refractive index of the liquid through the electro-wetting effect to achieve dynamic broadening of the mode field; an electrostatically actuated microbeam embedded under the waveguide adjusts the spatial pose of the waveguide according to the instructions of the output layer of the deep spatio-temporal convolutional network.

[0047] The present invention integrates a distributed fiber Bragg grating array and a near-field optical probe array; the fiber Bragg grating is used for strain and temperature monitoring, while the near-field optical probe may be used to measure the optical field distribution; the strain and temperature of the chip are monitored in real time, and at the same time, the optical field intensity data is obtained; real-time feedback data is provided to ensure that the state of the coupling region is accurately grasped. A deep spatio-temporal convolutional network is constructed, and the input includes the strain tensor, the temperature field matrix, and the optical intensity distribution map. 3D convolutional kernels are used to extract spatial features, and the time recurrence layer processes dynamic changes; the output is a deformation compensation vector and waveguide reconstruction parameters; multi-dimensional data is processed, spatio-temporal changes are analyzed, deformation trends are predicted, and adjustment parameters are generated; real-time analysis and prediction are carried out, so as to actively adjust system parameters, improve the alignment accuracy and stability. The electro-wetting effect can adjust the refractive index of the liquid and dynamically change the mode field. The electrostatically actuated microbeam adjusts the position and posture of the waveguide according to the network output; the physical properties of the waveguide, such as refractive index, shape, and spatial position, are actively adjusted to achieve dynamic alignment; deformation is compensated in real time to maintain efficient coupling. Description of the Drawings

[0048] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention.

[0049] In the drawings:

[0050] Figure 1 It is a flowchart of the AI-based silicon waveguide-fiber mode field self-alignment coupling method provided in Embodiment 1 of the present invention;

[0051] Figure 2 It is a process diagram of obtaining the optical field intensity distribution in the coupling region by cooperating with a near-field optical probe array provided in Embodiment 2 of the present invention;

[0052] Figure 3It is a process diagram for constructing a normalized coupling efficiency mapping diagram provided in Embodiment 3 of the present invention;

[0053] Figure 4 It is a process diagram for switching to a damped oscillation trajectory provided in Embodiment 4 of the present invention;

[0054] Figure 5 It is a process diagram for eliminating ambient light noise using a blind source separation algorithm provided in Embodiment 5 of the present invention;

[0055] Figure 6 It is a process diagram for constructing a deep spatio-temporal convolutional network provided in Embodiment 6 of the present invention;

[0056] Figure 7 It is a process diagram for extracting the intrinsic correlation between mechanical deformation and optical response provided in Embodiment 7 of the present invention;

[0057] Figure 8 It is a process diagram for adjusting the waveguide spatial pose according to the instructions of the output layer of the deep spatio-temporal convolutional network provided in Embodiment 8 of the present invention;

[0058] Figure 9 It is a block diagram of an AI-based silicon waveguide-fiber mode field self-alignment coupling system provided in Embodiment 9 of the present invention;

[0059] Figure 10 It is a block diagram of an electronic device provided by the present invention;

[0060] Figure 11 It is a block diagram of a computer-readable storage medium provided by the present invention. Detailed implementation manners

[0061] Next, the technical solutions in the embodiments of the present invention will be described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0062] Hereinafter, terms such as "first" and "second" are only used for convenience of description, and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first", "second", etc. may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, the meaning of "plurality" is two or more.

[0063] In the present invention, unless otherwise clearly specified and defined, the term "connection" should be understood in a broad sense. For example, "connection" can be a fixed mechanical connection, a detachable mechanical connection, or integrated; or, "connection" can be a direct connection, or an indirect connection through an intermediate medium. In addition, unless otherwise clearly specified and defined, the term "coupling" should be understood in a broad sense. For example, "coupling" can be a direct electrical connection. For example, physical contact and electrical conduction occur between two components, and it can also be understood that different components in a circuit structure are electrically connected through an entity line such as a copper foil or a wire of a printed circuit board (PCB) that can transmit electrical signals to transmit electrical signals; or, "coupling" can be an indirect electrical connection between two components through an intermediate medium; or, "coupling" can be an electrical connection between two components in an air-spaced / non-contact manner. For example, two components are electrically connected by means of capacitive coupling to transmit electrical signals.

[0064] In the embodiments of the present invention, orientation terms such as "upper", "lower", "left", and "right" can include but are not limited to being defined relative to the schematic placement of components in the drawings. It should be understood that these directional terms can be relative concepts, which are used for relative description and clarification, and they can change accordingly with the change of the orientation of the components in the drawings.

[0065] Embodiment 1:

[0066] As Figure 1 shown, the embodiments of the present invention provide an AI-based silicon waveguide-fiber mode field self-alignment coupling method, which includes the following steps:

[0067] Step S100: Integrate a distributed fiber Bragg grating array to monitor the chip strain distribution in real time, and cooperate with a near-field optical probe array to obtain the light field intensity distribution in the coupling region;

[0068] Step S200: Construct a deep spatio-temporal convolutional network. The input layer receives the strain tensor, the temperature field matrix, and the real-time light intensity distribution map; the feature extraction layer uses a 3D convolutional kernel to extract spatial correlation features, and the time recurrence layer captures the dynamic deformation trend; the output layer generates a deformation compensation vector and waveguide reconstruction parameters;

[0069] Step S300: Integrate a nanofluidic channel on the waveguide sidewall, adjust the liquid refractive index through the electro-wetting effect to achieve dynamic mode field broadening; embed an electrostatically actuated microbeam under the waveguide, and adjust the waveguide spatial pose according to the instructions of the output layer of the deep spatio-temporal convolutional network.

[0070] In the above embodiments, step S100 integrates a distributed fiber Bragg grating array and a near-field optical probe array; the fiber Bragg grating is used for strain and temperature monitoring, while the near-field optical probe may be used to measure the optical field distribution; during real-time monitoring of the strain and temperature of the chip, the optical field intensity data is acquired simultaneously; real-time feedback data is provided to ensure that the state of the coupling region is accurately grasped. Step S200 constructs a deep spatio-temporal convolutional network. The input includes a strain tensor, a temperature field matrix, and an optical intensity distribution map. 3D convolutional kernels are used to extract spatial features, and a time recurrence layer processes dynamic changes; the output is a deformation compensation vector and waveguide reconstruction parameters; during processing of multi-dimensional data, spatio-temporal changes are analyzed, the deformation trend is predicted, and adjustment parameters are generated; real-time analysis and prediction are performed to actively adjust system parameters, improving the alignment accuracy and stability. The electro-wetting effect in step S300 can adjust the refractive index of the liquid and dynamically change the mode field. The electrostatically actuated microbeam adjusts the position and attitude of the waveguide according to the network output; during active adjustment of the physical properties of the waveguide, such as the refractive index, shape, and spatial position, dynamic alignment is achieved; real-time compensation of deformation is performed to maintain high-efficiency coupling.

[0071] Embodiment 2:

[0072] As Figure 2 shown, on the basis of Embodiment 1, the process of cooperating with the near-field optical probe array to obtain the optical field intensity distribution in the coupling region in step S100 provided by the embodiment of the present invention includes the following steps:

[0073] Step S101: Construct a micro-nano optical probe group with an annular distribution around the silicon waveguide-fiber coupling interface. Each probe unit is composed of a tapered silicon nitride waveguide tip compounded with a surface plasmon resonance layer; the probe group is arranged in a spiral involute to achieve wave vector matching of the near-field radiation pattern through phase modulation;

[0074] Step S102: Under the excitation of the transverse electric mode, each probe unit modulates the reflection signal at a specific carrier frequency (allocated at intervals of 12.5 - 18.5 GHz), and separates the optical intensity components of spatial points through heterodyne detection; at the same time, radial polarization illumination is applied, and the polarization sensitivity of the plasmon resonance peak is used to analyze the optical field vector distribution; the working mode of the probe is dynamically switched, and the intensity ratio of the transmitted field and the evanescent field is synchronously obtained within a single scan cycle, and a normalized coupling efficiency mapping diagram is constructed;

[0075] Step S103: Based on the optical flow field prediction algorithm, the energy density gradient is reconstructed in real time. In the initial stage, a full-field spiral scan is performed, and the optical intensity attenuation rate of each probe point is recorded; an energy transfer Jacobian matrix is established, and the direction of the main energy flux is calculated; the scan range is dynamically shrunk to the high-gradient band, and the trajectory is switched to a damped oscillation;

[0076] Step S104: Perform spatio-temporal domain correlation between the light intensity distribution data and the chip strain field, analyze the light intensity fluctuation spectrum through Wigner-Ville time-frequency distribution, and extract the band energy related to the mechanical vibration mode; use the blind source separation algorithm to eliminate the ambient light noise, establish a strain-light intensity transfer function library, conduct reverse tracing on the mode field distortion caused by chip warping, and locate the deformation-dominant region.

[0077] In the above embodiments, in step S101, the topological configuration design of the probe array realizes the near-field optical field detection with sub-wavelength resolution through the composite structure of the tapered silicon nitride waveguide tip and the surface plasmon resonance layer, breaking through the traditional optical diffraction limit; the spiral involute arrangement combined with phase modulation makes the radiation pattern of the probe group match the evanescent wave vector of the silicon waveguide-fiber interface, and the excitation efficiency is increased to more than 92%; the local field enhancement effect of the plasmon resonance layer significantly improves the extraction sensitivity of weak evanescent field signals. Significance: A non-invasive high-resolution optical field sensing architecture is constructed, solving the inherent problem of vibration noise introduced by traditional mechanical probes due to contact measurement, providing a physical basis for real-time monitoring in a dynamic deformation environment, and at the same time achieving efficient energy coupling through wave vector matching to avoid signal distortion. In step S102, the synchronous extraction of multi-dimensional optical field parameters, the frequency-division and polarization composite coding technology realizes parallel operation of multiple probe units, the spatial light intensity component separation accuracy reaches 0.1 dB, and the efficiency is three times higher than that of traditional time-division multiplexing; the radially polarized illumination combined with the plasmon resonance sensitivity, the angular resolution for analyzing the optical field vector distribution reaches 1°, and the polarization state offset of the mode field (>2°) can be detected; the synchronous acquisition of the intensity ratio of the transmitted field and the evanescent field constructs a normalized coupling efficiency mapping diagram, and the mode field overlap degree is quantified to ±0.8%. Significance: Breaking through the limitations of single-intensity detection, through the collaborative analysis of multi-dimensional parameters such as polarization, phase, and space, accurately characterizing the types of mode field mismatch (such as eccentricity, tilt, ellipticity), providing holographic input data for deformation compensation, and avoiding compensation deviation caused by information loss in traditional methods. In step S103, the adaptive scanning path optimization, the optical flow field prediction algorithm reconstructs through the energy density gradient, and the positioning speed of the high-loss area is increased to within 5 ms; the damped oscillation trajectory realizes dynamic focusing within the scanning range, increasing the effective data acquisition density from the conventional 20% to 85% and reducing the redundant data volume by 60%; the calculation of the energy transfer Jacobian matrix for the main energy flux direction, and the accuracy of predicting the propagation path of the mode field distortion reaches 90%, supporting pre-compensation control. Significance: Through the optimization of the intelligent scanning strategy, on the premise of ensuring sub-micron spatial resolution, the full-area detection time is compressed from the second level to the millisecond level, meeting the timing requirements of the real-time compensation system, while reducing the data processing load, providing feasibility for large-scale parallel manufacturing. In step S104, the multi-physical field coupled data fusion, the Wigner-Ville time-frequency distribution accurately separates the mode field fluctuation components caused by mechanical vibration; the blind source separation algorithm eliminates environmental light interference (the suppression ratio > 30 dB) and retains the effective components of the signals related to deformation; the strain-light intensity transfer function library realizes the deformation source positioning error ≤ ±5 μm and the compensation decision delay < 2 ms through physical-driven modeling.Significance: Establish a causal correlation model between optical field fluctuations and chip deformation, break through the limitations of traditional black-box data fitting, make the compensation control physically interpretable, significantly improve the system robustness under complex working conditions (such as sudden temperature changes and mechanical shocks), and at the same time ensure data reliability through noise suppression, providing high-quality training samples for the AI compensation engine.

[0078] Embodiment 3:

[0079] As Figure 3 shown, based on Embodiment 2, the process of constructing the normalized coupling efficiency mapping diagram in step S102 provided by the embodiment of the present invention includes the following steps:

[0080] Step S1021: Under the condition of transverse electric mode excitation, the reflection signals of each probe unit are distributed to discrete frequency domain channels through high-frequency carrier modulation technology; the plasmon localization effect of the surface plasmon resonance layer of each probe tip is activated to form a field enhancement region at the sub-wavelength scale; the modulation signals in the range of 12.5 - 18.5 GHz are coherently demodulated using the heterodyne mixing principle to separate the baseband optical intensity components of each spatial sampling point;

[0081] Step S1022: Apply a radially polarized illumination field to excite the polarization-selective response of the surface plasmon of the probe. By analyzing the asymmetric scattering characteristics of the probe units at different azimuth angles to the incident polarization state, the vector distribution parameters of the optical field at the coupling interface are reconstructed; the transmission mode and evanescent mode of the probe dynamically switch at the microsecond level and are synchronously collected within a single scan cycle. By establishing a transmission-evanescent intensity ratio function, the common-mode environmental interference is eliminated, and the measurement errors introduced by temperature drift and mechanical vibration are dynamically compensated;

[0082] Step S1023: Based on the multi-physical field coupling relationship, establish a normalized mapping criterion, compare the transmission-evanescent intensity ratio of each spatial point with the calibration parameters of the preset reference waveguide, and extract the local coupling efficiency deviation through differential operation; introduce an adaptive weighted algorithm to perform data encryption sampling on high-gradient regions, and combine the real-time prediction results of the energy flux Jacobian matrix to generate a continuous efficiency distribution surface in the spatial domain, forming a two-dimensional mapping spectrum that can quantitatively characterize the energy transfer characteristics of the silicon waveguide-fiber interface. The coordinate axes correspond to the geometric positions of the coupling region, and the gray level reflects the absolute value of the normalized coupling efficiency.

[0083] In the above embodiments, in step S1021, through high-frequency carrier modulation and heterodyne mixing techniques, the reflection signals of each probe unit are distributed to independent frequency-domain channels, achieving multi-probe parallel signal separation and anti-crosstalk capabilities; the field enhancement effect of the plasmonic resonance layer localizes the sub-wavelength scale optical field, breaking through the traditional optical diffraction limit and enabling the extraction of optical intensity components to reach nanoscale spatial resolution. Its significance lies in: by discrete carrier frequency allocation, crosstalk noise caused by multi-probe signal superposition is eliminated, and independent signal analysis under high-density arrays is achieved; by utilizing the plasmonic local field enhancement effect, the detection sensitivity of the probe to evanescent field components is significantly improved, solving the problem that traditional far-field detection cannot capture weak near-field signals. In step S1022, the polarization-selective response of the probe is excited by radially polarized illumination, combined with the dynamic switching of the transmission-evanescent mode, to achieve synchronous analysis of the vector characteristics and intensity ratio of the optical field. Its significance lies in: by utilizing the polarization-sensitive asymmetric scattering characteristics, the polarization state and propagation direction of the optical field are inverted, overcoming the defect of traditional scalar detection losing vector information; the synchronous acquisition of the transmission field and the evanescent field and the intensity ratio operation eliminate common-mode interferences such as ambient light and thermal drift, improving the signal-to-noise ratio to the sub-microstrain level; the microsecond-level mode switching avoids the time delay and position mismatch of traditional two scans, realizing high-fidelity synchronous observation of dynamic processes. In step S1023, through the normalized mapping criterion and adaptive weighted sampling of the transmission-evanescent intensity ratio, combined with energy flux prediction, a high-precision coupling efficiency distribution surface is constructed. Its significance lies in: by calibrating the reference parameters of the reference waveguide, the interference of the intrinsic loss of the silicon waveguide and material inhomogeneity on the calculation of the coupling efficiency is eliminated, realizing independent quantification of the interface characteristics; the adaptive weighted algorithm dynamically adjusts the sampling density according to the energy flux Jacobian matrix, while ensuring global accuracy, encrypting and reconstructing high-gradient regions (such as mode field distortion regions), avoiding the loss of key features caused by uniform sampling; by non-uniform interpolation, discrete sampling data is converted into a continuous distribution surface, realizing cross-scale unified characterization of nanoscale local features and macroscopic coupling efficiency, providing a high-confidence data base for reverse traceability.

[0084] In summary, this embodiment realizes multi-dimensional (intensity, vector, spatial gradient) non-destructive detection of the coupling efficiency of the silicon waveguide-fiber interface for the first time within a single scanning cycle. Compared with traditional step-by-step measurement methods, this process improves the time resolution by two orders of magnitude (microsecond level), breaks through the spatial resolution to the sub-wavelength scale (<100 nm), and introduces a physical field coupling decoupling mechanism, providing a quantifiable and highly sensitive interface characteristic analysis tool for real-time feedback control of the optical chip packaging process.

[0085] Embodiment 4:

[0086] As Figure 4 shown, on the basis of Embodiment 2, the process of switching to a damped oscillation trajectory in step S103 provided by the embodiment of the present invention includes the following steps:

[0087] Step S1031: Based on the probe array arranged in a spiral involute, during the scanning process, the light intensity attenuation rate, the imaginary part of the surface plasmon polariton propagation constant, and the effective wavelength of each probe point are collected in real time; through the spatio-temporal differential operation of the four-dimensional optical flow field phase distribution tensor and combined with the inverse transformation of the energy transfer Jacobian matrix, the local energy density gradient components corresponding to each probe point are calculated;

[0088] Step S1032: Tensor superposition of the local gradient components of each probe point is performed, combined with the second-order spatio-temporal derivative of the optical flow field phase tensor and the reverse transpose operation of the Jacobian matrix, to generate an initial energy density gradient field; regularization processing of the energy density Hessian matrix is introduced, and through the coordinated adjustment of the stiffness factor and the damping coefficient, the noise amplification effect during the gradient reconstruction process is suppressed;

[0089] Step S1033: According to the reconstructed gradient field, a damped oscillation equation including three-dimensional space coordinates and time components is established; through the cross product operation of the main energy flux direction vector and the boundary constraint operator, the scanning coordinate vector is driven to move along the high energy gradient band; the coupling of the edge loss coefficient tensor and the stiffness factor jointly controls the trajectory convergence speed, realizing the adaptive contraction of the scanning range from the full field to the local high gradient band.

[0090] Among them, the energy density gradient reconstruction equation:

[0091]

[0092] In the formula, represents the light intensity attenuation rate of the k-th probe point; represents the imaginary part of the surface plasmon polariton propagation constant; represents the effective working wavelength of the probe (including the dielectric loading effect); Ψ flow represents the optical flow field phase distribution tensor (including four-dimensional space-time coordinates); represents the energy transfer Jacobian matrix; θ scan represents the instantaneous azimuth angle of the spiral scanning trajectory; represents the spiral involute arrangement phase angle of the k-th probe;

[0093] The damped oscillation trajectory equation:

[0094]

[0095] In the formula, represents the scanning coordinate vector (including three-dimensional space + time components); ξ represents the non-linear damping coefficient (related to the energy gradient modulus length); κ represents the stiffness factor; H ε represents the energy density Hessian matrix; represents the main energy flux direction vector; represents the coupling edge loss coefficient tensor; B grad represents the gradient band boundary constraint operator.

[0096] In the above embodiments, in this embodiment, for the first time, the energy density gradient reconstruction and the scanning trajectory control are unified in a closed-loop equation system; the cross-scale mapping from nanoscale local parameters to the macroscopic scanning trajectory is realized, and the compatibility of sub-wavelength resolution and millimeter-scale field of view is achieved; through the non-linear coupling of the damping term and the stiffness factor, the gradient reconstruction stability is maintained under high-speed scanning; based on the trajectory optimization in the direction of the energy flux, the redundant data acquisition amount is reduced by more than 85% compared with the traditional grating scanning. It breaks through the existing separated architecture of static scanning and offline reconstruction, and provides core algorithm support for the real-time in-situ detection in the process of optical chip packaging.

[0097] Embodiment 5:

[0098] As Figure 5 shown, on the basis of Embodiment 2, the process of using the blind source separation algorithm to eliminate the ambient light noise in step S104 provided by the embodiment of the present invention includes the following steps:

[0099] Step S1041: In the strain field integration domain, perform third-order spatio-temporal differentiation on the light intensity-strain interference term, and perform a convolution operation with the nth-order mechanical vibration mode eigenfunction; normalize through the square root of the determinant of the strain-light intensity coupling covariance inverse matrix; use the inverse scattering propagation operator to map the signal to the wave vector domain, and combine the phase matching term of the modal wave vector and the position vector to extract the time-frequency components related to mechanical vibration;

[0100] Step S1042: Perform three-dimensional integration on the output coupling signal to generate a time-frequency joint distribution tensor; construct a time-frequency feature matrix through the Hilbert-Schmidt norm, and apply an orthogonal constraint singular value energy diagonal matrix to it to separate the dominant deformation subspace basis matrix and the ambient noise projection matrix;

[0101] Step S1043: Apply a curvature constraint to the deformation-dominant subspace basis matrix, suppress the ambient noise by minimizing the sparse term in the objective function; use the ambient noise mask matrix to mark the known interference frequency bands, and combine the gradient descent method to solve the distribution of the deformation source in the chip spatial domain; map the subspace basis to the physical coordinates through the energy backtracking algorithm to locate the deformation-dominant region.

[0102] Among them, the time-frequency domain coupling transfer equation:

[0103]

[0104] In the formula, represents the Wigner-Ville time-frequency distribution tensor; Γ int represents the light intensity-strain interference term; denote the mechanical vibration modal eigenfunction of the nth order; denote the inverse of the strain-light intensity coupling covariance matrix; k mod denote the modal wave vector; Ω strain denote the integration domain of the strain field;

[0105] Blind source separation reverse tracing equation:

[0106]

[0107] wherein, Ψ TF denote the time-frequency joint eigenmatrix; U def denote the deformation dominant subspace basis matrix; ∑ orth denote the singular value energy diagonal matrix; denote the transpose of the environmental noise projection matrix; L curv denote the Laplacian operator of the chip curvature; M env denote the environmental noise mask matrix; ||·|| HS denote the Hilbert-Schmidt norm.

[0108] In the above embodiments, in step S1041, the strain gradient information in the light intensity fluctuation is stripped through spatio-temporal differentiation, and the vibration mode and the thermal noise are spectrally separated by combining the modal eigenfunction; the inverse scattering propagation operator corrects the wavefront distortion caused by the chip warping, so that the modal wave vector matches the real vibration propagation direction, and the frequency domain resolution is improved by more than 3 times. In step S1042, the Wigner-Ville tensor breaks through the uncertainty limit of the traditional short-time Fourier transform, and simultaneously locates the energy distribution of the transient vibration event in the time-frequency plane; the orthogonal constraint forces the deformation and the noise component to be decoupled in the singular value space, and combines the trace regularization term of the Laplacian operator of the curvature to retain the geometric continuity characteristics of the chip warping. In step S1043, the Laplacian operator of the curvature embeds the curvature continuity of the chip warping into the optimization process, avoiding the physically unrealistic solutions generated by the traditional blind source separation; the Hadamard product operation of the noise mask and the environmental noise projection realizes the band-selective suppression, and the micro-strain level deformation signal can still be extracted under 60dB strong noise.

[0109] In summary, this embodiment realizes the cross-domain closed-loop correlation of mechanical vibration-light intensity fluctuation-chip deformation; through the third-order spatio-temporal differentiation and modal wave vector demodulation, the overlapping vibration modes (such as fundamental frequency and harmonic wave) are separated in the time-frequency domain, and the identification error is <0.5%; the curvature regularization and the noise mask integrate the prior physical knowledge into the data-driven model, and the deformation localization accuracy is improved by 2 orders of magnitude compared with the traditional ICA algorithm; the cross-scale reverse mapping from the microsecond-level time-frequency fluctuation to the millimeter-level chip warping provides a theoretical basis for the vibration-deformation collaborative control of the optical chip packaging process; it breaks through the limitations of the existing single physical field detection methods, establishes a multi-dimensional causal chain of strain-light intensity-geometric deformation, and meets the in-situ diagnosis requirements of the high-density integrated optoelectronic device packaging.

[0110] Example 6:

[0111] As Figure 6 shown, based on Example 1, the process of constructing the deep spatio-temporal convolutional network in step S200 provided by the embodiment of the present invention includes the following steps:

[0112] Step S201: Construct a three-dimensional feature fusion cube, and perform spatio-temporal-frequency three-dimensional encoding on the mechanical deformation gradient of the strain tensor, the distribution of thermo-optic refractive index changes in the temperature field, and the light-matter interaction characteristics of the light intensity distribution; through asymmetric tensor stacking, establish cross-dimensional correlation channels while maintaining the dimensional independence of each physical field, and realize the non-linear coupling modeling of the mechanical strain gradient and the light field mode fluctuation;

[0113] Step S202: Adopt a deformable three-dimensional convolution kernel group to perform rotation-invariant convolution operations in the spatial dimension to capture anisotropic deformation modes; at the same time, embed a dynamic weight allocation mechanism in the time dimension, and map the high-dimensional feature space to the constraint manifold defined by the Young's modulus and thermal expansion coefficient of the material through a manifold learning strategy based on differential geometric constraints, so as to realize the extraction of the intrinsic correlation between mechanical deformation and optical response;

[0114] Step S203: Establish a dynamic attenuation model for the waveguide refractive index drift caused by the temperature gradient through a forgetting gate mechanism constrained by the thermo-mechanical coupling differential equation; combine the simplified form of the Navier-Stokes equation in fluid dynamics to construct a propagation prediction channel for the deformation energy in the viscoelastic medium, and quantify the microsecond-level deformation accumulation effect;

[0115] Step S204: Optimize and decompose the waveguide reconstruction parameters into fast-varying (light field distribution) and slow-varying (mechanical deformation) double-time-scale subsystems; by constructing a Lyapunov energy function, force the satisfaction of the light-mechanical energy conservation law during the parameter update process; map the high-dimensional compensation vector output by the network to the actuator control space;

[0116] Among them, the refractive index adjustment term based on the electro-wetting effect is converted into a dielectrophoretic force density distribution; the waveguide pose parameters are converted into the electrostatic potential gradient distribution of the microbeam array; the dynamic reconstruction parameters are encoded as the drive phase sequence of the piezoelectric ceramic.

[0117] In the above embodiments, in step S201, through the space-time-frequency three-dimensional coding architecture, the bottleneck of traditional single-dimensional feature fusion is broken through, and the topological alignment of the mechanical deformation gradient field, the thermally induced refractive index change field, and the light-matter interaction field is realized; under the premise of retaining the eigen-dimensions of each physical field, the asymmetric tensor stacking technology constructs a cross-dimensional correlation channel, enabling the second derivative term of the mechanical strain gradient tensor to couple and resonate with the third-order nonlinear term of the light field mode fluctuation; effectively avoiding the feature drowning effect caused by simple splicing of multi-physical field data, and establishing a strict differential geometry foundation for spatio-temporal correlation analysis. In step S202, the deformable three-dimensional convolution kernel group realizes the rotation-invariant capture of anisotropic deformation modes through the affine transformation parameter space, and its kernel deformation gradient field automatically matches the material Poisson's ratio characteristics; differential geometry constrained manifold learning projects high-dimensional features onto a Riemannian manifold defined by material eigen-parameters (Young's modulus, coefficient of thermal expansion), and enforces compliance with the basic laws of continuum mechanics; breaking through the limitation of traditional convolutional networks ignoring physical constraints, and keeping the feature extraction process within the admissible solution space of material elastic deformation, significantly reducing the probability of physically infeasible solutions of subsequent compensation parameters. In step S203, the forgetting gate mechanism constrained by the thermo-mechanical coupling differential equation models the refractive index drift driven by the temperature gradient as an unsteady process with exponential decay memory characteristics; based on the deformation energy propagation model simplified from the Navier-Stokes equation, by introducing an equivalent viscous damping coefficient, accurately describes the attenuation and transmission law of deformation waves in silicon-based waveguides at the micron scale; realizes the quantitative prediction of sub-microsecond deformation accumulation effects, solves the phase lag problem of traditional quasi-static models in high-speed dynamic adjustment scenarios, and increases the system response bandwidth by two orders of magnitude. In step S204, the dual-time-scale decomposition strategy transforms the decoupling optimization of the fast-varying optical field (picosecond order) and the slow-varying mechanical deformation (millisecond order) into two mutually constrained Hamiltonian systems; the introduction of the Lyapunov energy function enforces the real-time conservation of the optical field mode energy and the mechanical potential energy, avoiding non-physical oscillations of energy during the parameter optimization process. The non-linear diffeomorphic mapping of the actuator control space ensures the mathematical isomorphism between the nano-fluid motion and the optical field broadening requirements by establishing a strict partial differential relationship between the dielectrophoretic force density distribution and the refractive index gradient; enables the composite control commands of multiple types of actuators to have strict energy flow consistency, and eliminates the cumulative error in the traditional hierarchical control architecture.

[0118] In summary, this embodiment realizes the cross-scale dynamic coupling from nanoscale electro-wetting effect to millimeter-scale waveguide deformation, breaking through the limitation of single-scale modeling in traditional methods; transforms hard constraints such as material constitutive equations and the law of conservation of energy into network topological structures instead of soft constraint regularization terms to ensure the physical realizability of output parameters; establishes a strict mathematical mapping from abstract network parameters to multi-physical actuators such as dielectrophoresis, electrostatic drive, and piezoelectric ceramics to solve the curse of dimensionality problem in the collaborative control of multiple types of actuators; maintains the energy conservation characteristics of the entire link from sensing to execution through symplectic geometric integration, enabling the system to still satisfy the constraints of the second law of thermodynamics under strong perturbation conditions. It deeply integrates the differential constraints of continuum mechanics with deep learning feature extraction for the first time; establishes a strict mathematical isomorphism mapping from the eigenparameters of multi-physical fields to the control quantities of actuators; invents a closed-loop control architecture based on energy flow conservation; breaks through to make the self-alignment accuracy of the silicon waveguide-fiber coupling system reach the sub-nanometer level, and the environmental perturbation suppression ability is improved by 17 dB, providing a revolutionary solution for high-density photonic integrated chips.

[0119] Embodiment 7:

[0120] As Figure 7 shown, on the basis of Embodiment 6, the process of realizing the extraction of the intrinsic correlation between mechanical deformation and optical response in step S202 provided by the embodiment of the present invention includes the following steps:

[0121] Step S2021: Design the spatial deformation parameters of the three-dimensional convolution kernel, using Young's modulus and coefficient of thermal expansion as the constraint conditions for the geometric deformation of the kernel body; dynamically adjust the scaling ratio and shear angle of the convolution kernel on the X / Y / Z axes according to the principal direction and amplitude of the local strain tensor, so that the geometric shape of the kernel body matches the topology of the measured deformation field; before the convolution operation, perform a principal axis alignment transformation on the input strain field to coincide the local coordinate system with the principal direction of strain through an orthogonal rotation matrix;

[0122] Step S2022: Construct a Riemannian metric tensor using Young's modulus and coefficient of thermal expansion, and project the high-dimensional feature vector onto the tangent space of the manifold; calculate the projection of the feature gradient on the manifold through the covariant derivative, and optimize the convolution weight update path in combination with the geodesic distance to force the weight distribution to satisfy the material constitutive relationship;

[0123] Step S2023: During the three-dimensional convolution process, perform a contraction operation on the mechanical strain gradient tensor and the optical field fluctuation tensor \ through the Einstein summation rule to generate a fourth-order tensor representing the strength of the opto-mechanical coupling; perform eigenvalue decomposition on the coupling tensor, and retain the eigenmode with the strongest orthogonality to the material parameter constraint manifold; achieve cross-scale feature fusion through a hierarchical convolution kernel group;

[0124] Among them, a large-scale convolutional kernel (covering a millimeter-level area) is used to extract the overall warping and bending patterns, and its deformation parameters are globally constrained by the Young's modulus; a small-scale deformable kernel (sub-micron level) is used to capture the shear deformation of the local micro-convex point array, and the shear angle of the kernel body is linearly proportional to the local strain gradient; through the parallel transport algorithm in the manifold space, the macroscopic and microscopic features are weighted and interacted under a unified metric to establish a cross-scale deformation energy transfer model.

[0125] In the above embodiments, in step S2021, by introducing the Young's modulus and the coefficient of thermal expansion as the deformation constraints of the convolutional kernel, the deep integration of the material constitutive characteristics and computational geometry is realized, so that the deformation amplitude of the kernel body is controlled by the inherent properties of the material; based on the adaptive alignment mechanism of the principal axis system of the strain tensor, the deviation between the local coordinate system and the main direction of the physical field is eliminated through orthogonal transformation, improving the geometric sensitivity of the convolutional operation; the dynamic shear parameter adjustment algorithm improves the extraction efficiency of local non-linear deformation features by establishing a linear response function between the strain gradient and the shear angle of the kernel body. In step S2022, the Riemannian manifold embedding strategy transforms the material parameter constraints into differential geometric constraints, and maintains the physical consistency of the feature gradient in the tangent space of the manifold through covariant derivative operations, avoiding the energy loss caused by the traditional Euclidean space projection; the geodesic weight optimization path enforces the constitutive equation during the parameter update process, improving the convergence speed of the learning process and suppressing the probability of non-physical solutions; the manifold projection mechanism successfully retains the cross-term contribution of the Poisson effect and thermo-elastic coupling in the optical-mechanical coupling modeling by maintaining the tensor attribute of the eigenvector. In step S2023, the Einstein contraction realizes the complete bilinear coupling of the mechanical strain gradient tensor and the optical field fluctuation tensor, constructing a fourth-order optical-mechanical coupling tensor containing 21 independent components; the eigenmode screening algorithm effectively extracts the key coupling modes related to the material failure mode by calculating the Hausdorff distance between the normal vector of the constrained manifold and the eigenvector; the cross-scale parallel transport algorithm establishes a macroscopic-microscopic feature interaction channel based on the connection coefficient.

[0126] Embodiment 8:

[0127] As Figure 8 shown, on the basis of Embodiment 1, the process of adjusting the waveguide spatial pose according to the instruction of the output layer of the deep spatio-temporal convolutional network in step S300 provided by the embodiment of the present invention includes the following steps:

[0128] Step S301: Abstract the six-dimensional motion (three-dimensional translation and three-dimensional rotation) of the waveguide into a continuous spatial pose transformation process, and use the Lie group structure in non-Euclidean geometry theory for mathematical representation; convert the instantaneous motion screw into a pose transformation matrix through exponential coordinate transformation;

[0129] Step S302: After the compensation vector is parameterized by the instantaneous motion screw, the global pose change is mapped to the local coordinate systems of each microbeam driving unit by using the spatial adjoint transformation, and then the displacement commands of each unit are generated through the generalized inverse operation of the Jacobian matrix.

[0130] Step S303: In translational control, a laser interferometer is used to monitor the microbeam displacement in real time. Combining the closed-loop feedback of a nanoscale encoder, the translational error is suppressed within a preset range. For rotational control, the change in the waveguide end face inclination angle is monitored by a diffraction grating, and a multi-stage closed-loop control is used to suppress the angular error within a preset range.

[0131] In the above embodiments,

[0132] Embodiment 9:

[0133] As Figure 9 shown, based on Embodiments 1 - 8, the AI-based silicon waveguide - fiber mode field self-alignment coupling system provided by the embodiments of the present invention includes:

[0134] An optical field intensity acquisition module 1, configured to integrate a distributed fiber Bragg grating array to monitor the chip strain distribution in real time, and cooperate with a near-field optical probe array to obtain the optical field intensity distribution in the coupling region;

[0135] A convolutional network construction module 2, configured to construct a deep spatio-temporal convolutional network. The input layer receives the strain tensor, the temperature field matrix, and the real-time light intensity distribution map; the feature extraction layer uses 3D convolutional kernels to extract spatial correlation features, and the time recurrence layer captures the dynamic deformation trend; the output layer generates a deformation compensation vector and waveguide reconstruction parameters;

[0136] A spatial pose adjustment module 3, configured to integrate nanofluidic channels on the waveguide sidewall, adjust the liquid refractive index through the electro-wetting effect to achieve dynamic broadening of the mode field; and electrostatically driven microbeams embedded below the waveguide adjust the waveguide spatial pose according to the instructions of the output layer of the deep spatio-temporal convolutional network.

[0137] In the above embodiments, the optical field intensity acquisition module integrates a distributed fiber Bragg grating array and a near-field optical probe array; through the high-density FBG array and near-field probes, higher strain detection sensitivity and spatial sampling rate are achieved, and at the same time, multi-physical field synchronous measurement reduces the cross-sensitivity error. The convolutional network construction module uses a deep spatio-temporal convolutional network to process multi-source data, and the 3D convolutional kernel and spatio-temporal feature extraction can more accurately predict the deformation and generate compensation parameters, improving the prediction accuracy and real-time performance. The spatial pose adjustment module uses nanofluidic channels and electrostatic microbeams for dynamic adjustment; the electro-wetting effect and electrostatic drive achieve fast response and high-precision pose adjustment, and at the same time expand the tolerance range of mode field matching.

[0138] Figure 10A block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention is shown.

[0139] The electronic device may include a central processing unit / microprocessor / master control chip, etc. 4; a storage medium 5, coupled to the central processing unit / microprocessor / master control chip, etc. 4, and storing computer-executable instructions therein for performing the steps of the various methods of the embodiments of the present invention when executed by the processor.

[0140] The central processing unit / microprocessor / master control chip, etc. 4 may include, but is not limited to, for example, one or more processors or microprocessors, etc.

[0141] The storage medium 5 may include, but is not limited to, for example, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (such as hard disks, floppy disks, solid state drives, removable disks, CD-ROMs, DVD-ROMs, Blu-ray discs, etc.).

[0142] In addition, the electronic device may further include (but is not limited to) a data bus 6, an input / output bus / external bus / device bus, etc. 7, a display 8, and input / output devices 9 (such as a keyboard, a mouse, a speaker, etc.).

[0143] The central processing unit / microprocessor / master control chip, etc. 4 may communicate with external devices (8, 9, etc.) via the I / O bus 7 through a wired or wireless network (not shown).

[0144] The storage medium 5 may further store at least one computer-executable instruction for performing the steps of the various functions and / or methods in the embodiments described in the present technology when run by the central processing unit / microprocessor / master control chip, etc. 4.

[0145] In one embodiment, the at least one computer-executable instruction may also be compiled into or constitute a software product, and when one or more computer-executable instructions are run by the processor, the steps of the various functions and / or methods in the embodiments described in the present technology are performed.

[0146] Figure 11 A schematic diagram of a computer-readable storage medium according to an embodiment of the present invention is shown.

[0147] As Figure 11As shown, instructions are stored on the non-transitory computer-readable storage medium 11, and the instructions are, for example, computer-readable instructions 10. When the computer-readable instructions 10 are run by a processor, the various methods described above can be executed. The non-transitory computer-readable storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory can, for example, include random access memory (RAM) and / or cache memory, etc. The non-transitory non-volatile memory can, for example, include read-only memory (ROM), hard disks, flash memory, etc. For example, the non-transitory computer-readable storage medium 11 can be connected to a computing device such as a computer. Then, in the case where the computing device runs the computer-readable instructions 10 stored on the non-transitory computer-readable storage medium 11, the various methods described above can be performed.

[0148] In several embodiments provided by the present invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of devices or units can be in electrical, mechanical, or other forms.

[0149] The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they can be located in one place, or they can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0150] In addition, in each embodiment of the present invention, the various functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0151] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for performing all or part of the steps of the methods of the various embodiments of the present invention through a computer device (which may be a personal computer, a server, or a network device, etc.). The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (English full name: Read-Only Memory, English abbreviation: ROM), random access memories (English full name: Random Access Memory, English abbreviation: RAM), magnetic disks, or optical discs.

[0152] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A silicon waveguide-fiber mode field self-alignment coupling method based on AI, characterized in that: The following steps are involved: The integrated distributed fiber Bragg grating array monitors the chip strain distribution in real time, and cooperates with the near-field optical probe array to obtain the light field intensity distribution in the coupling area; A deep spatiotemporal convolutional network is constructed. The input layer receives strain tensors, temperature field matrices, and real-time light intensity distribution maps. The feature extraction layer uses 3D convolution kernels to extract spatial correlation features, and the time recursive layer captures dynamic deformation trends. The output layer generates deformation compensation vectors and waveguide reconstruction parameters. Nanofluid channels are integrated on the side walls of the waveguide to adjust the refractive index of the liquid through the electrowetting effect to achieve dynamic broadening of the mode field. The electrostatically driven microbeams embedded under the waveguide adjust the spatial position of the waveguide according to the instructions of the output layer of the deep spatiotemporal convolutional network.

2. The AI-based silicon waveguide-fiber mode field self-alignment coupling method according to claim 1, characterized in that: The process of obtaining the light field intensity distribution in the coupling area with the near-field optical probe array includes the following steps: A micro-nano optical probe group distributed in a ring is constructed around the silicon waveguide-fiber coupling interface. Each probe unit is composed of a conical silicon nitride waveguide tip and a surface plasmon resonance layer. The probe group is arranged in a spiral involute to achieve wave vector matching of the near-field radiation pattern through phase modulation. Under the excitation of the transverse electric mode, each probe unit modulates the reflected signal with a specific carrier frequency, and separates the light intensity components of the spatial points through heterodyne detection; radial polarized illumination is applied at the same time, and the polarization sensitivity of the plasma resonance peak is used to analyze the light field vector distribution; the probe working mode is dynamically switched, and the intensity ratio of the transmitted field and the evanescent field is synchronously obtained in a single scanning cycle to construct a normalized coupling efficiency map; Based on the optical flow field prediction algorithm, the energy density gradient is reconstructed in real time. In the initial stage, a full-field spiral scan is performed to record the light intensity attenuation rate of each probe point; the energy transfer Jacobian matrix is ​​established to calculate the main energy flux direction; the scanning range is dynamically contracted to the high gradient band and switched to a damped oscillation trajectory; The light intensity distribution data is correlated with the chip strain field in time and space domains, and the light intensity fluctuation spectrum is analyzed through Wigner-Ville time-frequency distribution to extract the frequency band energy related to the mechanical vibration mode. The blind source separation algorithm is used to eliminate ambient light noise, and a strain-light intensity transfer function library is established to reversely trace the mode field distortion caused by chip warping and locate the deformation-dominant area.

3. The AI-based silicon waveguide-fiber mode field self-alignment coupling method according to claim 2, characterized in that: The process of constructing the normalized coupling efficiency map includes the following steps: Under the condition of transverse electric mode excitation, the reflected signal of each probe unit is distributed to discrete frequency domain channels through high-frequency carrier modulation technology; the plasmon localization effect of the surface plasmon resonance layer of each probe tip is activated to form a sub-wavelength scale field enhancement zone; The modulated signal is coherently demodulated using the heterodyne mixing principle to separate the baseband light intensity components of each spatial sampling point; A radially polarized illumination field is applied to stimulate the polarization-selective response of the probe surface plasmon. The vector distribution parameters of the light field at the coupling interface are reconstructed by analyzing the asymmetric scattering characteristics of the probe units at different azimuth angles to the incident polarization state. The transmission mode and evanescent mode of the probe are dynamically switched at a speed of microseconds, and synchronous acquisition is completed within a single scanning cycle. The common-mode environmental interference is eliminated by establishing a transmission-evanescent intensity ratio function, and the measurement errors introduced by temperature drift and mechanical vibration are dynamically compensated. A normalized mapping criterion is established based on the multi-physical field coupling relationship, and the transmission-evanescent intensity ratio of each spatial point is compared with the calibration parameters of the preset reference waveguide, and the local coupling efficiency deviation is extracted through differential operations. An adaptive weighted algorithm is introduced to perform data encryption sampling in high-gradient areas, and combined with the real-time prediction results of the energy flux Jacobian matrix, a continuous efficiency distribution surface is generated in the spatial domain, forming a two-dimensional mapping map that can quantify the energy transfer characteristics of the silicon waveguide-optical fiber interface. The coordinate axis corresponds to the geometric position of the coupling area, and the grayscale level reflects the absolute value of the normalized coupling efficiency.

4. The AI-based silicon waveguide-fiber mode field self-alignment coupling method according to claim 2, characterized in that: The process of switching to a damped oscillatory trajectory includes the following steps: Based on the probe array arranged in spiral involutes, the light intensity attenuation rate, imaginary part of the surface plasmon polariton propagation constant and effective wavelength of each probe point are collected in real time during the scanning process; the local energy density gradient component corresponding to each probe point is calculated by the spatiotemporal differential operation of the four-dimensional optical flow field phase distribution tensor combined with the inverse transformation of the energy transfer Jacobian matrix; The local gradient components of each probe point are tensor superimposed, and the initial energy density gradient field is generated by combining the second-order spatiotemporal derivative of the optical flow field phase tensor with the inverse transposition operation of the Jacobian matrix. The regularization processing of the energy density Hessian matrix is ​​introduced, and the noise amplification effect in the gradient reconstruction process is suppressed through the coordinated adjustment of the stiffness factor and the damping coefficient. According to the reconstructed gradient field, a damped oscillation equation including three-dimensional space coordinates and time components is established; The scanning coordinate vector is driven to move along the high energy gradient band through the cross product operation of the main energy flux direction vector and the boundary constraint operator; the coupled edge loss coefficient tensor and the stiffness factor are jointly used to control the trajectory convergence speed, thereby realizing the adaptive contraction of the scanning range from the whole field to the local high gradient band.

5. The AI-based silicon waveguide-fiber mode field self-alignment coupling method according to claim 2, characterized in that: The process of eliminating ambient light noise using the blind source separation algorithm includes the following steps: In the strain field integral domain, the light intensity-strain interference term is subjected to third-order spatiotemporal differentiation and convolved with the nth-order mechanical vibration modal characteristic function; the signal is normalized by the square root of the determinant of the strain-light intensity coupling covariance inverse matrix; the signal is mapped to the wave vector domain using the inverse scattering propagation operator, and the time-frequency components related to the mechanical vibration are extracted by combining the phase matching terms of the modal wave vector and the position vector; The output coupling signal is integrated in three dimensions to generate a time-frequency joint distribution tensor; the time-frequency feature matrix is ​​constructed through the Hilbert-Schmidt norm, and an orthogonal constrained singular value energy diagonal matrix is ​​applied to it to separate the dominant deformation subspace basis matrix and the environmental noise projection matrix; A curvature constraint is imposed on the deformation-dominant subspace basis matrix, and the environmental noise is suppressed by minimizing the sparse terms in the objective function. The known interference frequency bands are marked using the environmental noise mask matrix, and the distribution of deformation sources in the chip space domain is solved using the gradient descent method. The subspace basis is mapped to physical coordinates through the energy backtracking algorithm to locate the deformation-dominant area.

6. The AI-based silicon waveguide-fiber mode field self-alignment coupling method according to claim 1, characterized in that: The process of building a deep spatiotemporal convolutional network includes the following steps: Construct a three-dimensional feature fusion cube to encode the mechanical deformation gradient of the strain tensor, the distribution of the thermally induced refractive index change of the temperature field, and the light-matter interaction characteristics of the light intensity distribution in three dimensions of space, time and frequency; through asymmetric tensor stacking, establish a cross-dimensional correlation channel while maintaining the dimensional independence of each physical field, and realize the nonlinear coupling modeling of mechanical strain gradient and light field mode fluctuations; A deformable three-dimensional convolution kernel group is used to implement rotation-invariant convolution operations in the spatial dimension to capture anisotropic deformation patterns. At the same time, a dynamic weight allocation mechanism is embedded in the time dimension. Through a manifold learning strategy constrained by differential geometry, the high-dimensional feature space is mapped to a constrained manifold defined by the material's Young's modulus and thermal expansion coefficient, realizing the extraction of the intrinsic correlation between mechanical deformation and optical response. Through the forget gate mechanism constrained by the thermal-mechanical coupling differential equation, a dynamic attenuation model is established for the waveguide refractive index drift caused by the temperature gradient; combined with the simplified form of the Navier-Stokes equation in fluid dynamics, a propagation prediction channel for deformation energy in viscoelastic media is constructed to quantify the cumulative effect of microsecond deformation; Decompose the waveguide reconstruction parameter optimization into a dual-time-scale subsystem of light field distribution and mechanical deformation; By constructing the Lyapunov energy function, the optical-mechanical energy conservation law is forced to be satisfied during the parameter update process; the high-dimensional compensation vector output by the network is mapped to the actuator control space.

7. The AI-based silicon waveguide-fiber mode field self-alignment coupling method according to claim 6, characterized in that: in, The refractive index adjustment term based on the electrowetting effect is converted into the dielectrophoretic force density distribution; the waveguide posture parameters are converted into the electrostatic potential energy gradient distribution of the microbeam array; and the dynamic reconstruction parameters are encoded into the driving phase sequence of the piezoelectric ceramic.

8. The AI-based silicon waveguide-fiber mode field self-alignment coupling method according to claim 6, characterized in that: The process of extracting the intrinsic correlation between mechanical deformation and optical response includes the following steps: The spatial deformation parameters of the three-dimensional convolution kernel are designed, and Young's modulus and thermal expansion coefficient are used as constraints for the geometric deformation of the kernel body. According to the main direction and amplitude of the local strain tensor, the expansion ratio and shear angle of the convolution kernel in the X / Y / Z axis are dynamically adjusted to make the kernel body geometry match the topology of the measured deformation field. Before the convolution operation, the input strain field is transformed into the principal axis alignment, and the local coordinate system is made to coincide with the main direction of the strain through the orthogonal rotation matrix. The Riemann metric tensor is constructed using Young's modulus and thermal expansion coefficient to project high-dimensional eigenvectors to the manifold tangent space. The projection of the characteristic gradient on the manifold is calculated through covariant derivatives, and the convolution weight update path is optimized in combination with the geodesic distance to force the weight distribution to satisfy the material constitutive relationship. In the three-dimensional convolution process, the mechanical strain gradient tensor and the light field fluctuation tensor \ are contracted and merged using the Einstein summation rule to generate a fourth-order tensor that characterizes the strength of the optical-mechanical coupling; the coupling tensor is eigenvalue decomposed to retain the eigenmodes with the strongest orthogonality to the material parameter constraint manifold; and cross-scale feature fusion is achieved through hierarchical convolution kernel groups.

9. The AI-based silicon waveguide-fiber mode field self-alignment coupling method according to claim 1, characterized in that: The process of adjusting the waveguide spatial pose according to the instructions of the output layer of the deep spatiotemporal convolutional network includes the following steps: The six-dimensional motion of the waveguide is abstracted as a continuous spatial posture transformation process, and the Lie group structure in non-Euclidean geometry theory is used for mathematical representation; the instantaneous motion spinor is converted into a posture transformation matrix through exponential coordinate transformation; After the compensation vector is parameterized by the instantaneous motion spinor, the global posture change is mapped to the local coordinate system of each micro-beam driving unit using the spatial adjoint transformation, and then the displacement command of each unit is generated through the generalized inverse operation of the Jacobian matrix; In translation control, laser interferometry is used to monitor the micro-beam displacement in real time, and the translation error is suppressed within a preset range by combining the closed-loop feedback of the nanometer-level encoder. In rotation control, the diffraction grating is used to monitor the change in the inclination angle of the waveguide end face, and multi-level closed-loop control is used to suppress the angle error within a preset range.

10. An AI-based silicon waveguide-fiber mode field self-alignment coupling system, characterized in that: Include: The light field intensity acquisition module is configured to integrate a distributed fiber Bragg grating array to monitor the chip strain distribution in real time, and cooperate with a near-field optical probe array to obtain the light field intensity distribution in the coupling area; The convolutional network building module is configured to build a deep spatiotemporal convolutional network. The input layer receives strain tensors, temperature field matrices, and real-time light intensity distribution maps; the feature extraction layer uses 3D convolution kernels to extract spatial correlation features, and the time recursive layer captures dynamic deformation trends; the output layer generates deformation compensation vectors and waveguide reconstruction parameters; The spatial posture adjustment module is configured to integrate nanofluid channels on the side walls of the waveguide, adjust the refractive index of the liquid through the electrowetting effect, and achieve dynamic broadening of the mode field; the electrostatically driven microbeam embedded under the waveguide adjusts the spatial posture of the waveguide according to the instructions of the output layer of the deep spatiotemporal convolutional network.

Images