Method for designing a diamond tool for fiber cleaving

Abstract

The application relates to the field of mechanical engineering and discloses a machining design method based on a diamond cutter for fiber cutting. The method comprises the following steps: collecting the three-dimensional morphology of a fiber end face cut by diamond cutters with different edge parameters and optical performance indexes thereof; extracting microscopic geometric features such as edge curvature gradient, crystal direction orientation, collapse density and surface power spectral density; constructing a physically interpretable neural network model embedded with fracture mechanics and optical waveguide constraints; reversely solving an edge parameter set meeting target optical performance; and generating an ultra-precision machining path to control a single-point diamond lathe to customize diamond cutter machining. Through closed-loop optimization and online fine adjustment, the application improves the stability of end face reflectivity and the consistency of mode fields, reduces the standard deviation of return loss, and meets the requirements of high-speed optical communication.

Description

Technical Field

[0001] This invention belongs to the field of mechanical engineering, specifically relating to a processing design method based on a diamond cutter for fiber optic cutting. Background Technology

[0002] With the rapid development of fiber optic communication, sensing, and precision manufacturing technologies, the requirements for fiber optic end-face processing quality are becoming increasingly stringent. As a pre-process for optical device packaging and interconnection, fiber optic cleaving directly determines the transmission efficiency, return loss, and connection reliability of the optical signal. In this process, the diamond cutter, as the cutting tool, has its cutting edge geometry as a fundamental factor affecting the cutting quality. Traditional diamond cutter design relies primarily on process experience and macroscopic mechanical analysis, focusing on mechanical properties such as tool strength and wear resistance, but failing to establish a quantitative mapping relationship between the microscopic geometric characteristics of the cutting edge and the final optical performance of the fiber end-face. This limitation means that even if the cut end-face appears "smooth" or "without chipping" under a microscope, its microscopic morphology may still contain nanoscale undulations, microcracks, or non-ideal fracture surfaces, leading to light scattering, mode mismatch, or enhanced Fresnel reflection, resulting in unpredictable fluctuations in optical performance.

[0003] The fiber optic cutting process based on diamond cutters is essentially a controlled brittle fracture behavior, and its end-face morphology is jointly determined by the local stress field distribution applied by the cutter, the crack initiation location, and the propagation path. While existing technologies can simulate the stress-strain response of the cutter-fiber contact area using the finite element method or observe the end-face morphology experimentally, these methods are isolated: mechanical simulations cannot predict optical consequences, and optical testing is difficult to use to deduce the direction of cutter shape optimization. The lack of a systematic design framework that couples the solid mechanical response with the light wave propagation characteristics has kept cutter development in an inefficient cycle of trial and error.

[0004] In existing technologies, there is no diamond cutter design method that establishes a closed-loop correlation between the mechanical formation mechanism of the fiber optic cutting end face and its optical transmission performance. Current solutions suffer from the following problems: the cutting edge parameter settings lack optical target guidance, only satisfying mechanical cutting feasibility; end face quality assessment lags behind the manufacturing process, relying on offline optical inspection and cannot be used for design iteration; and a calculable mapping model from cutting edge input to optical output is not constructed, making reverse design with the optimization goal of maximizing optical transmission efficiency or minimizing reflectivity impossible. In applications such as high-speed optical interconnects, quantum communication, and high-power laser transmission, which have high requirements for end face optical consistency, these shortcomings limit the performance ceiling and mass production yield of fiber optic devices. Therefore, an intelligent reverse design method integrating multiphysics simulation is urgently needed to achieve precise control of the micro-geometry of the diamond cutter edge and predictable assurance of optical performance. Summary of the Invention

[0005] This invention provides a processing design method based on a diamond cutter for fiber optic dicing, aiming to solve the technical problem that traditional diamond cutter edge design relies on experience and macroscopic mechanical analysis, making it impossible to establish a precise mapping relationship between the microscopic geometry of the cutting edge and the optical performance of the fiber end face. In existing technologies, the cutting edge of a diamond cutter is usually simplified to an idealized straight line or arc contour. The processing mainly relies on the operator's experience to set macroscopic parameters such as the rake angle, clearance angle, and cutting edge radius, while ignoring the true three-dimensional morphological characteristics of the cutting edge at the nanoscale, including factors such as the cutting edge curvature gradient, microscopic chipping, crystal orientation deviation, and surface roughness distribution. This coarse design method results in fluctuations in optical performance indicators such as end face reflectivity, return loss, and mode field matching degree during actual dicing, even though the fiber end face appears mechanically acceptable under a microscope (no cracks, no burrs), making it difficult to meet the consistency requirements of high-precision optical communication systems for end face quality.

[0006] This invention provides a machining design method based on a diamond cutter for fiber optic dicing, comprising:

[0007] End face samples were collected by cutting standard single-mode optical fibers with multiple diamond cutters of different micro-geometric characteristics.

[0008] Three-dimensional surface topography scanning is performed on each end face sample to obtain its surface height field data, and its optical performance indicators are measured simultaneously, including end face reflectivity, return loss and mode field distortion coefficient.

[0009] Microscopic geometric feature parameters of each diamond cutting edge are extracted. These microscopic geometric feature parameters include the radius of curvature of the cutting edge profile, the rate of change of curvature gradient, the atomic-level flatness of the cutting edge tip, the crystal orientation angle of the front and rear surfaces of the cutting edge, the microscopic chipping density and depth distribution of the cutting edge along the cutting direction, and the power spectral density function of the surface roughness of the cutting edge.

[0010] Based on the microscopic geometric feature parameters and the corresponding end-face optical performance indicators, a multi-input multi-output physical constraint neural network model is constructed. This physical constraint neural network model takes the microscopic geometric feature parameters of the cutting edge as input, the end-face optical performance indicators as output, and embeds the material fracture mechanics equation and the optical waveguide boundary conditions as regularization constraints.

[0011] A target optical performance index threshold range is set, and the target optical performance index threshold range is used as the inverse solution target of the physical constraint neural network model. The micro-geometric feature parameters of the input layer are iteratively optimized through the gradient backpropagation algorithm until the optical performance index of the output layer falls within the target optical performance index threshold range, thereby obtaining a set of solution sets of micro-geometric parameters of the cutting edge that meet the optical performance requirements.

[0012] The micro-geometric parameters of the cutting edge are converted into ultra-precision machining path instructions, and the nano-positioning platform of the single-point diamond lathe is controlled to perform flying cutting on the single-crystal diamond billet according to the ultra-precision machining path to form a cutting edge structure with a specified micro-geometric morphology.

[0013] Preferably, a three-dimensional surface topography scan is performed on each end face sample to obtain its surface height field data, and its optical performance indicators are measured simultaneously, including:

[0014] A white light interferometer was used to scan a circular region with a diameter of 100 micrometers at the center of the fiber end face to obtain surface height field data.

[0015] The reflectivity of the end face was measured by an optical time domain reflectometer, the return loss was determined by a polarization-independent return loss tester, and the mode field intensity distribution image was obtained by a near-field scanning optical microscope and the mode field distortion coefficient was calculated.

[0016] Preferably, the microscopic geometric feature parameters of each diamond cutting edge are extracted, including:

[0017] The cross-sectional image of the cutting edge tip was obtained by focusing ion beam scanning electron microscopy, and the contour points within a 10-nanometer range of the cutting edge tip were fitted to a quadratic curve to calculate the radius of curvature.

[0018] The change in radius of curvature per unit distance along the cutting edge length direction is calculated using the sliding window difference method as the rate of change of curvature gradient.

[0019] The surface height of the five atomic layers at the very tip of the cutting edge was observed by high-resolution transmission electron microscopy, and its standard deviation was calculated as atomic-level flatness.

[0020] The crystal orientation angles of the front and rear surfaces of the cutting edge were determined using electron backscatter diffraction.

[0021] The number of local depressions with a depth greater than 0.5 nanometers per micrometer of cutting edge length is counted using an automatic image recognition algorithm as the micro-fracture density, and the ratio of the maximum depth to the average depth is calculated as a characterization of the depth distribution.

[0022] After performing a fast Fourier transform on the surface height data of the cutting edge, the modulus is squared and normalized to obtain the power spectral density function of the surface roughness, and the energy proportion of high-frequency components with spatial frequencies greater than 1.5 cycles per micrometer is limited to below 15%.

[0023] Preferably, based on the microscopic geometric feature parameters and the corresponding end-face optical performance indicators, a multi-input multi-output physical constraint neural network model is constructed, including:

[0024] The input layer is set to contain 7 neurons, corresponding to the radius of curvature, rate of change of curvature gradient, standard deviation of atomic flatness, crystal orientation deviation angle of the rake face, crystal orientation deviation angle of the flank face, micro-fracture density, and proportion of high-frequency power spectral density, respectively.

[0025] The output layer is configured to contain three neurons, corresponding to end face reflectivity, return loss, and mode field distortion coefficient, respectively.

[0026] Two hidden layers are set up, and the activation function is the modified linear unit;

[0027] The stress intensity factor calculation module based on Griffith's fracture theory is introduced into the loss function as the first physical constraint term, and the continuity constraint of the electric field tangential component based on the boundary conditions of Maxwell's equations is introduced as the second physical constraint term.

[0028] Preferably, the construction of the first physical constraint term includes:

[0029] The equivalent crack length and geometric correction factor are calculated based on the cutting edge geometry, and the stress intensity factor is calculated in combination with the cutting stress.

[0030] The stress intensity factor is mapped to a predicted value of the subsurface damage depth of the fiber end face, and the square of the difference between the predicted value and the actual damage depth constitutes the first physical constraint loss term.

[0031] Preferably, the construction of the second physical constraint term includes:

[0032] Based on the predicted end-face height field, the normal derivative of the tangential component of the electric field at the air-quartz interface is calculated using the finite difference approximation.

[0033] Based on the difference in dielectric constant, the continuity residual of the tangential component of the electric field is constructed, and its sum of squares constitutes the second physical constraint loss term.

[0034] Preferably, a target optical performance index threshold range is set, and this target optical performance index threshold range is used as the inverse solution target of the physical constraint neural network model. The microscopic geometric feature parameters of the input layer are iteratively optimized through a gradient backpropagation algorithm, including:

[0035] Set the target threshold ranges for end-face reflectivity, return loss, and mode field distortion coefficient;

[0036] The constrained projective gradient descent algorithm is used for inverse solution;

[0037] After each iteration, the physical feasibility of the microscopic geometric parameters is verified, and infeasible solutions that are outside the feasible range are eliminated.

[0038] Preferably, converting the solution set of the micro-geometric parameters of the cutting edge into ultra-precision machining path instructions includes:

[0039] The three-dimensional profile curve of the cutting edge is reconstructed based on the radius of curvature and the rate of change of curvature gradient in the solution set;

[0040] The three-dimensional profile curve of the cutting edge is discretized into a three-dimensional coordinate sequence;

[0041] Generate machining path instructions that include a three-dimensional coordinate sequence of the tool path, feed rate, depth of cut, and spindle speed.

[0042] Preferably, after the diamond cutting process is completed, a verification step is also included:

[0043] Use the diamond cutter to cut a new fiber sample and repeat the optical performance measurement process.

[0044] If the measured optical performance index deviates from the target threshold by more than 5%, new data will be added to the training set, the physical constraint neural network model will be fine-tuned online, and the reverse solution and processing process will be re-executed until the verification is successful.

[0045] Preferably, the online fine-tuning adopts a transfer learning strategy, freezing the weights of the first two layers of the neural network and adjusting only the last layer and the output layer.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. This invention establishes a quantitative mapping relationship between the nanoscale micro-geometric features of the diamond cutting edge and the optical performance of the fiber end face, breaking through the traditional design paradigm that relies on trial and error based on experience.

[0048] 2. By introducing a physical constraint neural network model, material fracture mechanics and optical waveguide theory are embedded into the data-driven framework, ensuring the model's generalization ability and physical consistency.

[0049] 3. The reverse design mechanism allows the blade structure to be directly customized for the target optical performance, improving the stability of the fiber end face reflectivity.

[0050] 4. By strictly controlling the curvature gradient of the cutting edge, crystal orientation, and surface power spectral density, subsurface damage and high-frequency scattering on the end face are suppressed, and the mode field distortion coefficient is stabilized.

[0051] 5. The closed-loop verification and online fine-tuning mechanism ensures the robustness and repeatability of the processing technology, making it suitable for large-scale, high-consistency fiber optic device manufacturing scenarios. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0053] Figure 2This is a schematic diagram of the core principle framework of the reverse-drive cutting edge micro-geometry design in this invention, which takes optical performance as the optimization target;

[0054] Figure 3 This is a logical flowchart of the fiber optic end-face sample acquisition, three-dimensional morphology scanning and optical performance measurement in this invention;

[0055] Figure 4 This is a flowchart illustrating the logical process of extracting and characterizing the microscopic geometric features of the diamond cutting edge in this invention.

[0056] Figure 5 This is a flowchart illustrating the logical process of constructing a physical constraint neural network model and inversely solving for the cutting edge parameters in this invention.

[0057] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal and the cloud in this invention. Detailed Implementation

[0058] refer to Figures 1 to 6 This invention provides a processing design method based on a diamond cutter for fiber optic dicing. This method optimizes the optical performance of the fiber end face by constructing a physically interpretable mapping relationship between the microscopic geometric features of the cutting edge and the optical response of the end face, thereby achieving reverse customization design and ultra-precision machining of the nanoscale structure of the diamond cutter edge. The specific embodiments of this invention are described in detail below with reference to the accompanying drawings.

[0059] The method includes the following steps:

[0060] End face samples were collected by cutting standard single-mode optical fibers with multiple diamond cutters of different micro-geometric characteristics.

[0061] Three-dimensional surface topography scanning was performed on each end face sample to obtain its surface height field data, and its optical performance indicators were measured simultaneously.

[0062] Extract the microscopic geometric feature parameters of the cutting edge of each diamond cutting tool;

[0063] Based on the aforementioned microscopic geometric feature parameters and corresponding end-face optical performance indicators, a multi-input multi-output physical constraint neural network model is constructed.

[0064] A target optical performance index threshold range is set, and this target optical performance index threshold range is used as the inverse solution target of the physical constraint neural network model. The micro-geometric feature parameters of the input layer are iteratively optimized through the gradient backpropagation algorithm until the optical performance index of the output layer falls within the threshold range, thereby obtaining a set of solution sets of micro-geometric parameters of the cutting edge that meet the optical performance requirements.

[0065] The micro-geometric parameters of the cutting edge are converted into ultra-precision machining path instructions, and the nano-positioning platform of the single-point diamond lathe is controlled to perform fly cutting on natural or artificial single-crystal diamond blanks according to the ultra-precision machining path to form a cutting edge structure with a specified micro-geometric morphology.

[0066] End-face samples were collected by cutting standard single-mode optical fibers using multiple diamond cutters with different micro-geometric cutting edges. The standard single-mode optical fiber was a quartz-based fiber conforming to the ITU-T G652D standard, with a mode field diameter of 9.2 μm, a cladding diameter of 125 μm, and a coating outer diameter of 245 μm. Each diamond cutter was installed in the same type of fiber optic dicing equipment to ensure that process parameters such as cutting tension, cutting speed, and pre-stretch length were completely consistent, eliminating the influence of non-cutting edge factors on end-face quality. During the dicing process, the fiber was fixed in a high-rigidity fixture, with a constant tension of 1.5 Newtons, a cutting speed of 0.8 mm / s, and a pre-stretch length of 10 mm. Each diamond cutter continuously cut 50 fibers, and 30 end faces without macroscopic defects were selected as valid samples for subsequent data acquisition. All samples were immediately sealed and stored in a nitrogen atmosphere after dicing to prevent surface contamination or oxidation from affecting the accuracy of optical measurements.

[0067] Three-dimensional surface topography scanning was performed on each end-face sample to obtain its surface height field data, and its optical performance indicators were measured simultaneously. The three-dimensional surface topography scanning was achieved using a white light interferometer with a lateral resolution of 40 nm and a longitudinal resolution of 0.08 nm. The scanning area was a circular region with a diameter of 100 μm at the center of the fiber end-face, with 1024 × 1024 sampling points. During the scanning process, the light source wavelength was 530 nm, and the coherence length was 2 μm, ensuring sensitive capture of sub-nanometer height changes. The obtained height field data was stored in the form of a two-dimensional matrix, where each element represents the surface height value at the corresponding spatial location, in nanometers.

[0068] The simultaneous optical performance measurements included three indicators: end-face reflectivity, return loss, and mode field distortion coefficient. End-face reflectivity was measured using an optical time-domain reflectometer with an incident light wavelength of 1550 nm and a measurement accuracy of ±0.02 dB. Return loss was measured using a polarization-independent return loss meter with a dynamic range of not less than 70 dB. The mode field distortion coefficient was obtained by acquiring the mode field intensity distribution image using a near-field scanning optical microscope, calculating its correlation coefficient with an ideal Gaussian mode field, and then subtracting the absolute value of the correlation coefficient from one. The smaller the value, the higher the mode field fidelity.

[0069] Microscopic geometric feature parameters of each diamond cutting edge were extracted. These parameters include the radius of curvature of the cutting edge profile, the rate of change of curvature gradient, the atomic-level flatness of the cutting edge tip, the crystal orientation angles of the front and rear surfaces of the cutting edge, the micro-chipping density and depth distribution along the cutting direction, and the power spectral density function of the cutting edge surface roughness. The cutting edge profile was obtained using focused ion beam scanning electron microscopy (FEM) with an accelerating voltage of 30 kV and a beam current of 10 picoamperes. A cross-sectional cut was performed on the cutting edge tip, and the image resolution reached 0.5 nm. From the mode field intensity distribution image obtained by this near-field scanning optical microscope, a sequence of contour points within a 20 nm range of the cutting edge tip was extracted and fitted to a quadratic curve. , , , These are the fitting coefficients. For a set of associated measured parameters, radius of curvature .

[0070] The rate of change of curvature gradient is defined as the change in radius of curvature per unit distance along the cutting edge. It is calculated using the sliding window difference method: with a window length of 200 nanometers, the window slides along the entire length of the cutting edge, calculating the difference in radius of curvature at the center points of adjacent windows and dividing it by the window spacing to obtain the local gradient value. Finally, the root mean square value over the entire cutting edge length is taken as the overall rate of change of curvature gradient. The atomic-level flatness of the cutting edge tip is determined by observing the lattice arrangement of the five atomic layers at the very tip of the cutting edge using a high-resolution transmission electron microscope, extracting surface height data, and calculating its standard deviation. ,when A cutting edge with a diameter less than 0.3 nanometers is considered to have high flatness. The crystal orientation angles of the front and rear surfaces of the cutting edge are determined by electron backscatter diffraction. The sample tilt angle is 70 degrees and the step size is 10 nanometers. After collecting Kikuchi flower patterns, the samples are compared with the diamond crystal standard database to determine that the front cutting face extends along the 11 zero crystal direction and the rear cutting face is arranged along the 100 zero crystal direction, with the crystal orientation deviation angle controlled within ±0.5 degrees.

[0071] Micro-chipping density is defined as the number of local depressions deeper than 0.5 nanometers per micrometer of cutting edge length. It is determined by marking all depressions meeting the depth condition in the cutting edge contour image using an automatic image recognition algorithm, counting their number, and normalizing to the unit length. Chipping depth distribution is characterized by calculating the ratio of the maximum depth to the average depth of all chipping defects; a ratio closer to 1 indicates a more uniform chipping depth distribution. The power spectral density function of surface roughness is obtained by performing a Fast Fourier Transform on the cutting edge surface height data, taking the square of the modulus of the transform result, dividing by the data length, and then normalizing. High-frequency components refer to the portion with spatial frequencies greater than 1.5 periods per micrometer. Their energy percentage is defined as the ratio of the integral of the high-frequency power spectral density to the integral of the full-frequency power spectral density, and this energy percentage is limited to below 15%.

[0072] Based on the aforementioned microscopic geometric feature parameters and corresponding end-face optical performance indicators, a multi-input, multi-output physically constrained neural network model is constructed. This model takes the microscopic geometric feature parameters of the cutting edge as input and the end-face optical performance indicators as output, embedding material fracture mechanics equations and optical waveguide boundary conditions as regularization constraints. The input layer contains 7 neurons, corresponding to radius of curvature, rate of change of curvature gradient, standard deviation of atomic-level flatness, rake face crystal orientation deviation angle, flank face crystal orientation deviation angle, micro-chipping density, and high-frequency power spectral density ratio, respectively. The output layer contains 3 neurons, corresponding to end-face reflectivity, return loss, and mode field distortion coefficient, respectively. Two hidden layers are set, each containing 128 neurons, with modified linear units as the activation function.

[0073] The model training uses mean squared error as the basic loss function. At the same time, two physical constraint losses are introduced: the first physical constraint loss Based on Griffith's fracture theory, the stress intensity factor corresponding to the geometric characteristics of the cutting edge is calculated. Its expression is , For geometric correction factor, For cutting stress, For the equivalent crack length, this The value is mapped to the predicted subsurface damage depth of the fiber end face. , and the measured damage depth The square of the difference constitutes The second physical constraint loss Based on the continuity condition of Maxwell's equations at the medium interface, the predicted end-face height field is required. The tangential component of the electric field is continuous, i.e. , To indicate on the air side, and to indicate on the quartz side, , Where is the dielectric constant. As the normal vector, the continuity condition is approximated as a discrete constraint using finite difference, and its sum of squared residuals constitutes... Total loss function , Set it to 0.3. Set the value to 0.2. The model is trained using an adaptive moment estimator optimizer with an initial learning rate of 0.001, a batch size of 32, and 500 training epochs until the validation set loss converges.

[0074] A target optical performance index threshold range is set, and this threshold range is used as the inverse solution objective of the physically constrained neural network model. The microscopic geometric feature parameters of the input layer are iteratively optimized using a gradient backpropagation algorithm until the optical performance index of the output layer falls within the threshold range, thereby obtaining a set of solution parameters for the cutting edge microscopic geometry that meet the optical performance requirements. The target threshold range is set as follows: end-face reflectivity less than -70 dB, return loss greater than 65 dB, and mode field distortion coefficient less than 3%. The inverse solution employs a constrained projective gradient descent algorithm.

[0075] Initial input parameters are randomly generated, but must meet the following physical feasibility requirements: radius of curvature no less than 2 nm and no more than 50 nm; rate of change of curvature gradient no more than 0.05 nm; standard deviation of atomic-level flatness no more than 0.4 nm; crystal orientation deviation angle no more than 1 degree; micro-fragmentation density no more than 3 points per micrometer; high-frequency power spectral density ratio no more than 20%. In each iteration, the deviation gradient between the current output and the target threshold is calculated, backpropagated to the input layer, and the micro-geometric parameters are updated. Physical feasibility is immediately checked after the update; if any parameter exceeds the feasible range, it is projected back to the boundary value. The iteration terminates when the output indicators are within the target threshold range for 10 consecutive iterations, or when the number of iterations reaches 1000. Finally, all parameter combinations that meet the conditions are retained to form the solution set of the cutting edge micro-geometric parameters.

[0076] The solution set of micro-geometric parameters of the cutting edge is converted into ultra-precision machining path instructions. These instructions control the nano-positioning platform of a single-point diamond lathe to perform fly-cut machining on natural or synthetic single-crystal diamond blanks according to this ultra-precision machining path, forming a cutting edge structure with a specified micro-geometric morphology. The conversion process includes two sub-steps: geometric modeling and trajectory planning. In the geometric modeling stage, based on parameters such as the radius of curvature and the rate of change of curvature gradient from the solution set, the three-dimensional contour curve of the cutting edge is reconstructed. This three-dimensional contour curve exhibits a non-uniform curvature distribution along the length of the cutting edge, ensuring high-precision control in critical areas. In the trajectory planning stage, the contour curve is discretized into a three-dimensional coordinate sequence with a sampling interval of 10 nanometers. The machining path instructions include the three-dimensional coordinate sequence of the tool trajectory, feed rate, depth of cut, and spindle speed.

[0077] The depth of cut employs a progressively decreasing strategy: the initial roughing depth of cut is 500 nm, the intermediate transition depths of cut are 300 nm, 100 nm, 50 nm, and 20 nm respectively, and the final finishing depth of cut is 10 nm. The feed rate is adjusted in conjunction with the depth of cut: the feed rate is 5 μm / min during the roughing stage, decreasing sequentially to 3 μm / min, 1.5 μm / min, 0.8 μm / min, and 0.6 μm / min during the transition stage, and 0.5 μm / min during the finishing stage. The spindle speed is kept constant at 60,000 rpm to maintain a constant cutting linear velocity of 150 m / s, avoiding vibration noise caused by speed fluctuations. During machining, the ambient temperature is controlled at 20°C ± 0.1°C, the humidity is controlled at 40% ± 2%, and the displacement noise of the vibration isolation platform is less than 0.1 nm per hertz.

[0078] After the diamond cutter machining is completed, a verification step is also included: using the diamond cutter to cut new fiber samples, the aforementioned optical performance measurement process is repeated. If the measured optical performance index deviates from the target threshold by more than 5%, the new data is added to the training set, the physical constraint neural network model is fine-tuned online, and the reverse solution and machining process is re-executed until verification is successful. The number of verification samples is no less than 20, and the average value of the optical performance index is used for judgment. The online fine-tuning adopts a transfer learning strategy, freezing the weights of the first two layers of the neural network and only fine-tuning the last layer and the output layer. The learning rate is set to 0.0001, and the number of training rounds is 50 to avoid catastrophic forgetting. The closed-loop mechanism ensures the robustness and long-term consistency of the machining process.

[0079] The above method realizes a design paradigm shift from "experience-driven" to "performance-driven" by establishing a quantitative mapping between the micro-geometric features of the cutting edge and the optical performance of the end face.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machining design method based on a diamond cutter for fiber optic dicing, characterized in that, include: End face samples were collected by cutting standard single-mode optical fibers with multiple diamond cutters of different micro-geometric characteristics. Three-dimensional surface topography scanning is performed on each end face sample to obtain its surface height field data, and its optical performance indicators are measured simultaneously, including end face reflectivity, return loss and mode field distortion coefficient. Microscopic geometric feature parameters of each diamond cutting edge are extracted. These microscopic geometric feature parameters include the radius of curvature of the cutting edge profile, the rate of change of curvature gradient, the atomic-level flatness of the cutting edge tip, the crystal orientation angle of the front and rear surfaces of the cutting edge, the microscopic chipping density and depth distribution of the cutting edge along the cutting direction, and the power spectral density function of the surface roughness of the cutting edge. Based on the microscopic geometric feature parameters and the corresponding end-face optical performance indicators, a multi-input multi-output physical constraint neural network model is constructed. This physical constraint neural network model takes the microscopic geometric feature parameters of the cutting edge as input, the end-face optical performance indicators as output, and embeds the material fracture mechanics equation and the optical waveguide boundary conditions as regularization constraints. A target optical performance index threshold range is set, and the target optical performance index threshold range is used as the inverse solution target of the physical constraint neural network model. The micro-geometric feature parameters of the input layer are iteratively optimized through the gradient backpropagation algorithm until the optical performance index of the output layer falls within the target optical performance index threshold range, thereby obtaining a set of solution sets of micro-geometric parameters of the cutting edge that meet the optical performance requirements. The micro-geometric parameters of the cutting edge are converted into ultra-precision machining path instructions, and the nano-positioning platform of the single-point diamond lathe is controlled to perform flying cutting on the single-crystal diamond billet according to the ultra-precision machining path to form a cutting edge structure with a specified micro-geometric morphology.

2. The machining design method based on a diamond cutter for fiber optic cutting according to claim 1, characterized in that, Three-dimensional surface topography scanning was performed on each end face sample to obtain its surface height field data, and its optical performance indicators were measured simultaneously, including: A white light interferometer was used to scan a circular region with a diameter of 100 micrometers at the center of the fiber end face to obtain surface height field data. The reflectivity of the end face was measured by an optical time domain reflectometer, the return loss was determined by a polarization-independent return loss tester, and the mode field intensity distribution image was obtained by a near-field scanning optical microscope and the mode field distortion coefficient was calculated.

3. The machining design method based on a diamond cutter for fiber optic cutting according to claim 2, characterized in that, The microscopic geometric feature parameters of each diamond cutting edge are extracted, including: The cross-sectional image of the cutting edge tip was obtained by focusing ion beam scanning electron microscopy, and the contour points within a 10-nanometer range of the cutting edge tip were fitted to a quadratic curve to calculate the radius of curvature. The change in radius of curvature per unit distance along the cutting edge length direction is calculated using the sliding window difference method as the rate of change of curvature gradient. The surface height of the five atomic layers at the very tip of the cutting edge was observed by high-resolution transmission electron microscopy, and its standard deviation was calculated as atomic-level flatness. The crystal orientation angles of the front and rear surfaces of the cutting edge were determined using electron backscatter diffraction. The number of local depressions with a depth greater than 0.5 nanometers per micrometer of cutting edge length is counted using an automatic image recognition algorithm as the micro-fracture density, and the ratio of the maximum depth to the average depth is calculated as a characterization of the depth distribution. After performing a fast Fourier transform on the surface height data of the cutting edge, the modulus is squared and normalized to obtain the power spectral density function of the surface roughness, and the energy proportion of high-frequency components with spatial frequencies greater than 1.5 cycles per micrometer is limited to below 15%.

4. The machining design method based on a diamond cutter for fiber optic cutting according to claim 3, characterized in that, Based on the aforementioned microscopic geometric feature parameters and corresponding end-face optical performance indicators, a multi-input multi-output physically constrained neural network model is constructed, including: The input layer is set to contain 7 neurons, corresponding to the radius of curvature, rate of change of curvature gradient, standard deviation of atomic flatness, crystal orientation deviation angle of the rake face, crystal orientation deviation angle of the flank face, micro-fracture density, and proportion of high-frequency power spectral density, respectively. The output layer is configured to contain three neurons, corresponding to end face reflectivity, return loss, and mode field distortion coefficient, respectively. Two hidden layers are set up, and the activation function is the modified linear unit; The stress intensity factor calculation module based on Griffith's fracture theory is introduced into the loss function as the first physical constraint term, and the continuity constraint of the electric field tangential component based on the boundary conditions of Maxwell's equations is introduced as the second physical constraint term.

5. The machining design method based on a diamond cutter for fiber optic cutting according to claim 4, characterized in that, The construction of the first physical constraint term includes: The equivalent crack length and geometric correction factor are calculated based on the cutting edge geometry, and the stress intensity factor is calculated in combination with the cutting stress. The stress intensity factor is mapped to a predicted value of the subsurface damage depth of the fiber end face, and the square of the difference between the predicted value and the actual damage depth constitutes the first physical constraint loss term.

6. The machining design method based on a diamond cutter for fiber optic cutting according to claim 5, characterized in that, The construction of the second physical constraint term includes: Based on the predicted end-face height field, the normal derivative of the tangential component of the electric field at the air-quartz interface is calculated using the finite difference approximation. Based on the difference in dielectric constant, the continuity residual of the tangential component of the electric field is constructed, and its sum of squares constitutes the second physical constraint loss term.

7. The machining design method based on a diamond cutter for fiber optic cutting according to claim 6, characterized in that, A target optical performance index threshold range is set, and this target optical performance index threshold range is used as the inverse solution target of the physical constraint neural network model. The microscopic geometric feature parameters of the input layer are iteratively optimized through a gradient backpropagation algorithm, including: Set the target threshold ranges for end-face reflectivity, return loss, and mode field distortion coefficient; The constrained projective gradient descent algorithm is used for inverse solution; After each iteration, the physical feasibility of the microscopic geometric parameters is verified, and infeasible solutions that are outside the feasible range are eliminated.

8. The machining design method based on a diamond cutter for fiber optic cutting according to claim 7, characterized in that, The solution set of the micro-geometric parameters of the cutting edge is converted into ultra-precision machining path instructions, including: The three-dimensional profile curve of the cutting edge is reconstructed based on the radius of curvature and the rate of change of curvature gradient in the solution set; The three-dimensional profile curve of the cutting edge is discretized into a three-dimensional coordinate sequence; Generate machining path instructions that include a three-dimensional coordinate sequence of the tool path, feed rate, depth of cut, and spindle speed.

9. The machining design method based on a diamond cutter for fiber optic cutting according to claim 8, characterized in that, After the diamond cutting process is completed, a verification step is also included: Use the diamond cutter to cut a new fiber sample and repeat the optical performance measurement process. If the measured optical performance index deviates from the target threshold by more than 5%, new data will be added to the training set, the physical constraint neural network model will be fine-tuned online, and the reverse solution and processing process will be re-executed until the verification is successful.

10. The machining design method based on a diamond cutter for fiber optic cutting according to claim 9, characterized in that, The online fine-tuning employs a transfer learning strategy, freezing the weights of the first two layers of the neural network and adjusting only the last layer and the output layer.

Images