Method for improving glass through hole yield

Abstract

The invention discloses a method for improving the yield of glass through holes, and the method comprises the steps: obtaining the internal morphology of a laser-induced modification region through three-dimensional point cloud data, and precisely calculating the modification depth; fusing the laser process parameters and the material attributes by adopting an improved convolutional neural network model, and predicting the etching rate and the standard deviation of each modified area; based on the same set of three-dimensional detection data, a rework compensation decision and an etching process are driven to optimize double closed loops, point-level laser secondary induction is carried out on an insufficient-depth area, array ultrasonic power distribution is dynamically adjusted according to a rate prediction result, and ultrasonic power is improved for a slow-speed area; and continuously fine-tuning model parameters through an online meta-learning adaptation mechanism to realize adaptive evolution of the process model. According to the method, detection, prediction, reworking and regulation are integrated in the same process track, the problems that internal induction defects cannot be recognized, the etching rate is uneven and the process depends on experience in a traditional process are solved, and the preparation yield and the process consistency of the TGV through holes are remarkably improved.

Description

Technical Field

[0001] This invention relates to a method for improving the yield of glass through-holes. Background Technology

[0002] Through-glass vias (TGVs) are vertical electrical interconnect structures that penetrate glass substrates and have broad application prospects in fields such as radio frequency devices, microelectromechanical systems (MEMS) packaging, and optoelectronic system integration. Currently, the mainstream TGV forming process employs a two-step method: "laser-induced + wet etching." First, a high-energy pulsed laser beam (picosecond or femtosecond laser) is focused onto the glass surface or interior to locally modify the glass material, forming a modified region at a predetermined location. Then, the laser-treated glass substrate is placed in an etching solution. Because the dissolution rate of the modified glass in the HF or KOH-based etching solution is significantly higher than that in the unmodified area, selective etching is used to form the through-hole.

[0003] The main problems with existing technologies are:

[0004] (1) Limited detection dimensions make it impossible to predict the etching effect;

[0005] In the existing process, the detection after laser induction is only for two-dimensional morphological observation of the glass surface (such as observing the surface aperture and edge smoothness), and cannot obtain the depth uniformity inside the modified area, whether there are microcracks or local insufficient energy and other hidden defects. The laser induction effect directly determines the subsequent etching quality. If the modified area is not deep enough, or there are internal fractures or uneven density, even if the surface detection is normal, problems such as irregular aperture of the through hole and residual glass layer at the bottom (not etched through) will still occur after etching. As shown in Figure 3, this figure is the actual measurement result of the through hole after laser induction. The apertures of the four points [1] to [4] in the figure are 41.13μm, 22.72μm, 47.28μm and 35.29μm, respectively. The maximum aperture deviation is 25.06μm, and some points have obvious un-etched areas at the bottom, which directly reflects the destructive effect of laser-induced internal defects on the formation of through holes. Such defects cannot be intercepted by the existing surface detection and can only be discovered after the etching is completed, resulting in a large amount of waste and material waste.

[0006] (2) The etching process lacks real-time control, resulting in poor rate consistency;

[0007] Wet etching is a typical diffusion-controlled reaction. Even slight fluctuations in the depth and morphology of the modified region, as well as the concentration and temperature of the etching solution, can lead to differences in the etching rate of each via. Current technologies use a fixed etching time (usually 15-25 minutes), which cannot dynamically adjust process parameters according to actual rate differences. This results in some vias being over-etched (rough walls, severe side etching) while others are under-etched, making it difficult to guarantee the uniformity of etching across the entire surface. Although traditional ultrasonic-assisted stirring can improve mass transfer efficiency, it uses a uniform power distribution and cannot specifically compensate for slow-moving areas.

[0008] (3) Process decisions rely on experience and lack data-driven capabilities;

[0009] The mapping relationship between laser-induced parameters (energy density, repetition rate, defocusing amount) and the internal morphology and etching rate of the modified region is highly nonlinear and influenced by material properties. Existing technologies rely on engineers' experience to set fixed parameter combinations, which cannot adaptively optimize based on real-time detection data. When glass material batches change or the laser ages, repeated trial and error adjustments are required, which is inefficient and difficult to guarantee consistency. In addition, laser, etching, and detection data are isolated from each other, failing to form a closed-loop feedback, resulting in long process optimization cycles. Summary of the Invention

[0010] The present invention provides a method for improving the yield of glass through-holes in order to solve the problems existing in the prior art.

[0011] The technical solutions adopted in this invention are as follows:

[0012] A method for improving the yield of glass through-holes includes the following steps:

[0013] S1: A modified region is formed at a preset position by inducing the formation of a modified region on a glass substrate using a pulsed laser beam;

[0014] S2: Obtain the three-dimensional point cloud data of the modified region and calculate the actual modification depth d of each modified region. i ;

[0015] S3: Input the three-dimensional point cloud data into the improved CNN model to predict the etching rate v of each modified region. i Generate an etching rate distribution map;

[0016] S4: Determine whether each modified region satisfies d. i ≥d0, where d0 is a preset threshold:

[0017] If d exists i For modified regions less than d0, output the coordinates and depth compensation Δd of that modified region. i = d0- d i Return the glass substrate to S1, according to Δd iAdjust the laser parameters to induce a second wave of radiation at the non-conforming points, repeating S2-S4 until all d points are reached. i ≥d0;

[0018] If all conditions are met, then execute S5;

[0019] S5: Based on the etching rate distribution map generated in S3, dynamically adjust the power distribution of the ultrasonic field during the wet etching process, increase the ultrasonic power for the modified areas whose predicted etching rate is lower than the arithmetic mean of the etching rates of all modified areas on the glass substrate, and perform etching until the end point.

[0020] S6: After etching is completed, check the status of the through hole and output a pass or fail signal.

[0021] Furthermore, the improved CNN model is an improvement on the original CNN model, and the improvements are as follows:

[0022] (1) The original CNN model's single serial feature extraction architecture is reconstructed into a parallel dual-branch architecture;

[0023] (2) Add a multiphysics coding module to the input layer of the original CNN model, and encode the laser pulse energy density, repetition frequency, defocus amount and glass substrate material properties in S1 into implicit physical field vectors, and fuse them with the three-dimensional point cloud data through tensor product operation;

[0024] (3) Add a differentiable etching dynamics constraint layer to the loss function of the original CNN model;

[0025] (4) Replace the output layer of the original CNN model with a Bayesian nested variational inference architecture. By adding a random deactivation layer after each convolutional layer, the mean and standard deviation of the etching rate are output simultaneously through 10 random forward propagations during inference. i ;

[0026] (5) In the inference stage of the original convolutional neural network model, an online meta-learning adaptation mechanism is added. The actual etching rate and prediction deviation fed back by S6 are used to quickly fine-tune the weights of the last two layers of the network through 5-step gradient updates.

[0027] Furthermore, the parallel dual-branch architecture includes a micro-morphology branch and a macro-geometry branch. The micro-morphology branch adds a lightweight U-Net sub-network to capture roughness features of the modified region surface at the 0.5-5μm scale, while the macro-geometry branch adds a global deep gradient coding layer to extract the anisotropic features of the overall morphology of the modified region. The two branches are cascaded and fused in a high-dimensional feature space.

[0028] Furthermore, the constraint layer constructs physical residual terms based on the Arrhenius equation:

[0029] L physics = λ||v i - f(C HF , T, d i )||²,

[0030] Among them, L physics The physical residual term is used to quantify the deviation between the CNN prediction and the theoretical physical value; λ is the weight coefficient; v i C represents the etching rate of the i-th modified region. HF is the molar concentration of HF in the S5 etching solution; T is the absolute temperature of the S5 etching process.

[0031] f(·) is the theoretical function for calculating the etching rate, constructed based on the Arrhenius kinetic equation, and has the following specific form:

[0032]

[0033] Where k is the pre-exponential factor. For activation energy, The gas constant is for Reaction order, This is the modification depth index.

[0034] Furthermore, the online meta-learning adaptation mechanism adopts a two-layer optimization strategy based on model-independent meta-learning (MAML). The inner layer performs five steps of gradient descent on the weights of the last two layers of the network with a learning rate of 0.001 to quickly adapt to the current batch of data. The outer layer updates the network's shared initial parameters with a meta-learning rate of 0.01, so that the model can maintain its generalization ability across different glass substrate batches.

[0035] Furthermore, in S5, the dynamic adjustment of the ultrasonic field power distribution is achieved through an 8×8 array of ultrasonic transducers. The power adjustment range of a single transducer is 50-500W, and the adjustment response time is ≤1 second. The ultrasonic power calculation formula is as follows:

[0036] ,

[0037] Where P base The reference power is η, and the modulation coefficient is η. σ is the arithmetic mean of the etching rates of all modified regions on the glass substrate. i Let be the standard deviation of the i-th modified region.

[0038] Furthermore, the wet etching in S5 uses an HF or KOH solution system, a temperature of RT-50℃, an ultrasonic frequency of 40kHz and 80kHz alternating, a power of 500-1000W, and a process time of 15-25min.

[0039] Furthermore, after the S5 etching is completed, a multi-step cleaning and drying process is performed, which includes ultrapure water cleaning, ozone water cleaning, megasonic cleaning, final megasonic deep cleaning, and vacuum Marangoni drying.

[0040] Ultrapure water cleaning removes soluble salts from the surface, temperature RT, spray time 30-60s, overflow bubbling time 30-180s;

[0041] Ozone water cleaning decomposes carbon compounds at a concentration of 10-20 ppm, with spraying time of 30-60 seconds and overflow bubbling time of 30-180 seconds.

[0042] Megasonic cleaning removes SiO2 debris at a frequency of 1MHz, a power of 0.5~0.8W / cm², and a time of 30-180s.

[0043] Finally, megasonic cleaning was used to remove particle residues. The solution was UPW + 0.1~0.2% TMAH, the temperature was 40±2℃, and the time was 120-240s. Vacuum Marangoni drying was used to prevent water residues. IPA steam replacement, slow lifting speed of 1mm / s, vacuum of -80kPa, and hot N2 purging at 50-60℃ were employed.

[0044] Furthermore, the detection of the through-hole status in S6 includes hole diameter and morphology detection, hole depth and aspect ratio detection, hole wall roughness detection, and detection of residual etching solution and microcrack defects.

[0045] Furthermore, the three-dimensional point cloud data in S2 is acquired by confocal microscopy or structured light three-dimensional scanning; the preset threshold is 60%-80% of the thickness of the glass substrate.

[0046] The present invention has the following beneficial effects:

[0047] (2) Obtain the internal morphology of the modified area through three-dimensional point cloud data and accurately calculate the modification depth d. i This technology upgrades the traditional two-dimensional inspection of glass surfaces to three-dimensional inspection of internal structures, effectively identifying hidden defects caused by insufficient laser energy and focus shift before the etching process, thus avoiding ineffective etching and subsequent material waste.

[0048] (2) An improved convolutional neural network model is used to integrate laser process parameters and material properties to predict the local etching rate and standard deviation of each modified area, transforming the traditional blind etching process that relies on fixed time into a rate prediction process based on physical laws and data-driven, thereby improving the accuracy of etching endpoint judgment.

[0049] (3) The same set of three-dimensional detection data simultaneously drives the laser rework compensation decision of S4 and the ultrasonic power spatial modulation decision of S5, realizing the synergistic linkage between defect interception and process optimization, and changing the open-loop mode in which detection, rework and etching are isolated from each other in the traditional process.

[0050] (4) By using the online meta-learning adaptation mechanism, the model parameters are continuously fine-tuned using the actual data after etching, so that the process model can adapt to dynamic changes such as laser aging and material batch fluctuations, reducing the debugging cycle of manual intervention and repeated trial and error. Attached Figure Description

[0051] Figure 1 This is a flowchart of the present invention.

[0052] Figure 2 To improve the architecture of CNN models.

[0053] Figure 3 The measured results of the through hole after laser-induced anomaly. Detailed Implementation

[0054] The invention will now be further described with reference to the accompanying drawings.

[0055] This invention relates to TGV (Through Glass Via) fabrication technology in the field of semiconductor packaging, and is particularly suitable for 2.5D / 3D packaging scenarios with high requirements for via yield and etching uniformity, such as the mass production process of TGV substrates for RF MEMS devices and optoelectronic integrated modules. This embodiment uses a 500μm thick borosilicate glass substrate as the processing object, aiming to fabricate TGV vias with a diameter of 50μm. The specific implementation steps are as follows:

[0056] Step S1: Laser-induced formation of the modified region.

[0057] A femtosecond laser induction device (model: Trumpf TruMicro 5000, wavelength 1030nm, pulse width 350fs) is used, along with a six-axis motion platform (positioning accuracy ±5μm); the automatic feeding and positioning of the glass substrate is achieved by a robotic arm (model: ABB IRB 1200, load 3kg).

[0058] Based on the material properties of borosilicate glass, the following parameters are set to balance the modification efficiency and the heat-affected zone (controlled within 5μm):

[0059] Pulse energy density: 8J / cm² (to avoid glass cracking due to excessive energy, and failure to form an effective modification zone due to insufficient energy).

[0060] Repetition frequency: 200 kHz;

[0061] Defocusing amount: -200μm (the focal point is located 200μm inside the glass substrate, forming a "from the inside out" modification zone to reduce surface damage).

[0062] Scanning speed: 500 mm / s (laser interaction time of 100 μs for a single modified zone to ensure continuous and uniform modification).

[0063] The process of forming the modified region is a conventional existing processing technology, which will not be described in detail in this invention. The femtosecond laser breaks the SiO2 bonds in the glass through photochemical effects (multiphoton absorption), forming a modified region with a diameter of about 45 μm and an initial depth of about 300 μm (not completely penetrating the substrate, to be etched later).

[0064] Step S2: 3D detection and depth calculation of the modified area.

[0065] The robotic arm transports the laser-induced glass substrate to the inspection area (using a confocal laser scanning microscope) to perform a three-dimensional scan on each modified area.

[0066] Scanning range: 100μm × 100μm area around each modified region, axial scan step size 0.5μm;

[0067] Data processing: Generate 3D point cloud data for each modified area (1000×1000×600 data points, including horizontal, vertical and axial coordinates), and remove noise using Gaussian filtering (standard deviation σ=1.2).

[0068] Using the Zeiss ZEN 3.0 3D analysis software built into the microscope, with the upper surface of the glass substrate as the reference plane, the "maximum axial depth" (the distance from the bottom of the modified area to the reference plane) of each modified region was calculated and denoted as d. i In this embodiment, the initial di values ​​of the 100 modified regions are distributed between 280-320 μm (due to the depth difference caused by slight fluctuations in laser energy).

[0069] Step S3: Improve the etch rate prediction of the CNN model.

[0070] Training dataset: 1000 sets of historical data were collected. Each set includes "laser parameters (energy density, repetition frequency, defocusing amount) + glass material properties (refractive index, coefficient of thermal expansion) + 3D point cloud features of the modified area (depth, surface roughness, volume) + actual etching rate". The actual etching rate is calculated by "difference in via depth before and after etching / etching time".

[0071] Model Framework: An improved CNN model is constructed, with the input layer consisting of a 512×512×3 3D point cloud feature map and a 4D physical field vector (laser energy density 8J / cm², repetition frequency 200kHz, defocusing amount -200μm, glass refractive index 1.52), and the output layer consisting of the average etching rate v of each modified region. i(Unit: μm / min) + Standard Deviation σ i ".

[0072] The parallel dual-branch architecture is as follows:

[0073] Micromorphology branch: Lightweight U-Net sub-network (4 layers of encoder, 4 layers of decoder, ReLU activation function) to extract roughness features (such as Ra value) on the surface of the modified region at the scale of 0.5-5μm.

[0074] Macroscopic geometry branch: Global deep gradient encoding layer (calculates the L1 norm of the depth gradient of the modified region) to extract anisotropic features (such as the modified region being “conical” or “cylindrical”).

[0075] Fusion method: The feature maps output by the two branches (both 64×64×128) are concatenated through channels (the dimension becomes 64×64×256) and then input into the subsequent convolutional layer.

[0076] To prevent model predictions from deviating from the etching dynamics, a differentiable physical residual term based on the Arrhenius equation is introduced into the loss function, as shown in the following formula:

[0077] L physics = λ||v i - f(C HF , T, d i )||²,

[0078] Among them, L physics The physical residual term is used to quantify the deviation between the CNN prediction and the theoretical physical value; λ is the weight coefficient; v i C represents the etching rate of the i-th modified region. HF is the molar concentration of HF in the S5 etching solution; T is the absolute temperature of the S5 etching process.

[0079] f(·) is the theoretical function for calculating the etching rate, constructed based on the Arrhenius kinetic equation, and has the following specific form:

[0080] ,

[0081] Where k is the pre-exponential factor. For activation energy, The gas constant is for Reaction order, This is the modification depth index.

[0082] The values ​​of key parameters in this embodiment are:

[0083] Weighting coefficient λ: 0.3 (determined through cross-validation of 10 validation sets to balance data fitting with physical laws and ensure that the deviation between the predicted value and the physical theory is ≤5%).

[0084] Pre-exponential factor k: 2.5 × 10 -3 um / (min·mol -m ·K) (Fitting of static etching experiment of borosilicate glass with 40% HF solution, goodness of fit R) 2 =0.98).

[0085] Activation energy Ea: 45 kJ / mol.

[0086] Gas constant R: 8.314 J / (mol・K) (standard physical constant).

[0087] HF reaction order m: 1.2 (by measuring the etching rate by changing the HF concentration to 8-15 mol / L and fitting the power relationship between concentration and rate, m=1.2).

[0088] Modification depth index n: 0.8 (by changing the modification depth to 300-400μm, measuring the etching rate, and fitting the power relationship between depth and rate, n=0.8).

[0089] The total loss function of the model is L = L data + L physics L data To predict v i The MSE loss relative to the actual rate ensures that "data-driven" and "physical laws" work in tandem.

[0090] Bayesian Inference and Meta-Learning:

[0091] Random deactivation layer: Add a Dropout layer after each convolutional layer, with a dropout probability of 0.2;

[0092] Inference process: Perform 10 random forward propagations for each modified region, and output v. i The mean (distributed in this embodiment between 1.8-2.2 μm / min) and σ i (Distributed in the range of 0.05-0.12 μm / min).

[0093] Meta-learning fine-tuning: Using the MAML algorithm with Adam as the inner optimizer (learning rate 0.001), 5 gradient updates were performed, reducing the model prediction error from the initial 8% to 4.5%.

[0094] Step S4: Determination of the quality of the modified area and secondary induction.

[0095] Preset threshold d0 setting: Based on the etching requirements of a 500μm thick glass substrate, d0 = 70% of the glass thickness = 350μm.

[0096] Secondary induction operation:

[0097] Non-conforming point identification: Among 100 modified zones, 12 modified zones have di < 350 μm (range 280-345 μm), and the depth compensation Δd is calculated. i =d0-d i (Range 5-70μm);

[0098] Laser parameter adjustment: based on Δd i Linear relationship with laser energy density (experimental fit: Δd) i =0.05×ΔE, where ΔE is the energy density increment), increasing the laser energy density at the non-compliant points from 8 J / cm² to 8.1-9.4 J / cm², and repeating the laser induction of S1 only at the non-compliant points (to avoid damaging the compliant areas).

[0099] Re-examination: After secondary induction, the d values ​​of the 12 modified regions were re-examined using a confocal microscope. i All were improved to 355-370μm, and all met the d... i If ≥d0, proceed to the etching process.

[0100] Step S5: Dynamic ultrasonic-assisted wet etching, cleaning and drying.

[0101] The etching parameters are set as follows:

[0102] Etching solution: 40% HF solution (molar concentration 12 mol / L), temperature controlled in a closed loop by heating rod and thermocouple to 40℃±0.5℃;

[0103] Ultrasonic transducer: 8×8 array transducer (single power 50-500W, response time 0.8 seconds, dual-frequency 40kHz / 80kHz alternating operation, switching cycle 10s);

[0104] Dynamic power parameters: base power Pbase=200W, modulation coefficient η=50 (determined experimentally to ensure appropriate power compensation in low-rate areas and prevent over-etching).

[0105] The dynamic power adjustment process is as follows:

[0106] Calculate the arithmetic mean of etching rates (In this embodiment) =2.0um / min, N=100).

[0107] For each modified region, according to the formula Calculate the target ultrasonic power:

[0108] Calculation example: If a certain modified region v i = 1.8 um / min, σ i = 0.1 um / min, then P i = 200 + 30 × ((2.0 - 1.8) / 0.1 = 260W; if another modified region v i = 2.2 um / min, σ i = 0.08 um / min, then P i = 125W.

[0109] The etching machine control system is based on P i Adjust the transducer power at the corresponding position in real time, and etch for 20 minutes (to ensure penetration of a 500μm thick substrate).

[0110] The multi-step cleaning and drying process is as follows:

[0111] Ultrapure water cleaning: temperature RT (25℃), spraying for 45s (pressure 0.2MPa), overflow bubbling for 120s (bubbling volume 5L / min) to remove soluble salts from the surface;

[0112] Ozone water cleaning: ozone concentration 15ppm, spraying for 40s, overflow bubbling for 90s, decomposing laser-induced residual amorphous carbon;

[0113] Megasonic cleaning: frequency 1MHz, power 0.6W / cm², time 90s, to remove SiO2 debris;

[0114] TMAH megasononic cleaning: The solution is UPW + 0.15% TMAH (tetramethylammonium hydroxide), temperature 40℃, time 180s, neutralizing HF residue (pH adjusted to 10.5), dissolving H2SiF6 crystals;

[0115] Chelating agent cleaning: 1% EDTA-2Na (disodium ethylenediaminetetraacetate) + 5% citric acid, temperature 50℃, time 300s, complexing Ca²⁺ / Fe³⁺ ions;

[0116] Final mega-sonic cleaning: Same as step 3, time 60s, to remove residual particles in deep pores;

[0117] Vacuum Marangoni drying: IPA (isopropanol) vapor replacement (concentration 99.9%), lifting speed 1mm / s, vacuum degree -80kPa, hot N2 purging (55℃, flow rate 10L / min) to avoid water residue causing pore wall contamination.

[0118] Step S6: Through-hole detection and closed-loop feedback.

[0119] A comprehensive testing system (integrating an optical profilometer, AFM, and ultrasonic flaw detector) was used to test the core quality indicators of TGV through-holes.

[0120] Aperture and morphology: The top / waist / bottom apertures were measured to be 50±2μm using an optical profilometer (Zygo NewView 9000), with roundness ≥0.95 and perpendicularity ≤0.5°.

[0121] Hole depth and aspect ratio: The ultrasonic flaw detector (Panametrics-NDT) measures a hole depth of 500±5μm and an aspect ratio of 10:1 (compliant with TGV packaging requirements).

[0122] Hole wall roughness: AFM (Bruker Dimension Icon) measured Ra=5nm (lower than 15nm in traditional processes);

[0123] Defect inspection: X-ray fluorescence spectroscopy showed F⁻ content < 1ppm (no residual etching solution), and ultrasonic testing showed no abnormal reflection signals (no microcracks).

[0124] Closed-loop feedback:

[0125] Acceptance criteria: 98 out of 100 through holes meet the above criteria, yield rate 98%; 2 NG products (due to minor cracks in the hole wall) are sent to the NG Port for scrapping.

[0126] Model fine-tuning: The actual etching rates (1.98-2.02 μm / min) of 98 qualified vias were collected and compared with the predicted v of S3. i Calculate the deviation (average deviation 0.03 μm / min);

[0127] Parameter update: Input the deviation data into the improved CNN model and perform a 5-step gradient update according to the MAML strategy. After the update, the prediction error of the model for the next batch of substrates is reduced to 3.8%, realizing continuous process optimization.

[0128] Implementation effect verification.

[0129] Compared with traditional techniques, the following table shows:

[0130]

[0131] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A method of improving yield of glass via, the method comprising: The method comprises the following steps: ​ S1: inducing a modification region on a glass substrate at a preset position by using a pulsed laser beam; S2: Obtain the three-dimensional point cloud data of the modification zone, and calculate the actual modification depth d of each modification zone i ; S3: input the three-dimensional point cloud data into the improved CNN model, predict the etching rate v of each modified area i , and generate an etching rate distribution map; S4: determining whether each modified region satisfies d i ≥ d0, where d0 is a preset threshold value: If there is a modified region of d i < d0, output the coordinates of the modified region and the depth compensation amount Δd i = d0- d i , return the glass substrate to S1, and adjust the laser parameters according to Δd i to induce the unqualified point again, and repeat S2-S4 until all d i ≥d0; If all the conditions are met, S5 is performed; S5: according to the etching rate distribution map generated in S3, the power distribution of the ultrasonic field is dynamically adjusted in the wet etching process, the ultrasonic power of the modification region with a predicted etching rate lower than the arithmetic mean of the etching rates of all modification regions on the glass substrate is increased, and etching is performed until the end point; S6: after the etching is completed, the state of the through hole is detected and a qualified or unqualified signal is output.

2. The method of claim 1, wherein: The improved CNN model is improved on the basis of the original CNN model, and the improvement points are: (1) the single serial feature extraction architecture of the original CNN model is reconstructed into a parallel double-branch architecture; (2) a multi-physical field encoding module is added to the input layer of the original CNN model, the laser pulse energy density, repetition frequency, defocusing amount and material properties of the glass substrate in S1 are encoded into an implicit physical field vector, and tensor product operation is used for fusion with three-dimensional point cloud data; (3) a differentiable etching dynamics constraint layer is added to the loss function of the original CNN model; (4) Replace the output layer of the original CNN model with a Bayesian nested variational inference architecture, and add a random dropout layer after each convolutional layer. During inference, the mean and standard deviation σ of the etching rate are output simultaneously through 10 random forward propagations i ; (5) an online meta-learning adaptation mechanism is added to the inference stage of the original convolutional neural network model, the actual etching rate and the prediction deviation fed back by S6 are used to quickly fine-tune the weights of the last two layers of the network through 5-step gradient update.

3. The method of claim 2, wherein: the glass via is formed by a method comprising: forming a glass via hole in a glass substrate; and forming a glass cap on the glass via hole. The parallel double-branch architecture includes a micro-topography branch and a macro-geometry branch, wherein the micro-topography branch adds a lightweight U-Net subnetwork to capture the roughness features of the modification region surface at a scale of 0.5-5μm, the macro-geometry branch adds a global depth gradient encoding layer to extract the anisotropy features of the overall topography of the modification region, and the double branches are fused at the channel level in a high-dimensional feature space.

4. The method of claim 2, wherein: the glass via is formed by a method selected from the group consisting of laser drilling, mechanical drilling, and combinations thereof. The constraint layer constructs a physical residual term based on the Arrhenius equation: L physics = λ||v i - f(C HF , T, d i )||², Wherein, L physics is a physical residual term for quantifying the deviation between the CNN prediction value and the physical theoretical value; λ is a weight coefficient; v i is the etching rate of the i th modified region; C HF is the molar concentration of HF in the S5 etching solution; T is the absolute temperature of the S5 etching process; f(·) is a theoretical calculation function of the etching rate, which is constructed based on the Arrhenius dynamics equation, and the specific form is: where k is a pre-exponential factor, is the activation energy, is the gas constant, is the reaction order, is the modification depth index.

5. The method of claim 2, wherein: the glass via is formed by a method selected from the group consisting of laser drilling, mechanical drilling, and combinations thereof. The online meta-learning adaptation mechanism adopts a double-layer optimization strategy based on model-agnostic meta-learning (MAML), the inner layer performs 5-step gradient descent on the weights of the last two layers of the network with a learning rate of 0.001 to quickly adapt to the current batch of data, and the outer layer updates the shared initial parameters of the network with a meta-learning rate of 0.01 to maintain the generalization ability of the model between different batches of glass substrates.

6. The method of claim 1, wherein: The dynamic adjustment of the power distribution of the ultrasonic field in S5 is realized by an 8x8 array ultrasonic transducer, the power adjustment range of a single transducer is 50-500W, the adjustment response time is ≤1 second, and the ultrasonic power calculation formula is: ​ ， where P base is the reference power, η is the modulation factor, is the arithmetic mean of the etching rates of all modified regions on the glass substrate, σ i is the standard deviation of the i-th modified region.

7. The method for improving the yield of glass through-holes as described in claim 1, characterized in that: The wet etching in S5 uses HF or KOH liquid system, the temperature is RT-50℃, the ultrasonic frequency is 40kHz and 80kHz, the power is 500-1000W, and the process time is 15-25min.

8. The method of claim 1, wherein: After the etching in S5 is completed, a multi-step cleaning and drying process is performed, which includes ultrapure water cleaning, ozone water cleaning, megasonic cleaning, final megasonic deep cleaning and vacuum maringoni drying in sequence. ​ 9. The method for improving the yield of glass through-holes as described in claim 1, characterized in that: The detection of the through hole state in S6 includes aperture and topography detection, hole depth and aspect ratio detection, hole wall roughness detection, residual etching liquid and micro-crack defect detection.

10. The method of claim 1, wherein: The three-dimensional point cloud data in S2 is acquired by a confocal microscope or a structured light three-dimensional scanner; and a preset threshold is 60%-80% of the thickness of the glass substrate. ​

Images