Ultra-thin diffractive lens nanoimprint mass production method and system

Abstract

The application relates to the technical field of nano-imprinting, and discloses a nano-imprinting mass production method and system for ultrathin diffractive lenses, which forms a partitioned dispensing pattern and a process pre-configuration parameter set through structure density field analysis and reverse optimization of partitioned dispensing quantity; two-stage prediction and closed-loop pre-compensation of whole-process alignment drift during pressing are carried out by using a physical information fusion neural network; UV curing and demolding are completed by using a prescription self-adaptive curing path optimization strategy, and full-aperture metering data is output online; online defect identification is completed by using a diffractive pattern difference field driven multi-scale convolution classification network with warpage compensation correction; input and output warehouse permissions and process risk labels are output by comprehensive judgment of multi-dimensional gating rules; and manufacturing data packets containing optical qualification verification fields are generated by reasoning full-aperture optical performance by using a graph neural network. The application can realize multi-prescription mixed-line production and full-process quality closed-loop management and control of nano-imprinting of ultrathin diffractive lenses.

Description

Technical Field

[0001] This invention relates to the field of nanoimprint technology, and more specifically, to a method and system for mass production of ultrathin diffractive lenses using nanoimprint technology. Background Technology

[0002] Ultrathin diffractive prescription lenses have wide applications in vision correction due to their thinness and lightness. Nanoimprint technology is the main process route for mass replication of their diffraction microstructures. The diffraction zone density of ultrathin diffractive lenses spans several orders of magnitude from the inner to the outer ring. The diffraction tooth period of the inner ring is on the micrometer scale, while the diffraction tooth period of the outer ring can be compressed to the nanometer scale. The resin filling mechanism of different zones is fundamentally different. Uniform resin application cannot compensate for the uneven residual layer thickness caused by the difference in resin demand in each zone, making it difficult to guarantee the consistency of the phase substrate across the entire aperture.

[0003] In multi-formula mass production scenarios, the radial gradient of the diffraction structure density field of different formulations varies greatly. During the pressing process, the spatial pattern of in-plane displacement of the substrate driven by the capillary pressure difference of each zone changes fundamentally with the formulation switching. Fixed empirical compensation amount cannot adapt to the alignment drift changes after the formulation switching. In addition, the ultrathin substrate undergoes elastic bending under the pressing pressure. The in-plane displacement of the edge mark caused by bending is superimposed with the actual alignment drift, which further increases the difficulty of alignment control.

[0004] Furthermore, ultra-thin molded parts exhibit full-diameter warpage. This warpage causes a deflection of the normal direction at various locations on the substrate, resulting in non-uniformity of the actual incident angle across the entire diameter. If the theoretical diffraction pattern is calculated directly using the assumption of perpendicular plane wave incidence, a large amount of systematic error caused by warpage will be mixed into the difference field, making it difficult to distinguish from the actual defect signal and reducing the accuracy of online defect detection. Current technologies lack systematic solutions for residual layer homogenization, alignment drift prediction, and warpage compensation defect detection in the aforementioned multi-formula mixed-line scenarios. Summary of the Invention

[0005] This invention provides a method and system for mass production of ultrathin diffractive lenses using nanoimprinting, which solves the technical problems of uneven residual layer thickness, unpredictable alignment drift, and warping interference defect detection in multi-formula mixed-line scenarios in related technologies.

[0006] This invention provides a method for mass production of ultrathin diffractive lenses using nanoimprinting, comprising the following steps: S1. Obtain the prescription optical design file, perform structural density field analysis and inverse optimization iteration of partitioned dispensing amount to form a spatially uniform partitioned dispensing pattern and process pre-configuration parameter set; S2, based on the partitioned dispensing pattern and the pre-configured process parameter set, completes precise partitioned dispensing of adhesive. The physical information fusion neural network is used to perform two-stage prediction and closed-loop pre-compensation for the alignment drift throughout the pressing process, and records the pressing alignment status data. S3, based on the pre-configured process parameter set and pressing alignment status data, adopts a prescription adaptive curing path optimization strategy to complete UV curing and demolding, and outputs the imprinted part and full-diameter measurement data through online metering; S4, based on the embossed parts and full-diameter measurement data, uses a warp-compensated and corrected diffraction pattern difference field-driven multi-scale convolution classification network to complete online defect identification, and generates defect maps and reworkability assessment conclusions; S5 integrates full-caliber measurement data, defect maps, and reworkability assessment conclusions. It comprehensively judges various quality indicators through multi-dimensional gating rules and outputs warehousing permission and process risk labeling. S6 combines the data from adhesive application verification and pressing alignment, full-aperture measurement data, defect maps, and process risk annotations. It then uses a graph neural network to infer the full-aperture diffraction efficiency distribution and the root mean square value of the phase wavefront error, generating a manufacturing data package containing optical conformity verification fields.

[0007] Furthermore, in S1, the structural density field analysis includes: The radial start and end coordinates of each diffraction zone and the period, height and cross-sectional shape parameters of the diffraction teeth are extracted from the prescription optical design file. The total volume of the diffraction groove on the unit substrate projection area of ​​each zone is defined as the structural density value. The structural density values ​​of all zones on the full aperture are organized into a two-dimensional structural density field. For partitions where the diffraction tooth period is greater than the preset period threshold, the residual layer thickness is calculated based on the thin-layer flow theory and the volume conservation of each partition. For partitions where the diffraction tooth period is less than the preset period threshold, the capillary permeation drive mechanism is used as the main mechanism. The capillary permeation drive pressure is calculated based on the diffraction tooth period and contact angle of each partition. The actual filling depth during the pressure holding time is calculated according to the permeation equation based on the groove depth and effective viscosity. The residual layer thickness deviation value is obtained by subtracting the actual consumption volume from the amount of adhesive received by each partition. The calculation results of the two types of partitions are combined and the residual layer thickness deviation field is output uniformly across the entire aperture.

[0008] Furthermore, in S1, the inverse optimization iterative solution of the partitioned dispensing amount includes: The amount of adhesive applied to each diffraction zone is set as an optimization variable. The objective function is to minimize the spatial standard deviation of the residual layer thickness at the full aperture. The constraints are set as follows: the amount of adhesive applied to each zone is non-negative, the total amount is equal to the total amount of adhesive applied in the process setting, and the difference between the amounts of adhesive applied to adjacent zones does not exceed the upper limit of the single accuracy of the dispensing system. The optimal partitioning dispensing pattern is determined by iteratively solving the gradient descent method until convergence.

[0009] Furthermore, in step S1, the process pre-configuration parameter set includes three types of parameters: The first category is the initial value of the alignment compensation coefficient, which estimates the expected alignment drift based on the radial gradient of the substrate thickness and the structural density field; the second category is the pressing pressure distribution parameter, which estimates the pressure requirements of each region based on the spatial distribution of the partitioned adhesive pattern. The third category is the curing path seed parameter, which selects the initial type of the curing path spatial strategy based on the radial gradient characteristics of the structural density field; the process pre-configuration parameter set is stored with the prescription version number and design file hash as dual indexes.

[0010] Furthermore, in step S2, the precise application of adhesive in designated zones includes: The dispensing control module reads the dispensing pattern of each zone and distributes resin droplets according to the optimal dispensing amount for each zone. The actual dispensing amount for each zone is verified online. If the difference from the target value exceeds the dispensing accuracy threshold, compensation dispensing is triggered. After the adhesive is applied, an actual adhesive amount distribution vector is generated and synchronously transmitted to the alignment drift prediction module. During the alignment execution phase, the visual alignment module simultaneously images the multi-scale alignment marks on the working model and the substrate, calculates the superposition error components in translation and rotation using the phase correlation method, and drives the precision platform to converge the superposition error to below the alignment accuracy threshold, thus completing the initial alignment.

[0011] Furthermore, in S2, the backbone structure of the physical information fusion neural network consists of a convolutional neural network branch and a long short-term memory network branch. The convolutional neural network branch uses the partitioned dispensing pattern as a two-dimensional feature map to extract the spatial distribution of resin quantity, and the long short-term memory network branch processes temporal features. The two-stage prediction includes: in the first stage, before the pressing is triggered, the initial superposition error triaxial component, substrate thickness and elastic modulus parameters, partition dispensing amount distribution vector, ambient temperature and the initial value of the alignment compensation coefficient in the process pre-configuration parameter set are received, and the initial pre-compensation amount is output; in the second stage, after the pressing is triggered, the long short-term memory network branch receives the timing signal of the contact force sensor in real time and updates the alignment drift prediction in a rolling manner.

[0012] Furthermore, in S2, the closed-loop pre-compensation includes: During the pressing process, in-plane displacement decoupling is performed simultaneously. The in-plane displacement component caused by bending is calculated according to the thin plate bending equation. The pure rigid body drift is obtained by subtracting the in-plane displacement from the measured displacement. This drift is compared with the time-series prediction value of the physical information fusion neural network. If the deviation exceeds the online update threshold, the pre-compensation amount is dynamically corrected. The training of the physical information fusion neural network incorporates the force balance relationship between the elastic stretching in the base plane and the capillary pressure difference in each zone as a soft constraint term into the loss function. The soft constraint term requires that the gradient distribution of the drift predicted by the physical information fusion neural network in space maintains the same directionality as the capillary pressure gradient distribution calculated from the structural density field.

[0013] Furthermore, in S2, the training of the physical information fusion neural network adopts a two-stage strategy: The first stage uses finite element simulation data for pre-training. The finite element simulation data consists of alignment drift time series data generated by substrate-resin-mold coupled mechanical simulation, covering the structural density field parameter space of different formulations. The second stage uses the alignment error collected after demolding in the actual production process for transfer learning. The training label is decoupled by in-plane displacement of the bending surface to ensure that the label reflects the alignment drift of pure rigid body.

[0014] Furthermore, in S3, the formulation adaptive curing path optimization strategy includes: The curing path seed parameters are read from the pre-configured process parameter set, and the shrinkage stress field of each diffraction zone is calculated using the structural density field and the resin curing shrinkage coefficient. The final alignment residual triaxial component in the pressing alignment state data is read, and stress deflection correction is applied to the shrinkage stress field to obtain the corrected shrinkage stress field. The curing time of each diffraction zone is used as an optimization variable, and the objective function is to minimize the maximum value of the full-aperture warping. A greedy sequence scheduling algorithm is used to solve the spatial allocation scheme of the curing time. When the structural density of the outer ring zone exceeds a preset multiple of the structural density of the inner ring zone, the alternating fan-shaped curing strategy is embedded as a hard constraint into the greedy sequence scheduling algorithm. The outer ring zone is divided into several fan-shaped sub-regions along the circumference. The odd-numbered fan-shaped sub-regions are cured first, followed by the even-numbered fan-shaped sub-regions.

[0015] Furthermore, in S3, the UV curing includes: The UV curing light source is equipped with a multi-zone addressable scanning unit, which applies UV irradiation to each zone sequentially according to the optimized curing time sequence. Each zone adopts a two-segment light intensity curve. The first segment uses a preset low light intensity for pre-gel irradiation, and the second segment uses a fully curing light intensity to complete cross-linking curing. The demolding includes: a precision platform lifting the working mold at a set uniform speed to separate it from the cured structure. The online measurement includes: using a white light interferometer to measure the thickness distribution of the entire aperture to form a thickness deviation field, and using a displacement sensor array to measure the warping distribution of the entire aperture to form a warping field, which serves as the measurement data for the entire aperture.

[0016] Furthermore, in S4, the warpage compensation correction includes: The deflection angle of the base normal direction at each measurement point is calculated from the warp field in the full-caliber measurement data. The deflection angle and the diffraction tooth period at that position are substituted into the grating equation to calculate the diffraction peak direction under the actual incident angle. The theoretical diffraction pattern after warp compensation is obtained through coordinate transformation. The difference field of the diffraction pattern is obtained by subtracting the theoretical diffraction pattern from the measured diffraction pattern pixel by pixel.

[0017] Furthermore, in S4, the input of the multi-scale convolutional classification network is obtained by frequency domain decomposition of the diffraction pattern difference field through prescription adaptive spatial frequency threshold. The first spatial frequency threshold and the second spatial frequency threshold are determined according to the minimum and maximum values ​​of the diffraction tooth period on the full aperture, respectively, and the difference field is decomposed into three component images of high frequency, medium frequency and low frequency as parallel input channels. The multi-scale convolutional classification network adopts a feature pyramid network architecture, outputs the probability distribution of defect type labels, and uses the maximum probability value of the probability distribution of each spatial location as the confidence level. Locations with confidence levels higher than the preset classification threshold are marked as defect points to generate the defect map. The reworkability assessment conclusion is obtained by calculating the remaining reworkable time based on the defect type, the interval between the current time and the UV curing completion time, and the historical imprinting cycle number of the working mold.

[0018] Furthermore, in S5, the multidimensional gating rules include: Thickness consistency gating: if the maximum absolute value of the thickness deviation field does not exceed the first thickness threshold, it passes; warpage gating: if the maximum absolute value of the warpage field does not exceed the first warpage threshold, it passes; defect gating is divided into three levels: the first level checks whether there are granular or scratch-type defects in the optical effective area; the second level counts whether the number of bubble-type and underfilled defects exceeds the first defect number threshold; the third level adopts a layered judgment: the first layer judges whether the composite vector magnitude of the final alignment residual in the pressing alignment state data exceeds the alignment deviation threshold; the second layer judges whether the average deviation of alignment deviation type defects in the defect map exceeds the preset proportion of the alignment deviation threshold. If all departmental controls pass, the entry permit is granted. If there are abnormal events in the contact force timing record, the process risk label is added.

[0019] Furthermore, in S5, if any gate fails and the reworkability assessment conclusion is reworkable, then rework is performed and the process pre-configured parameter set is updated in a targeted manner according to the gate failure type: if the thickness consistency gate fails, then the dispensing amount of the corresponding partition in the partition dispensing pattern is corrected. If the warp gate fails, the spatial distribution characteristics of the warp field are analyzed to diagnose the warp type. If the radially symmetrical distribution is determined to be curing shrinkage stress type warp, the curing path seed parameter is adjusted. If the edge asymmetrical distribution is determined to be demolding stress type warp, the demolding speed setting value is reduced. If the spatial distribution of the warp field is related to the thickness deviation field, it is determined to be residual layer non-uniform warp, and the corresponding partition glue amount is corrected. If the Level 3 defect gate fails, an attribution analysis is performed by comparing the global alignment residual direction in the pressing alignment state data with the spatial distribution direction of the local alignment deviation defect in the defect map. If the directions are consistent, the pre-compensation parameters of the physical information fusion neural network are updated; if the directions are inconsistent, the curing path seed parameters or demolding speed setting value are adjusted.

[0020] Furthermore, in S6, each diffraction zone is used as a node, and each node carries six types of features: actual adhesive application amount, final alignment residual, residual layer thickness deviation, warpage amount, defect mark, and contact force abnormality degree. Edge connections are established according to physical adjacency relationships, and a global convergence node is introduced to form a graph structure. After multi-layer graph convolution propagation of the graph neural network, the diffraction efficiency prediction value output by each zone node is organized into a full-aperture diffraction efficiency distribution. The global convergence node outputs the root mean square value of the full-aperture phase wavefront error, which is compared with the prescription design tolerance to generate an optical qualification verification field and written into the manufacturing data packet. The training of the graph neural network adopts a progressive data accumulation strategy. In the early stage of production, it is pre-trained with finite element optical simulation data. After steady-state production, it is periodically sampled and measured. The optical performance measurement is completed simultaneously in the same batch and serves as the supervision label of the graph neural network. The alignment error measurement is decoupled by in-plane displacement and serves as the training label of the physical information fusion neural network for transfer learning. The weights of the two networks are updated synchronously through incremental learning.

[0021] Furthermore, an ultrathin diffractive lens nanoimprint mass production system is provided for performing the steps in the aforementioned ultrathin diffractive lens nanoimprint mass production method, including: The analysis and optimization module is used to obtain the prescription optical design file, perform structural density field analysis and inverse optimization of the partitioned dispensing amount iterative solution, and form a spatially uniform partitioned dispensing pattern and a set of pre-configured process parameters. The adhesive application and alignment pressing module is used to complete precise adhesive application in zones according to the zoned adhesive pattern and the pre-configured process parameter set. It uses a physical information fusion neural network to perform two-stage prediction and closed-loop pre-compensation for alignment drift throughout the pressing process and records the pressing alignment status data. The curing and metering module is used to complete UV curing and demolding based on the pre-configured process parameter set and pressing alignment status data, and to output the imprinted parts and full-diameter metering data through online metering. The defect identification module is used to identify defects online based on the embossed parts and full-diameter measurement data, using a multi-scale convolution classification network driven by the diffraction pattern difference field with warp compensation correction, and generating defect maps and reworkability assessment conclusions. The quality assessment module integrates full-caliber measurement data, defect maps, and reworkability assessment conclusions. It then uses multi-dimensional gating rules to comprehensively assess various quality indicators and outputs warehousing permission and process risk labeling. The optical verification and data package generation module combines adhesive application verification and pressing alignment status data, full-aperture measurement data, defect maps and process risk labels, and uses graph neural networks to infer the full-aperture diffraction efficiency distribution and the root mean square value of phase wavefront error to generate a manufacturing data package containing optical conformity verification fields.

[0022] The beneficial effects of this invention are as follows: For the mass production scenario of multi-formula mixed production of ultra-thin diffractive lenses, the uneven residual layer thickness caused by the difference in resin demand in each zone is eliminated from the source of adhesive application by using dual-mechanism modeling of structural density field and reverse optimization of adhesive application amount in each zone. The physical information fusion neural network is used to embed the balance relationship between capillary pressure difference and elastic tensile force into the training constraints. Combined with in-plane displacement decoupling, two-stage prediction and closed-loop pre-compensation of alignment drift throughout the pressing process are realized. This solves the problem of the failure of fixed compensation amount caused by the fundamental change of alignment drift spatial pattern after formula switching, and realizes high-precision alignment control in multi-formula mixed production scenarios. By correcting the diffraction pattern through warpage compensation and calculating the difference field, and then performing frequency domain decomposition using the adaptive spatial frequency threshold of the formulation, the systematic detection error introduced by the warpage of ultra-thin molded parts is eliminated, ensuring the accuracy of defect detection in the high-density area of ​​the outer ring. A graph neural network is used to realize cross-formulation process-optical performance mapping by inductive bias using graph structure. Combined with multidimensional gating rules and rework attribution analysis, a full-process quality closed loop is constructed, realizing adaptive optimization of process parameters and online prediction and verification of optical conformity during mass production. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method for mass production of ultrathin diffractive lenses using nanoimprinting according to the present invention; Figure 2 This is a flowchart of a method for mass production of ultrathin diffractive lenses using nanoimprinting according to the present invention. Detailed Implementation

[0024] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0025] At least one embodiment of the present invention discloses a method for mass production of ultrathin diffractive lenses using nanoimprinting, such as... Figures 1 to 2 As shown, it includes the following steps: S1. Obtain the prescription optical design file, perform structural density field analysis and inverse optimization iteration of partitioned dispensing amount to form a spatially uniform partitioned dispensing pattern and process pre-configuration parameter set; The prescription optical design file is processed by a standardized analysis module to extract the following structural parameters: prescription power (including spherical power, cylindrical power, and axial angle), diffraction zone division across the entire aperture (with the radial start and end coordinates of each zone as boundaries), and morphological parameters of the diffraction teeth within each diffraction zone (including period, height, and cross-sectional shape). The analysis module supports the output formats of mainstream optical design software such as Zemax and CODE V, and converts design files of different formats into a unified representation of structural parameters through a format adaptation layer. After format adaptation, the structural density value of each diffraction zone is calculated one by one, defined as the total volume of the diffraction grooves per unit substrate projection area of ​​that zone; using the center coordinates of each zone as a spatial index, the structural density values ​​of all zones across the entire aperture are organized into a two-dimensional structural density field D(x,y). For high-power prescriptions, the value of D(x,y) in the outer region is significantly higher than that in the inner region, and its radial gradient is positively correlated with the prescription power; for low-power prescriptions, the radial gradient of D(x,y) is relatively gentle.

[0026] Based on the structural density field D(x,y), a positive prediction model for the residual layer thickness distribution is established. This model addresses the unique scenario where the diffraction structural density field spans three orders of magnitude in the radial direction by modeling the filling behavior separately into two physical mechanism regions, rather than using a single macroscopic flow model. For the inner and middle zones with diffraction tooth periods greater than 500 nm, the resin filling behavior is framed by the Hele-Shaw thin-layer flow theory, describing the resin filling behavior between the mold and the substrate as a non-uniform viscous flow process modulated by the structural density field. Its core assumption is that, given a global pressing pressure and total adhesive application, the residual layer thickness of each zone is determined by the conservation of resin volume in that zone. That is, the amount of adhesive received by the zone minus the resin volume consumed in filling the diffraction grooves of that zone, with the remaining resin spreading in the pressing gap to form the residual layer. For the outer ring partitions where the diffraction tooth period is compressed to below 500 nm, the gyration radius of the resin molecular chain and the feature size of the groove are on the same order of magnitude. The assumption of continuous medium viscous flow no longer applies. At this time, the dominant mechanism of groove filling becomes capillary permeation drive: the capillary pressure of the resin at the groove inlet is determined by the Young-Laplace relation. The magnitude of the capillary pressure is inversely proportional to the groove width. The narrower the groove, the stronger the capillary driving force. However, at the same time, the effective viscosity of the resin in the nano-confined space is significantly higher than the bulk viscosity due to the confinement of the molecular chain conformation.

[0027] In one embodiment of the present invention, the outer nanoscale partitioning uses a dynamic filling volume conservation method to calculate the residual layer thickness deviation, ensuring that its output is consistent with the inner partitioning as a continuous δt value. This guarantees that gradient calculation can be performed across the entire aperture during inverse optimization; specifically as follows: Based on the diffraction period d of the partition and the contact angle θ between the resin and the mold anti-stick layer, the capillary permeation driving pressure is calculated. Its value is equal to the resin surface tension γ multiplied by twice the cosine of the contact angle, then divided by the groove width; that is, the driving pressure is proportional to 2γcosθ / d. Next, based on the groove depth h and the effective viscosity η_eff of the resin in the nano-confined space (the default value of η_eff is 1.5 to 3 times the bulk viscosity; the specific multiple is determined by the groove width-to-diameter ratio h / d according to the pre-calibrated curve), the capillary permeation driving pressure is calculated using the Washburn permeation equation within the current pressing and holding time t_hold. The actual filling depth h_fill that capillary penetration can reach is calculated. h_fill is equal to the square root of the product of capillary driving pressure, holding time, and effective viscosity, multiplied by half the groove width. Then, the actual resin consumption volume of the groove in this zone is calculated using the ratio of the actual filling depth h_fill to the groove depth h, i.e., the actual consumption volume is equal to the groove cross-sectional area multiplied by h_fill. Finally, the amount of resin received by this zone is subtracted from the actual consumption volume, and the remaining resin spreads in the pressing gap to form a residual layer, thus obtaining the residual layer thickness deviation δt value of this outer ring zone. Wherein, when h_fill is less than h, the groove is not completely filled, which is equivalent to the residual layer of this zone being too thin, and δt is a negative value; when h_fill is equal to h, the groove is completely filled, degenerating into the same volume conservation calculation as the inner ring. The dynamic filling volume conservation calculation results of the outer ring zone are combined with the static volume conservation calculation results of the inner ring zone, and the residual layer thickness deviation field δt(x,y) is uniformly output across the entire diameter as the basis for calculating the objective function of inverse optimization. If h_fill is significantly less than h in a certain outer ring zone during the current holding time, then the zone is marked as having insufficient filling risk in the process pre-configuration parameter set, prompting an appropriate extension of the holding time in step three. Under the assumption of uniform adhesive distribution, high-density zones have thinner residual layers due to higher resin consumption in the grooves, while low-density zones have thicker residual layers due to lower resin consumption in the grooves, thus generating a residual layer thickness deviation field δt(x,y). The forward prediction model performs the above-mentioned interval mechanism calculation on each zone, and under the premise of any partition adhesive distribution scheme, outputs the corresponding full-diameter residual layer thickness deviation field δt(x,y), which serves as the basis for calculating the objective function of the inverse optimization.

[0028] Supported by a forward prediction model, an inverse optimization problem for the dispensing amount in each diffraction zone is constructed. The dispensing amount {q_k} of each diffraction zone is set as the set of optimization variables, and the objective function is to minimize the spatial standard deviation σ(δt) of the full-aperture residual layer thickness. Constraints are set as follows: the dispensing amount of each zone is not negative, the sum of the dispensing amounts of all zones is equal to the total amount of adhesive applied in the process setting, and the difference between the dispensing amounts of adjacent zones does not exceed the single-shot accuracy limit of the dispensing system, so as to ensure the physical feasibility of the adhesive application pattern. The gradient descent method is used to iteratively solve the above-mentioned constrained optimization problem: In each iteration, the current dispensing amount allocation scheme is used as input to run a forward prediction model, calculating the spatial standard deviation of the current δt(x,y). This standard deviation is then used to calculate the gradient of the dispensing amount for each partition, and the dispensing amount for each partition is updated in the opposite direction of the gradient. The updated dispensing amount is then projected onto the constraint set. Iteration continues until the difference between two adjacent objective function values ​​is lower than the convergence criterion, at which point the current dispensing amount allocation scheme is determined as the optimal partition dispensing pattern. The convergence criterion is set to 1% of the thickness gate threshold τ_t from step five; that is, iteration stops when the change in σ(δt) between two adjacent iterations is lower than 1% of τ_t. This setting ensures that the optimization accuracy is directly linked to the final quality gate threshold, avoiding excessive iteration that consumes computation time. This pattern corresponds to a larger amount of adhesive in high-density partitions and a smaller amount in low-density partitions. Its spatial distribution is positively correlated with the shape of the structural density field D(x,y), eliminating the uneven residual layer thickness caused by differences in resin demand in each partition from the source of adhesive application.

[0029] After completing the calculation of the partitioned dispensing pattern, the pre-configured parameter set for the process is calculated simultaneously based on the structural density field D(x,y) and the formulation parameters. This parameter set contains three types of parameters: The first type is the initial value of the alignment compensation coefficient, which estimates the uneven tensile stress generated in the substrate surface by the capillary pressure difference of each partition during the pressing process and the expected alignment drift caused by the in-plane displacement of the substrate relative to the mold, based on the radial gradient of the substrate thickness and the structural density field D(x,y) (which is positively correlated with the level of capillary pressure difference of each partition during the filling stage). This serves as the prior bias input for the physical information fusion neural network inference in step two. The second type is the pressing pressure distribution parameter, which estimates the pressure demand difference of each region during the filling stage based on the spatial distribution of the partitioned dispensing pattern. This is used for the partitioned pressure adjustment and control of the pressing module. The third type is the curing path seed parameter, which selects the initial type of the curing path spatial strategy based on the radial gradient characteristics of D(x,y). For high gradients (height number formulations), an inside-out progressive strategy is selected, and for low gradients (low degree number formulations), a center-radial strategy is selected. This serves as the initial input for the curing path optimization in step three. The pre-configured parameter set of the process is stored with a dual index of prescription version number and design file hash to ensure that the parameter set is strictly bound to the prescription design version. Steps two and three call the required parameter subsets according to the index.

[0030] S2, based on the partitioned dispensing pattern and the pre-configured process parameter set, completes precise partitioned dispensing of adhesive. The physical information fusion neural network is used to perform two-stage prediction and closed-loop pre-compensation for the alignment drift throughout the pressing process, and records the pressing alignment status data. During the adhesive application phase, the dispensing control module reads the partitioned dispensing pattern output from step one, uses the center coordinates of each diffraction partition as the target position, and precisely distributes resin droplets according to the optimal amount of adhesive for each partition. The dispensing system arranges the resin partitions on the substrate surface via piezoelectric inkjet or pneumatic jetting. The actual dispensing amount for each partition is verified online using a weight sensor or image area measurement. If the difference between the actual dispensing amount and the target value for any partition exceeds the dispensing accuracy threshold, compensation dispensing is triggered until the deviation is eliminated. The dispensing accuracy threshold is determined by the single-jet volume accuracy of the dispensing system, with a default value of 5% of the target dispensing amount for each partition. This proportion ensures that the residual layer thickness deviation introduced by the dispensing deviation does not exceed 10% of the thickness gate threshold τ_t in step five. After dispensing is completed, the actual dispensing amount for each partition is verified online to form an actual dispensing amount distribution vector. This vector is synchronously transmitted to the subsequent alignment drift prediction module as a core input feature. The actual amount of adhesive applied to each zone directly determines the magnitude of the capillary force in that zone during the lamination process. The spatial distribution pattern of the adhesive application determines the spatial pattern of the capillary pressure difference, which in turn determines the direction and amplitude of the alignment drift. Therefore, adhesive application and alignment lamination constitute an inseparable process unit in the data stream. The substrate is then transferred to the alignment and lamination station.

[0031] During the alignment execution phase, the vision alignment module simultaneously images the multi-scale alignment marks on the working model and the substrate. The alignment marks employ a three-layer nested structure: the outer layer consists of millimeter-scale crosshair marks for large-scale coarse alignment; the middle layer consists of hundred-micrometer-scale concentric ring marks for medium-precision alignment; and the inner layer consists of micrometer-scale barcode pattern marks for fine alignment and rotation angle detection. After acquiring the alignment mark images, the vision system uses phase correlation to achieve sub-pixel-level registration between the center of the working model's mark and the center of the substrate's mark. It calculates the superposition error components in translation (x-direction, y-direction) and rotation (θ-direction) between the two, driving the precision platform to adjust in three degrees of freedom, converging the superposition error to below the set alignment accuracy threshold, thus completing the initial alignment. During initial alignment, the pressing pressure is zero and the substrate is not bent; the superposition error measured by vision is the pure rigid body alignment error, directly used as input for subsequent neural network inference, along with the ambient temperature at the time of image acquisition.

[0032] The physical information fusion neural network is activated before the pressing is triggered, and performs the first stage of inference: This stage only uses the static features of the convolutional neural network branch, and receives the triaxial components of the initial superposition error (from the initial alignment result, at which time the pressing pressure is zero, the substrate is not bent, and the visual measurement value is the pure rigid body alignment error, without the need for bending decoupling), substrate thickness and elastic modulus parameters (from the substrate material database, indexed by material batch), partitioned glue distribution vector (from the actual verification value of the glue application stage in this step, with the partitioned glue pattern in step one as the target reference), ambient temperature (from real-time sensor readings), and the initial value of the alignment compensation coefficient in the process pre-configuration parameter set; the long short-term memory network branch in this stage uses the contact force time-series average value in the historical production data of similar prescriptions as the initial hidden state input, and merges it with the output of the convolutional network branch in the fully connected layer, outputting the initial pre-compensation amount before the pressing is triggered, which is used to apply the reverse pre-compensation displacement through the precision platform before pressing, so that the substrate is in a state of intentional offset relative to the working mode at the initial contact moment. The backbone of the network consists of a convolutional neural network and a long short-term memory network: the convolutional network branch processes two-dimensional spatial distribution features (extracting the spatial distribution of resin amount in each partition as a two-dimensional feature map, reflecting the influence of the structural density field difference of different formulations on capillary force distribution); the long short-term memory network branch processes temporal features, enters the second stage of inference after the pressing begins, receives the temporal signals of the contact force sensor in real time, models the temporal evolution of capillary force and viscous pressure, and updates the alignment drift prediction in a rolling manner.

[0033] The physical information fusion of the network is reflected in the explicit constraint embedding of the unique physical process of ultrathin substrate lamination during training. In this scenario, the main driving force of alignment drift is the non-uniform tensile stress generated in the substrate surface by the capillary pressure difference between different zones during resin filling, which drives the substrate to undergo in-plane displacement relative to the mold: the capillary pressure of each zone is proportional to the structural density of that zone (the denser the grooves, the greater the capillary force), and the capillary pressure difference between different zones forms a tensile stress gradient in the substrate surface, driving the substrate to undergo in-plane displacement relative to the mold. This mechanism makes the spatial distribution of alignment drift directly related to the radial gradient of the structural density field D(x,y), and the radial gradient of D(x,y) differs significantly for different formulations, which is the core difficulty in alignment drift prediction in multi-formulation mixed-line scenarios. With each change of formulation, the spatial pattern of drift changes fundamentally and cannot be handled with fixed empirical compensation amounts. To address this mechanism, the force balance relationship between the elastic tensile amount in the substrate surface and the capillary pressure difference in each zone is incorporated as a soft constraint term into the loss function. This constraint term requires that the gradient distribution of the drift predicted by the network in space maintains the same directionality as the capillary pressure gradient distribution calculated from the structural density field, preventing the network from generating physically unreasonable drift predictions in prescription intervals with sparse training data. The network training adopts a two-stage strategy: the first stage uses finite element simulation data (alignment drift time-series data generated by substrate-resin-mold coupled mechanical simulation, covering the structural density field parameter space of different prescriptions) for pre-training, enabling the network to have preliminary cross-prescription prediction capabilities before being put into production; the second stage uses the alignment error measured after demolding collected during actual production for transfer learning, where the training labels need to undergo in-plane displacement decoupling processing (the decoupling method is the same as the decoupling method for real-time measurement during the pressing process described below), ensuring that the labels reflect pure rigid body alignment drift rather than a mixture of bending and drift, so that the network prediction accuracy continuously improves with the accumulation of production data.

[0034] In one embodiment of the present invention, addressing the unique problem of elastic bending occurring during the lamination process of ultrathin substrates with a total thickness of less than 1 mm, the in-plane displacement decoupling is performed synchronously each time the vision system measures the real-time drift during lamination. The decoupled pure rigid body drift is then compared with the time-series prediction value of the neural network to drive closed-loop correction. Under lamination pressure, the ultrathin substrate bends, causing in-plane displacement of the alignment marks at the substrate edge. This displacement is superimposed on the rigid body translation of the substrate relative to the mold (i.e., the actual alignment drift). If not decoupled, the closed-loop update mechanism will misjudge the edge displacement caused by bending as alignment drift and apply incorrect compensation. The decoupling process is as follows: Based on the substrate's elastic modulus E, Poisson's ratio ν, thickness h_s, and the current contact force sensor reading (used as a proxy for the current pressing pressure P), the bending deflection field of the substrate under the current pressure is calculated according to the thin plate bending equation. Then, the in-plane displacement component at the edge marker location caused by bending is calculated. The real-time displacement of the edge marker measured by the vision system through the transparent observation window is subtracted from the aforementioned in-plane displacement caused by bending to obtain the pure rigid body drift after eliminating the bending effect. This pure rigid body drift is compared with the time-series prediction value of the neural network to determine whether to trigger an online update. For a typical substrate with a diameter of 50 mm and a thickness of 0.8 mm, the bending deflection difference between the center and the edge under the pressing pressure can reach the order of several micrometers, and the corresponding in-plane displacement of the edge marker can reach the order of sub-micrometers, which is on the same order of magnitude as the alignment accuracy requirement. Therefore, this decoupling step is necessary to ensure the alignment accuracy of the high-density outer ring area.

[0035] After pressing is triggered, the Long Short-Term Memory (LSTM) network branch enters the second stage of inference, starting to receive the temporal signals from the contact force sensor in real time. These signals are then merged with the features from the convolutional network branch to continuously update the alignment drift prediction. The vision system continuously captures alignment mark images of the substrate edge region through a transparent observation window at the edge of the working model. At fixed time intervals, it measures the actual drift and performs in-plane displacement decoupling to obtain the pure rigid body drift, which is then compared with the temporal prediction value from the neural network. If the deviation between the measured pure rigid body drift and the predicted value exceeds the online update threshold, the pre-compensation amount is dynamically corrected, driving the platform to fine-tune, forming a closed-loop rolling update mechanism of prediction-measurement-correction. The default value of the online update threshold is 50% of the alignment accuracy threshold. This setting ensures that correction is triggered before the measured deviation accumulates to the point of affecting the final alignment accuracy, while avoiding frequent invalid adjustments caused by measurement noise. After filling is complete and the alignment error falls within the gating threshold, the current pressing alignment status data (including the final superposition error triaxial components, the entire contact force temporal record, and the pre-compensation amount history) is recorded as input for steps three and six.

[0036] S3, based on the pre-configured process parameter set and pressing alignment status data, adopts a prescription adaptive curing path optimization strategy to complete UV curing and demolding, and outputs the imprinted part and full-diameter measurement data through online metering; Curing path optimization is completed before UV irradiation. Curing path seed parameters are read from the pre-configured process parameter set to obtain the initial value of the strategy type corresponding to the current formulation (either inside-outward progressive or center-radial) and the curing priority sorting seed for each zone. Subsequently, based on the structural density field D(x,y) and the resin curing shrinkage coefficient (obtained from the material database by resin batch index), the shrinkage stress field σ(x,y) of each diffraction zone in the fully cured state is calculated: the magnitude of the local stress in each zone is proportional to the product of the resin volume of that zone (i.e., the product of the structural density value and the zone area) and the curing shrinkage coefficient. The direction of the local stress is determined by the zone geometry (circular zones generate circumferential shrinkage stress, and irregular zones need to be decomposed into radial and tangential components according to their actual shape).

[0037] Based on the shrinkage stress field σ(x,y), the final alignment residual triaxial component from the compression alignment state data output in step two is first read, and an alignment residual correction is applied to the shrinkage stress field. The alignment residual causes a slight misalignment of the diffraction structure of each zone relative to the substrate. This misalignment alters the constraint boundary conditions of the resin-substrate interface during curing shrinkage: in the alignment residual direction, the interface constraint stiffness of the resin shrinkage decreases due to the structural misalignment, causing a deflection of the shrinkage stress release path in that direction. The correction method involves superimposing a stress deflection correction amount related to the alignment residual direction onto the shrinkage stress direction component of each zone. The magnitude of this correction amount is proportional to the product of the absolute value of the alignment residual and the structural density of that zone. Zones with larger alignment residuals and higher structural density exhibit more significant changes in interface constraints and more pronounced stress deflection. The corrected shrinkage stress field σ'(x,y) replaces the original σ(x,y) as the input for subsequent curing path optimization. For typical cases where the alignment residual is within the submicron range, this correction is a second-order small quantity; however, for ultrathin substrates with a total thickness of less than 1 mm, the substrate stiffness is extremely low, and a small stress direction deflection can be amplified in the warping field. Therefore, this correction has a marginal improvement effect on the warping control of height number prescriptions.

[0038] Based on the corrected shrinkage stress field σ'(x,y), a curing path optimization problem is established: the curing time of each diffraction zone is used as the optimization variable, and the objective function is to minimize the maximum value of the full-aperture warpage. The relationship between the warpage and the curing time sequence is described by a thin-plate bending mechanics model: the first-cured zone provides constraints on the shrinkage of the later-cured zone, and the degree of constraint is related to the distance and stiffness ratio between the two zones. The final warpage distribution w(x,y) is calculated using the thin-plate model under a given curing time sequence. Since there are many zones, exhaustively arranging all curing times is not feasible in terms of computation time. A greedy sequence scheduling algorithm is used to solve for an approximate optimal solution: starting from the strategy type specified by the seed parameter, at each step, the next curing zone that reduces the current warpage prediction value the most is selected and added to the curing sequence, until all zones are included in the curing sequence, thus obtaining the spatial allocation scheme of the curing times.

[0039] For high-density formulations (high outer ring structural density, stress concentration in the outer ring), optimization results typically indicate a progressive curing sequence from the inside out: the low-density inner ring sections cure first, forming an inner ring support framework with a certain rigidity, followed by the sequential curing of the high-density outer ring sections. Their radial shrinkage stress is released outwards to the unconstrained direction, rather than compressing the already cured inner ring structure inwards, thus suppressing warping caused by compressive buckling of the inner ring. However, for the high-density annular outer ring sections, in addition to radial shrinkage stress, there is also a significant tangential shrinkage stress: when the entire annular outer ring section cures simultaneously, circumferential shrinkage creates uniform compressive stress in the circumferential direction. If this compressive stress exceeds the bonding strength between the resin and the substrate interface, circumferential cracks may occur in the outer ring.

[0040] In one embodiment of the present invention, for the problem of tangential stress concentration in the outer ring high-density annular partition, the sector alternating solidification strategy is embedded as a hard constraint into the greedy sequence scheduling algorithm: when the structural density of the outer ring partition exceeds twice the structural density of the inner ring partition, the greedy algorithm restricts the candidate set of the outer ring partition to the next sector sub-region in the current odd-even alternation sequence, rather than all outer ring partitions, when selecting the next solidification partition; the greedy algorithm selects the partition that reduces the current warping prediction value the most from this constrained candidate set and adds it to the solidification sequence, thereby optimizing radial warping (soft target) while ensuring tangential stress dispersion (hard constraint). Specifically, the outer ring is divided into several sector-shaped sub-regions along the circumference (the default number of sector-shaped sub-regions is 8). Odd-numbered sector-shaped sub-regions are cured first, followed by even-numbered sub-regions. During curing shrinkage, the odd-numbered sectors cured first have uncured resin on both sides, allowing for free tangential shrinkage. During curing, the even-numbered sectors cured later have already cured odd-numbered sectors on both sides providing dispersed constraint points, dispersing tangential stress into multiple local constraints rather than concentrated constraints across the entire ring, thus reducing the peak tangential stress below the interfacial bonding strength. For low-density formulations (outer ring density not exceeding twice that of the inner ring, with a relatively uniform stress field), alternating sector hard constraints are not triggered. The greedy algorithm freely selects from the entire candidate partition set, resulting in an optimization order close to a central radial uniform curing sequence, maintaining the symmetry of stress release in all directions and suppressing the generation of asymmetric warping components.

[0041] The UV curing light source is equipped with a multi-zone addressable scanning unit. Following an optimized curing time sequence, it sequentially applies UV irradiation to each zone according to their order in the sequence. The UV irradiation of each zone employs a two-stage intensity curve: the first stage uses low intensity (default value is 30% of the full curing intensity) for pre-gelation irradiation, the duration of which is equal to the gelation time of the resin formulation under this low intensity condition (gelation time is obtained from the material database by resin batch and irradiation intensity index, with calibration conditions consistent with the first stage irradiation intensity). This allows the resin to reach its gel point while retaining some fluidity, allowing the curing shrinkage stress to be spontaneously released through resin flow before the gel point. The second stage uses full curing intensity to complete cross-linking and cure, locking in the final shape. Compared to single-stage full-power irradiation, the two-stage intensity curve reduces residual stress at the curing lock-in moment, thereby further reducing warpage. The total UV irradiation dose for each zone is set based on the exposure required for complete resin curing (indexed from the material database), and the irradiation dose is controlled in a closed loop by a light power sensor. After all zones have been irradiated with UV and the resin has been fully cured, the demolding operation is performed: the precision platform separates the working mold from the cured structure at a set uniform lifting speed. The lifting speed is determined by the anti-stick layer status parameters (the higher the degree of anti-stick layer degradation, the slower the lifting speed, so as to avoid excessive demolding force that could damage the structure). After demolding, the stamped part is transferred to the online metering station.

[0042] The online metrology module performs two full-diameter measurements on the embossed part: First, it uses a white light interferometer to measure the thickness distribution t(x,y) of the entire diameter using a matrix scanning method, with a sampling density of no less than one measurement point per millimeter. The difference between the thickness value at each measurement point and the target thickness t* is organized into a thickness deviation field. Second, it uses a displacement sensor array to simultaneously measure the warpage distribution w(x,y) of the entire diameter. Warpage is defined as the difference between the measured height at each measurement point and the average height of the entire diameter, with the maximum warpage representing the degree of warpage of the embossed part. After metrology is completed, the t(x,y) deviation field and the w(x,y) warpage field are numerically stored and written into the process log along with the pressing alignment status data, as part of the process feature set of step six.

[0043] S4, based on the embossed parts and full-diameter measurement data, uses a warp-compensated and corrected diffraction pattern difference field-driven multi-scale convolution classification network to complete online defect identification, and generates defect maps and reworkability assessment conclusions; The diffraction feature detection unit uses the diffraction optical properties of the embossed part itself as the physical medium for defect sensing. The detection optical route consists of a standard collimated light source, a support platform for the part under test, and a far-field diffraction pattern acquisition camera: the collimated light source illuminates the part under test in a direction perpendicular to the substrate surface, and the incident light is diffracted to a specific direction by the diffraction microstructure; the far-field acquisition camera captures the intensity and phase distribution of the diffracted light field at a fixed position away from the part under test, forming a measured diffraction pattern Im(x', y') (x', y' are the camera image plane coordinates). Due to the full-aperture warping of ultra-thin embossed parts with a total thickness of less than 1 mm, the warping causes the normal direction at various positions on the substrate to deflect relative to the vertical direction, resulting in non-uniformity of the actual incident angle across the entire aperture. If the theoretical diffraction pattern is calculated directly using the assumption of perpendicular plane wave incidence, a large amount of systematic error caused by warping will be mixed into the difference field, making it difficult to distinguish from the real defect signal.

[0044] In one embodiment of the present invention, the calculation of the theoretical diffraction pattern incorporates a warp compensation correction step. First, from the structural parameters of each diffraction zone obtained in step one, the periodic values ​​of the diffraction teeth of each zone are mapped to the coordinate system of the molded part surface according to the radial start and end coordinates of each zone, forming a full-aperture diffraction period distribution map d(x,y); the coordinate mapping adopts the same transformation relationship as the subsequent difference field coordinate transformation to ensure that d(x,y) is aligned with the detection coordinate system. Second, starting from the full-aperture warp field w(x,y) obtained by online measurement in step three, the deflection angle of the base normal direction relative to the vertical direction at each measurement point is calculated. The magnitude of the deflection angle is equal to the arctangent value of the local gradient of the warp field at that point. Then, the deflection angle at each object plane coordinate (x, y) and the diffraction tooth period d(x, y) at that position are substituted into the grating equation to calculate the diffraction peak direction at that position under the actual incident angle. The grating equation is: the sine of the diffraction peak direction is equal to the sine of the incident angle plus the quotient of the detection wavelength divided by the diffraction period d(x, y) at that position, where the incident angle is determined by the deflection angle. The diffraction peak directions at each object plane coordinate are mapped to the camera image plane coordinates through coordinate transformation, resulting in the warp-compensated point-by-point theoretical diffraction pattern It_corrected(x', y'). Finally, the difference between It_corrected(x', y') and the measured pattern Im(x', y') is calculated pixel-by-pixel to obtain the diffraction pattern difference field ΔI(x', y') after eliminating warp system errors. For molded parts with maximum warpage on the order of micrometers, the normal deflection angle of the edge region can reach the order of 0.1 degrees. For a diffraction structure with a period of 100 nanometers, the diffraction peak displacement caused by this deflection angle is on the same order of magnitude as the difference field signal of granular defects. Therefore, warpage compensation is a necessary step to ensure the accuracy of defect detection in the high-density area of ​​the outer ring.

[0045] By transforming coordinates, the difference field ΔI(x', y') is mapped from the camera image plane back to the coordinates of the molded part, resulting in the difference field Δφ(x, y) corresponding to the spatial position of the molded part. The difference field Δφ(x, y) carries structural deviation information at various positions on the molded part. However, different types of defects exhibit different spatial frequency characteristics in the difference field. Therefore, a two-dimensional Fourier transform is performed on Δφ(x, y) to extract its frequency domain representation, and the frequency domain components are decomposed into three intervals according to the prescription-adaptive spatial frequency threshold.

[0046] In one embodiment of the present invention, the spatial frequency threshold is determined by a prescription-adaptive method rather than a fixed value. The reason is that the period span of the diffraction structure from the inner circle to the outer circle is different for different prescriptions. If a fixed threshold is used, the high-frequency diffraction signal generated by the 100-nanometer periodic structure in the outer circle itself will be misjudged as the high-frequency difference component of the particulate defect, resulting in a significant increase in the false alarm rate as the prescription degree increases. The steps for determining the adaptive spatial frequency threshold of the prescription are as follows: From the diffraction zone structure parameters obtained in step one, extract the minimum value d_min (corresponding to the densest outer ring zone) and the maximum value d_max (corresponding to the sparsest inner ring zone) of the diffraction tooth period across the entire aperture; the first spatial frequency threshold f_1 is set to 1 / d_min multiplied by a scaling factor α less than 1 (the default value of α is 0.5), making f_1 lower than the inherent spatial frequency of the normal outer ring diffraction structure, thereby excluding the normal outer ring structure signal from the high-frequency defect judgment range; the second spatial frequency threshold f_2 is set to 1 / d_max multiplied by a scaling factor β greater than 1 (the default value of β is 2.0), making f_2 higher than the inherent spatial frequency of the normal inner ring diffraction structure, thereby excluding the normal inner ring structure signal from the low-frequency system error judgment range. f_1 and f_2 determined in this way are stored in the process pre-configured parameter set along with the prescription version number and are automatically updated each time the prescription is switched, without the need for manual parameter adjustment. High-frequency components (spatial frequencies above f_1) correspond to drastic local changes and are associated with the point-like or linear physical morphology of particulate contamination or scratch-type defects; mid-frequency components (spatial frequencies between f_1 and f_2) correspond to continuous changes at medium scales and are associated with the sheet-like physical morphology of bubble or resin-insufficient defects; low-frequency components (spatial frequencies below f_2) correspond to gradual overall changes and are associated with the gradual characteristics of systematic residual layer deviations or alignment errors. Frequency domain decomposition decomposes Δφ(x,y) into component images of three different frequency bands, providing multi-scale input for the subsequent classification network.

[0047] The multi-scale convolutional classification network adopts a feature pyramid network architecture, using the three frequency component images of Δφ(x,y) as parallel input channels to simultaneously extract defect features at three different spatial receptive field scales. The network structure consists of a backbone feature extraction module and a multi-scale feature fusion module. The backbone module is composed of multiple convolutional and pooling operations, extracting spatial features from the input difference map layer by layer, from fine to coarse-grained. The feature pyramid fusion module merges feature maps of different levels through lateral connections, so that each output position simultaneously integrates information from a small receptive field (suitable for detecting point defects such as particles / scratches) and a large receptive field (suitable for detecting sheet defects such as bubbles / insufficiency). Finally, the classification head outputs the probability distribution of five types of defect labels (particles, bubbles, scratches, insufficiency, and alignment deviation) and background labels at each spatial position of the fused feature map. The maximum probability value in the probability distribution at that position is taken as the confidence level of that position, and the defect type corresponding to the maximum probability is taken as the candidate type. When the confidence level is higher than the preset classification threshold, the position is marked as the defect point corresponding to the candidate type. The type labels, object coordinates, and confidence levels of all defect points are organized into a defect map. The network training data comes from defect samples that have been manually or scanned electron microscope confirmed and labeled in historical production. The training set is expanded with data augmentation (rotation, brightness perturbation, noise injection) to improve the network's robustness to changes in defect morphology under different imaging conditions.

[0048] After obtaining the defect map, a reworkability assessment is performed. The reworkability assessment is based on the following three types of information: First, the type of defect in the defect map. Particulate defects (located at the mold-resin interface) can usually be removed by solvent-assisted peeling, while scratch-type defects (substrate surface damage) are beyond the reworkability range. Second, the interval between the current time and the completion time of UV curing. The shorter the curing time, the lower the degree of resin crosslinking, the easier the solvent peeling, and the wider the reworkability time window. Third, the historical number of stamping cycles of the working mold. The anti-stick layer degrades with the increase of the number of cycles, the contact angle decreases, and the adhesion between the mold and the resin increases, causing the reworkability time window to narrow with the increase of the number of mold cycles. This degradation state is calculated from the number of cycles in the mold usage record according to the pre-calibrated degradation curve. Based on the above three types of information, calculate the remaining reworkable time for the current molded part (i.e., the maximum time after the current moment when solvent stripping can still be performed without damaging the substrate). If the remaining reworkable time is greater than the minimum time required for rework, the evaluation conclusion is "reworkable," and a recommended stripping solvent type and soaking time are given; otherwise, the evaluation conclusion is "not reworkable." The reworkability evaluation conclusion, along with the defect map, is used as input for step five.

[0049] S5 integrates full-caliber measurement data, defect maps, and reworkability assessment conclusions. It comprehensively judges various quality indicators through multi-dimensional gating rules and outputs warehousing permission and process risk labeling. The multi-dimensional gating judgment module simultaneously receives the full-diameter thickness deviation field and warping field data output from step three, as well as the defect map and reworkability assessment conclusion output from step four. After performing independent judgments on the four types of indicators, it comprehensively draws a quality conclusion.

[0050] Thickness Consistency Gating: Calculate the thickness deviation field max|t(x,y)-t*|, which is the maximum absolute value of the thickness deviation across the entire diameter. If this value does not exceed the first thickness threshold τ_t, the thickness gating is deemed successful. τ_t is calculated from the optical design tolerance of each formulation. The calculation method is as follows: the phase basis error introduced by the thickness deviation δt is equal to δt multiplied by the difference between the resin refractive index and the air refractive index, and then divided by the detection wavelength. This phase error does not exceed the specified proportion of the allowable wavefront error budget of the formulation (default value is 20%). τ_t is then calculated from this and stored according to the formulation version number index. Warpage Gating: Calculate the warpage field max|w(x,y)|, which is the maximum absolute value of the warpage across the entire diameter. If this value does not exceed the first warpage threshold τ_w, the warpage gating is deemed successful. τ_w is determined by the allowable range of the maximum deformation of the molded part due to the assembly process. The default value of τ_w is 5 micrometers.

[0051] Defect gating is divided into three levels: Level 1 defect gating checks whether there are granular or scratch-type defects located in the optically effective area (within a specified radius centered on the optical aperture) in the defect map. If present, Level 1 defect gating fails. Level 2 defect gating counts the number of bubble-type and underfilled defects within the optically effective area. If the number exceeds the first defect quantity threshold N_d, Level 2 defect gating fails. N_d is determined as follows: the impact of a single bubble-type or underfilled defect on diffraction efficiency is approximately the ratio of the defect's occlusion area to the total area of ​​the optically effective area multiplied by the nominal diffraction efficiency of that zone. The sum of the diffraction efficiency losses of N_d defects is set to not exceed the maximum allowable diffraction efficiency loss tolerance Δη_max of the prescription (the default value of Δη_max is 5% of the prescription's nominal diffraction efficiency). Thus, N_d is equal to the quotient of Δη_max divided by the average diffraction efficiency loss of a single defect, rounded down, and stored in categories according to prescription version number and average defect size. Level 3 defect gating uses a layered judgment mechanism to address alignment deviations. The first layer is a rapid screening of global alignment residuals: directly read the three-axis components of the final alignment residuals recorded in step two, calculate the magnitude of its composite vector, and if the magnitude exceeds the alignment deviation threshold τ_align, the level three defect gate is directly judged as failing, without waiting for the defect map analysis results in step four; the physical basis for this layer's judgment is that the alignment residuals in step two reflect the global rigid body alignment error between the substrate and the mold at the end of pressing. If this error has exceeded the tolerance, the overall offset of the diffraction structure on the molded part is unacceptable. The second layer is a fine screening for local alignment deviations: If the first layer passes (global rigid alignment error is within tolerance), then the full-aperture distribution of alignment deviation defects in the defect map of step four is further statistically analyzed within the optically effective area. The average deviation at the marked position of the alignment deviation defect is calculated. If this average exceeds 80% of τ_align, then the level three defect gating is deemed a failure. This layer captures local non-rigid alignment shifts introduced during curing shrinkage and demolding. These shifts cannot be covered by the global measurement in step two and can only be detected by the diffraction pattern difference field detection in step four. τ_align is determined by the optical design tolerance of the formulation for alignment errors and is stored indexed by the formulation version number.

[0052] The comprehensive judgment logic is as follows: If all three gates—thickness gate, warpage gate, first-level defect gate, second-level defect gate, and third-level defect gate—pass, the part is deemed acceptable for warehousing. Under this condition, the system additionally checks the contact force timeline recorded in step two for any abnormal events (an abnormal event is defined as the rate of change of contact force within any time window exceeding three times the standard deviation of the historical average of the formulation, or the absolute value of contact force exceeding the specified deviation range of the pressing pressure setting). If an abnormal contact force event exists, the molded part is still deemed acceptable for warehousing, but a process risk labeling field is added to the process log, indicating the time interval and corresponding partition coordinate range of the abnormal event. This information is used as additional risk feature input for the graph neural network in step six when predicting optical performance, and is prioritized for inclusion in the sampled test batch selection to verify whether the abnormal contact force actually affects optical performance. The molded part proceeds to step six. If any gate fails and the reworkability assessment conclusion is "reworkable," it is determined to be reworked; if any gate fails and the reworkability assessment conclusion is "non-reworkable," it is determined to be scrapped.

[0053] When the rework path is executed, the system performs solvent-assisted resin layer peeling on the molded part according to the peeling method recommended in step four. After peeling is completed, the substrate is cleaned. After confirming that the surface cleanliness of the substrate meets the requirements for re-imprinting, the substrate is re-arranged into the production queue of step two. Simultaneously, the pre-configured process parameter set is updated based on the gate type of this failure: If the failure is due to thickness consistency gate, the dispensing amount of the corresponding partition in the partition dispensing pattern is corrected, and the correction amount is written into the parameter set of this formulation version; if the failure is due to warpage gate, the warpage type is first diagnosed by analyzing the spatial distribution characteristics of the warpage field w(x,y): if the warpage is radially distributed with the lens center as the axis of symmetry (bowl-shaped or saddle-shaped), it is determined to be curing shrinkage stress type warpage, and the response is to adjust the strategy weight of the curing path seed parameter; if the warpage is asymmetrically distributed with local edge warping, it is determined to be demolding stress type warpage, and the response is to reduce the demolding speed setting value; if the spatial distribution of the warpage field is highly correlated with the thickness deviation field of this failure (the area with large warpage is exactly the area with large thickness deviation), it is determined to be residual layer non-uniformity type warpage, and the response is to correct the dispensing amount of the corresponding partition; if the failure is due to first-level or second-level defect gate, the process characteristics of this failure (contact force) are updated. The curve and ambient temperature are marked as high-risk states and included in the training samples for the next iteration of the neural network. If the system fails due to the level 3 defect gating (alignment deviation), it compares the global alignment residual direction in step 2 with the spatial distribution direction of the local alignment deviation defect in step 4 for attribution analysis. If the two directions are consistent, the deviation is mainly determined to be due to insufficient alignment in the pressing stage. The current alignment deviation is used as a correction to update the pre-compensation parameters of the physical information fusion neural network in step 2. At the same time, the contact force timing and alignment drift record of this pressing process are included in the training samples of the neural network transfer learning to improve the alignment drift prediction accuracy of subsequent similar prescriptions. If the two directions are inconsistent (for example, the residual in step 2 points to a certain translational direction, but the local deviation in step 4 is concentrated in the outer circle and is radially distributed), the deviation is mainly determined to be due to curing shrinkage or local displacement introduced in the demolding stage. The response is to adjust the curing path seed parameter in step 3 or reduce the demolding speed setting. The updated parameter set is annotated with this rework record, and the version number is incremented to ensure the traceability of parameter updates. All judgment conclusions, detailed gating data, and rework operation records are written into the process log, serving as the source of content for the manufacturing data package in step six.

[0054] S6 combines the data of adhesive application verification and pressing alignment status, full-aperture measurement data, defect spectrum and process risk labeling, and uses graph neural network to infer the full-aperture diffraction efficiency distribution and the root mean square value of phase wavefront error to generate a manufacturing data package containing optical conformity verification fields. For each molded part that is approved for warehousing in step five, process parameter records generated throughout steps one through five are collected from the process log to form a complete set of process features: including the actual amount of adhesive applied to each diffraction zone (from the adhesive application verification record in step two), the triaxial components of the final alignment residual (from the pressing alignment status record in step two), the residual layer thickness deviation of each zone (sampled from the thickness deviation field in step three according to the zone coordinates), the warpage of each zone (sampled from the warpage field in step three according to the zone coordinates), and the defect markings of each zone (merged from the defect map in step four according to the zone range). In addition, if a process risk labeling field is added during the warehousing approval determination in step five, the zone coordinate range and the degree of abnormality (the multiple by which the rate of change of contact force deviates from the historical average) corresponding to the abnormal contact force event are written as a sixth type of feature into the corresponding zone node, and the feature value is set to zero for unlabeled zones; this feature enables the graph neural network to perceive the potential impact of mechanical abnormalities during the pressing process on local optical performance, especially providing additional risk correction information for the diffraction efficiency prediction in the contact force mutation region. The aforementioned features are organized into node units based on diffraction zones. Each node carries a numerical vector of the six types of features, forming a node feature matrix. Edge connections are established based on the physical adjacency relationships between diffraction zones (radial and circumferential adjacent zones), forming a graph structure. The design of the graph structure is based on the locality of the influence of process parameters on optical performance: the deviation in adhesive application, alignment residual, and residual layer deviation of a zone has the greatest impact on the diffraction efficiency of that zone itself, followed by the impact on adjacent zones, and the impact on distant zones decreases rapidly with increasing distance. This locality allows the edge connections between adjacent zones to capture the spatial transmission effect of process parameters without the need to establish a globally fully connected structure. At the same time, the full-aperture wavefront error is a global superposition of the phase errors of each zone. To enable the graph neural network to capture this global effect, an additional global convergence node is introduced into the graph structure. This node establishes edge connections with all zone nodes and aggregates the process features of each zone of the full aperture during graph convolution propagation. Its output feature vector is used to predict the root mean square wavefront error of the full aperture, which, together with the local diffraction efficiency predictions output by each zone node, constitutes the complete optical performance prediction result.

[0055] The process-optical performance mapping model is implemented using a graph neural network. The network consists of multiple layers of graph convolutional propagation modules: each layer aggregates the features of its neighboring partition nodes at each partition node, combining them with the current node's own features through weighted summation and nonlinear transformation, allowing each node's features to progressively incorporate process influence information from neighboring nodes. The edge connections between the global convergence node and all partition nodes are unidirectional aggregation directions. During the forward feature propagation of the graph convolution, the features of each partition node only flow along the edges from the partition node to the global convergence node. The features aggregated by the global convergence node are not written back to each partition node to ensure that the locality assumption of the partition nodes is not destroyed by global information. The features of the partition nodes are determined only by the process parameters of adjacent partitions, reflecting local process influences. The global convergence node unidirectionally aggregates the features of all partition nodes in each layer of graph convolution, and its final output feature vector predicts the root mean square value W_rms of the full-aperture phase wavefront error through the global output head. After multi-layer graph convolution propagation, the receptive field of each partition node expands with the increase of the number of layers, eventually covering the influence of process parameters across the entire aperture. The diffraction efficiency η_k of each partition is predicted through the output head of each partition node. The predicted diffraction efficiency values ​​of all partitions are organized into a full-aperture diffraction efficiency distribution η(x,y), which, together with W_rms, constitutes the optical performance prediction result. The reason for choosing a graph neural network instead of a fully connected network is that graph neural networks utilize the inductive bias of the graph structure, which can achieve cross-prescription generalization through structure sharing in the early stages of production when training data is limited (the sample size of different prescriptions is small). In contrast, fully connected networks are prone to overfitting to the process parameter distribution of a specific prescription when the sample size is insufficient, and cannot maintain prediction accuracy after prescription switching.

[0056] The training of the graph neural network employs a progressive data accumulation strategy: In the early stages of production, the network is pre-trained using finite element optical simulation data (theoretical diffraction efficiency and wavefront error calculated under different process parameter perturbations) to establish initial process-optical mapping capabilities. After entering steady-state production, sampling measurements are performed on incoming products at fixed batches. These measurements simultaneously complete two types of measurements for the same batch: first, measuring the actual optical performance (diffraction efficiency and wavefront error) of each zone using a wavefront sensor and grating diffraction efficiency meter, serving as the supervisory label for the graph neural network; second, measuring the alignment error after demolding, which, after in-plane displacement decoupling processing, serves as the training label for the physical information fusion neural network transfer learning in step two. Both types of measurements are completed within the same sampling batch to avoid inconsistencies in the training data of the two networks due to batch sampling. Incremental learning updates the weights of the two networks separately, ensuring that the prediction accuracy of both models improves synchronously with the accumulation of measured data.

[0057] After completing the optical performance prediction, the system compares the predicted η(x,y) and W_rms with the prescription design tolerance: the minimum value of η(x,y) across the entire aperture is not lower than the lower limit of diffraction efficiency τ_η (determined by the optical design specifications of the prescription), and W_rms does not exceed the upper limit of phase wavefront error τ_W (converted from the image quality requirements of the wearer's vision correction). If both of the above conditions are met, the optical qualification mark is given as "prediction qualified"; otherwise, it is marked as "prediction unqualified" and manual review is triggered.

[0058] The manufacturing data package consists of the following fields: a hash digest of the formulation version number and optical design file (used to bind the data package to a unique formulation design version); a summary of the entire process parameters (partition dispensing volume vector, final alignment residual, curing path type, and demolding speed setting); full-aperture metrological results (statistical summary and partitioned sample values ​​of thickness deviation field and warpage field); a complete record of defect maps (including the type, object plane coordinates, and confidence level of each defect) and rework history; partitioned sample values ​​of the predicted diffraction efficiency distribution η(x,y) and the root mean square value of the full-aperture phase wavefront error W_rms; an optical conformity mark; and a hash digest calculated in a deterministic serialization order for integrity verification during cross-production site replication. The receiving site recalculates the hash of the received data package and compares it with the hash digest field. If they match, it verifies that the data package has not been tampered with, serving as a technical guarantee for cross-site mass production consistency.

[0059] In one embodiment of the present invention, the present invention focuses on the field of nanoimprint mass production of ultrathin diffractive prescription lenses. The following takes the mass production scenario of an ultrathin diffractive prescription lens with a diopter of -6.00D (high myopia prescription), a diameter of 50 mm, and a target total thickness of 0.8 mm as an example to illustrate the data flow and specific values ​​of key process parameters in the six steps.

[0060] The diffraction pattern of this prescription lens is divided into five concentric ring regions from the center to the edge. The structural density of each region increases sequentially, with the outer ring (region 5) having a structural density approximately 3 to 4 times that of the inner ring (region 1), representing a typical high-number prescription structural density field morphology. After analyzing the structural density field D(x,y) of the prescription in step one, the diffraction tooth period of the outer ring region 5 is compressed to approximately 120 nanometers, belonging to the capillary penetration driven filling range (period less than 500 nanometers). The pre-configured process parameters are concentrated on marking the risk of insufficient filling in this region and setting an extended holding time. The curing path seed parameters select a gradual strategy from the inside out because the structural density of the outer ring is more than twice that of the inner ring, triggering a fan-shaped alternating curing hard constraint (the outer ring is divided into 8 fan-shaped sub-regions along the circumference). Example data are shown in Tables 1 and 2: Table 1. Comparison of dispensing amount before and after optimization in each diffraction zone and examples of residual layer thickness.

[0061] Note: For regions 1 and 2, the diffraction tooth period is higher than 500 nm, and the residual layer thickness is calculated using the Hele-Shaw viscous flow model; for regions 3 to 5, the period is lower than 500 nm, and the thickness is calculated using the capillary permeation driven model. Region 5 has a risk of insufficient filling within the standard holding time, and the holding time has been extended in the pre-configured process parameters. Under uniform adhesive application conditions, the residual layer deviation in the outer ring can reach -40 nm to -60 nm (the difference between the inner and outer rings is approximately 100 nm). After optimization, the residual layer deviation across the entire aperture is significantly narrowed, and the absolute value of the deviation in each region is below 10 nm, resulting in a significant improvement in phase substrate consistency. During the adhesive application in step two, the actual adhesive amount in each region is verified online to form an actual adhesive amount distribution vector, which is synchronously transmitted to the alignment drift prediction module. In this example, the verified value of the actual adhesive amount in region 5 is 16.8 nL, which is within the 5% accuracy threshold of the target value of 17 nL, and no compensation dispensing is required.

[0062] Table 2. Examples of multidimensional gating determination results for 5 samples

[0063] Note: Sample B's warpage of 3.2μm did not exceed the default value of τ_w (5μm), but for the purpose of illustrating the rework scenario, it is assumed that the τ_w of the formulation is tightened to 3μm according to the assembly tolerance, thus exceeding the threshold. The warpage field is distributed radially in a bowl shape, attributed to curing shrinkage stress-type warpage. After rework, the seed parameter strategy weight of the curing path will be adjusted. The primary defect (outer ring particles) of Sample C is located within the optically effective area. It is still within the reworkable window within 15 minutes after curing. Solvent-assisted peeling will be performed, followed by re-imprinting. Sample D's thickness deviation exceeds τ_t and has exceeded the reworkable time window, so it is deemed scrap. At the same time, the correction amount of adhesive in Zone 4 of this batch is written into the process pre-configuration parameter set. The number of secondary defects (2) of Sample E is lower than the N_d threshold, so it is deemed acceptable for warehousing. However, the contact force change rate during the pressing process in Zone 4 exceeds 3 times the historical average standard deviation. An additional process risk labeling field is added. The neural network in step six uses the abnormal contact force degree of this zone as the input of the sixth type of node feature, and prioritizes the inclusion of this product in the next sampling batch for actual measurement and verification.

[0064] For samples A and E, which are deemed acceptable for warehousing, step six extracts six types of node features from the process log to construct a graph structure. After inference by a graph neural network, the diffraction efficiency distribution η(x,y) of each zone and the root mean square value of the full-aperture phase wavefront error W_rms are output. For sample A, the actual amount of adhesive applied to each zone deviates from the target value by less than 5%, the synthetic modulus of the alignment residual is lower than τ_align, and the maximum absolute value of the residual layer deviation field is 8 nm. The graph neural network predicts that the lowest value of the full-aperture diffraction efficiency is higher than τ_η and W_rms is lower than τ_W, so the optical qualification is marked as "predicted qualified". For sample E, due to the risk of abnormal contact force in zone 4, the graph neural network adds a risk correction to the predicted diffraction efficiency value of this zone. The prediction result still meets the τ_η requirement, so the optical qualification is marked as "predicted qualified". However, the process risk label field is retained in the manufacturing data package, and the network weights will be updated after sampling and actual measurement verification.

[0065] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for mass production of ultrathin diffractive lenses using nanoimprinting, characterized in that, Includes the following steps: S1. Obtain the prescription optical design file, perform structural density field analysis and inverse optimization iteration of partitioned dispensing amount to form a spatially uniform partitioned dispensing pattern and process pre-configuration parameter set; S2, based on the partitioned dispensing pattern and the pre-configured process parameter set, completes precise partitioned dispensing of adhesive. The physical information fusion neural network is used to perform two-stage prediction and closed-loop pre-compensation for the alignment drift throughout the pressing process, and records the pressing alignment status data. S3, based on the pre-configured process parameter set and pressing alignment status data, adopts a prescription adaptive curing path optimization strategy to complete UV curing and demolding, and outputs the imprinted part and full-diameter measurement data through online metering; S4, based on the embossed parts and full-diameter measurement data, uses a warp-compensated and corrected diffraction pattern difference field-driven multi-scale convolution classification network to complete online defect identification, and generates defect maps and reworkability assessment conclusions; S5 integrates full-caliber measurement data, defect maps, and reworkability assessment conclusions. It comprehensively judges various quality indicators through multi-dimensional gating rules and outputs warehousing permission and process risk labeling. S6 combines the data from adhesive application verification and pressing alignment, full-aperture measurement data, defect maps, and process risk annotations. It then uses a graph neural network to infer the full-aperture diffraction efficiency distribution and the root mean square value of the phase wavefront error, generating a manufacturing data package containing optical conformity verification fields.

2. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In S1, the structural density field analysis includes: The radial start and end coordinates of each diffraction zone and the period, height and cross-sectional shape parameters of the diffraction teeth are extracted from the prescription optical design file. The total volume of the diffraction groove on the unit substrate projection area of ​​each zone is defined as the structural density value. The structural density values ​​of all zones on the full aperture are organized into a two-dimensional structural density field. For partitions where the diffraction tooth period is greater than the preset period threshold, the residual layer thickness is calculated based on the thin-layer flow theory and the volume conservation of each partition. For partitions where the diffraction tooth period is less than the preset period threshold, the capillary permeation drive mechanism is used as the main mechanism. The capillary permeation drive pressure is calculated based on the diffraction tooth period and contact angle of each partition. The actual filling depth during the pressure holding time is calculated according to the permeation equation based on the groove depth and effective viscosity. The residual layer thickness deviation value is obtained by subtracting the actual consumption volume from the amount of adhesive received by each partition. The calculation results of the two types of partitions are combined and the residual layer thickness deviation field is output uniformly across the entire aperture.

3. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In step S1, the reverse optimization iterative solution of the partitioned dispensing amount includes: The amount of adhesive applied to each diffraction zone is set as an optimization variable. The objective function is to minimize the spatial standard deviation of the residual layer thickness at the full aperture. The constraints are set as follows: the amount of adhesive applied to each zone is non-negative, the total amount is equal to the total amount of adhesive applied in the process setting, and the difference between the amounts of adhesive applied to adjacent zones does not exceed the upper limit of the single accuracy of the dispensing system. The optimal partitioning dispensing pattern is determined by iteratively solving the gradient descent method until convergence.

4. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In step S1, the process pre-configuration parameter set includes three types of parameters: The first category is the initial value of the alignment compensation coefficient, which estimates the expected alignment drift based on the radial gradient of the substrate thickness and the structural density field; the second category is the pressing pressure distribution parameter, which estimates the pressure requirements of each region based on the spatial distribution of the partitioned adhesive pattern. The third category is the curing path seed parameter, which selects the initial type of the curing path spatial strategy based on the radial gradient characteristics of the structural density field; the process pre-configuration parameter set is stored with the prescription version number and design file hash as dual indexes.

5. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In step S2, the precise application of adhesive in designated zones includes: The dispensing control module reads the dispensing pattern of each zone and distributes resin droplets according to the optimal dispensing amount for each zone. The actual dispensing amount for each zone is verified online. If the difference from the target value exceeds the dispensing accuracy threshold, compensation dispensing is triggered. After the adhesive is applied, an actual adhesive amount distribution vector is generated and synchronously transmitted to the alignment drift prediction module. During the alignment execution phase, the visual alignment module simultaneously images the multi-scale alignment marks on the working model and the substrate, calculates the superposition error components in translation and rotation using the phase correlation method, and drives the precision platform to converge the superposition error to below the alignment accuracy threshold, thus completing the initial alignment.

6. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In S2, the backbone structure of the physical information fusion neural network consists of a convolutional neural network branch and a long short-term memory network branch. The convolutional neural network branch uses the partitioned dispensing pattern as a two-dimensional feature map to extract the spatial distribution of resin amount, and the long short-term memory network branch processes temporal features. The two-stage prediction includes: the first stage, before the pressing is triggered, receiving the initial superposition error triaxial component, substrate thickness and elastic modulus parameters, partition dispensing amount distribution vector, ambient temperature and the initial value of the alignment compensation coefficient in the process pre-configuration parameter set, and outputting the initial pre-compensation amount; In the second stage, after the pressing is triggered, the long short-term memory network branch receives the timing signal from the contact force sensor in real time and updates the alignment drift prediction in a rolling manner.

7. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In S2, the closed-loop pre-compensation includes: During the pressing process, in-plane displacement decoupling is performed simultaneously. The in-plane displacement component caused by bending is calculated according to the thin plate bending equation. The pure rigid body drift is obtained by subtracting the in-plane displacement from the measured displacement. This drift is compared with the time-series prediction value of the physical information fusion neural network. If the deviation exceeds the online update threshold, the pre-compensation amount is dynamically corrected. The training of the physical information fusion neural network incorporates the force balance relationship between the elastic stretching in the base plane and the capillary pressure difference in each zone as a soft constraint term into the loss function. The soft constraint term requires that the gradient distribution of the drift predicted by the physical information fusion neural network in space maintains the same directionality as the capillary pressure gradient distribution calculated from the structural density field.

8. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In step S2, the training of the physical information fusion neural network employs a two-stage strategy: The first stage uses finite element simulation data for pre-training. The finite element simulation data consists of alignment drift time series data generated by substrate-resin-mold coupled mechanical simulation, covering the structural density field parameter space of different formulations. The second stage uses the alignment error collected after demolding in the actual production process for transfer learning. The training label is decoupled by in-plane displacement of the bending surface to ensure that the label reflects the alignment drift of pure rigid body.

9. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In step S3, the formulation adaptive curing path optimization strategy includes: The curing path seed parameters are read from the pre-configured process parameter set, and the shrinkage stress field of each diffraction zone is calculated using the structural density field and the resin curing shrinkage coefficient. The final alignment residual triaxial component in the pressing alignment state data is read, and stress deflection correction is applied to the shrinkage stress field to obtain the corrected shrinkage stress field. The curing time of each diffraction zone is used as an optimization variable, and the objective function is to minimize the maximum value of the full-aperture warping. A greedy sequence scheduling algorithm is used to solve the spatial allocation scheme of the curing time. When the structural density of the outer ring zone exceeds a preset multiple of the structural density of the inner ring zone, the alternating fan-shaped curing strategy is embedded as a hard constraint into the greedy sequence scheduling algorithm. The outer ring zone is divided into several fan-shaped sub-regions along the circumference. The odd-numbered fan-shaped sub-regions are cured first, followed by the even-numbered fan-shaped sub-regions.

10. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In step S3, the UV curing includes: The UV curing light source is equipped with a multi-zone addressable scanning unit, which applies UV irradiation to each zone sequentially according to the optimized curing time sequence. Each zone adopts a two-segment light intensity curve. The first segment uses a preset low light intensity for pre-gel irradiation, and the second segment uses a fully curing light intensity to complete cross-linking curing. The demolding includes: a precision platform lifting the working mold at a set uniform speed to separate it from the cured structure. The online measurement includes: using a white light interferometer to measure the thickness distribution of the entire aperture to form a thickness deviation field, and using a displacement sensor array to measure the warping distribution of the entire aperture to form a warping field, which serves as the measurement data for the entire aperture.

11. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In step S4, the warpage compensation correction includes: The deflection angle of the base normal direction at each measurement point is calculated from the warp field in the full-caliber measurement data. The deflection angle and the diffraction tooth period at that position are substituted into the grating equation to calculate the diffraction peak direction under the actual incident angle. The theoretical diffraction pattern after warp compensation is obtained through coordinate transformation. The difference field of the diffraction pattern is obtained by subtracting the theoretical diffraction pattern from the measured diffraction pattern pixel by pixel.

12. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In S4, the input of the multi-scale convolutional classification network is obtained by frequency domain decomposition of the diffraction pattern difference field through prescription adaptive spatial frequency threshold. The first spatial frequency threshold and the second spatial frequency threshold are determined according to the minimum and maximum values ​​of the diffraction tooth period on the full aperture, respectively, and the difference field is decomposed into three component images of high frequency, medium frequency and low frequency as parallel input channels. The multi-scale convolutional classification network adopts a feature pyramid network architecture, outputs the probability distribution of defect type labels, and uses the maximum probability value of the probability distribution of each spatial location as the confidence level. Locations with confidence levels higher than the preset classification threshold are marked as defect points to generate the defect map. The reworkability assessment conclusion is obtained by calculating the remaining reworkable time based on the defect type, the interval between the current time and the UV curing completion time, and the historical imprinting cycle number of the working mold.

13. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In S5, the multidimensional gating rules include: Thickness consistency gating: if the maximum absolute value of the thickness deviation field does not exceed the first thickness threshold, it passes; warpage gating: if the maximum absolute value of the warpage field does not exceed the first warpage threshold, it passes; defect gating is divided into three levels: the first level checks whether there are granular or scratch-type defects in the optical effective area; the second level counts whether the number of bubble-type and underfilled defects exceeds the first defect number threshold; the third level adopts a layered judgment: the first layer judges whether the composite vector magnitude of the final alignment residual in the pressing alignment state data exceeds the alignment deviation threshold; the second layer judges whether the average deviation of alignment deviation type defects in the defect map exceeds the preset proportion of the alignment deviation threshold. If all departmental controls pass, the entry permit is granted. If there are abnormal events in the contact force timing record, the process risk label is added.

14. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In step S5, if any gate fails and the reworkability assessment conclusion is reworkable, then rework is performed and the process pre-configuration parameter set is updated accordingly based on the gate failure type: if the thickness consistency gate fails, then the dispensing amount of the corresponding partition in the partition dispensing pattern is corrected. If the warp gate fails, the spatial distribution characteristics of the warp field are analyzed to diagnose the warp type. If the radially symmetrical distribution is determined to be curing shrinkage stress type warp, the curing path seed parameter is adjusted. If the edge asymmetrical distribution is determined to be demolding stress type warp, the demolding speed setting value is reduced. If the spatial distribution of the warp field is related to the thickness deviation field, it is determined to be residual layer non-uniform warp, and the corresponding partition glue amount is corrected. If the Level 3 defect gate fails, an attribution analysis is performed by comparing the global alignment residual direction in the pressing alignment state data with the spatial distribution direction of the local alignment deviation defect in the defect map. If the directions are consistent, the pre-compensation parameters of the physical information fusion neural network are updated; if the directions are inconsistent, the curing path seed parameters or demolding speed setting value are adjusted.

15. The method for mass production of ultrathin diffractive lenses by nanoimprinting according to claim 1, characterized in that, In S6, each diffraction zone is used as a node. Each node carries six types of features: actual adhesive application amount, final alignment residual, residual layer thickness deviation, warpage amount, defect mark, and contact force abnormality degree. Edge connections are established according to physical adjacency relationships, and a global convergence node is introduced to form a graph structure. After multi-layer graph convolution propagation of the graph neural network, the diffraction efficiency prediction value output by each zone node is organized into a full-aperture diffraction efficiency distribution. The global convergence node outputs the root mean square value of the full-aperture phase wavefront error, which is compared with the prescription design tolerance to generate an optical conformity verification field and written into the manufacturing data package. The training of the graph neural network adopts a progressive data accumulation strategy. In the early stage of production, it is pre-trained with finite element optical simulation data. After steady-state production, it is periodically sampled and measured. The optical performance measurement is completed simultaneously in the same batch and serves as the supervision label of the graph neural network. The alignment error measurement is decoupled by in-plane displacement and serves as the training label of the physical information fusion neural network for transfer learning. The weights of the two networks are updated synchronously through incremental learning.

16. A nanoimprint mass production system for ultrathin diffractive lenses, used to perform the steps in the nanoimprint mass production method for ultrathin diffractive lenses as described in any one of claims 1-15, characterized in that, include: The analysis and optimization module is used to obtain the prescription optical design file, perform structural density field analysis and inverse optimization of the partitioned dispensing amount iterative solution, and form a spatially uniform partitioned dispensing pattern and a set of pre-configured process parameters. The adhesive application and alignment pressing module is used to complete precise adhesive application in zones according to the zoned adhesive pattern and the pre-configured process parameter set. It uses a physical information fusion neural network to perform two-stage prediction and closed-loop pre-compensation for alignment drift throughout the pressing process and records the pressing alignment status data. The curing and metering module is used to complete UV curing and demolding based on the pre-configured process parameter set and pressing alignment status data, and to output the imprinted parts and full-diameter metering data through online metering. The defect identification module is used to identify defects online based on the embossed parts and full-diameter measurement data, using a multi-scale convolution classification network driven by the diffraction pattern difference field with warp compensation correction, and generating defect maps and reworkability assessment conclusions. The quality assessment module integrates full-caliber measurement data, defect maps, and reworkability assessment conclusions. It then uses multi-dimensional gating rules to comprehensively assess various quality indicators and outputs warehousing permission and process risk labeling. The optical verification and data package generation module combines adhesive application verification and pressing alignment status data, full-aperture measurement data, defect maps and process risk labels, and uses graph neural networks to infer the full-aperture diffraction efficiency distribution and the root mean square value of phase wavefront error to generate a manufacturing data package containing optical conformity verification fields.

Images