A machine learning based dark field imaging simulation and illumination parameter optimization method and system

Abstract

The application provides a kind of dark field imaging simulation and illumination parameter optimization method and system based on machine learning, belongs to optical detection field. In order to solve the problems of low optimization efficiency caused by artificial experience in the illumination parameters of the existing dark field detection technology, huge search calculation amount of theoretical simulation exhaustive method and lack of engineering practicability, the application establishes a "defect-illumination-imaging" parameter and scattering intensity correlation data set through wave optics physical simulation, and trains a machine learning agent model based on this, which replaces the time-consuming and laborious direct simulation process, thereby realizing the instantaneous prediction of scattering intensity. The application realizes the rapid optimization of illumination parameters, and through the systematic data-driven method, it can find the globally optimal illumination parameter combination, improve the signal-to-noise ratio and sensitivity of detection, and has high flexibility, which is suitable for different application scenarios.

Description

Technical Field

[0001] This invention relates to the field of optical detection technology, and more specifically, to a method and system for dark-field imaging simulation and illumination parameter optimization based on machine learning. Background Technology

[0002] Precision optical components have extremely high requirements for surface condition, such as surface damage and cleanliness. If surface damage is not detected in time, it will not only generate harmful scattered and stray light, severely reducing the imaging contrast and beam quality of the optical system, but also induce irreversible laser-induced damage (LIDT) in high-energy laser applications, or even lead to the failure of the entire system. Therefore, the fundamental purpose of surface inspection is to effectively intercept these potential hazards before assembly and application, eliminate failure risks through strict quality control, and ensure the wavefront integrity, energy transmission efficiency and long-term operational stability of the optical system.

[0003] In defect detection, especially for weakly scattering surfaces at the micrometer or even submicrometer scale, dark-field illumination detection is a crucial and widely used technique. The principle behind this technique is to use illumination optics at a specific angle to prevent normally reflected light from entering the imaging system, allowing only light scattered by the defect to enter and be received by the detector. This creates a bright image of the defect against a dark background. Dark-field detection, with its high signal-to-noise ratio and scattering amplification effect, makes the detection of micro-defects much easier. Figure 1 The diagram shown illustrates the principle of dark-field imaging.

[0004] However, the implementation of existing dark field detection technologies faces the following technical bottlenecks:

[0005] 1. Detection results are highly sensitive to illumination parameters: The signal-to-noise ratio (SNR) of dark-field imaging, i.e., the intensity of the defect scattering signal, is not constant but closely related to parameters such as the incident angle, wavelength, and polarization state of the illumination light. The optimal combination of illumination parameters varies significantly depending on the size and morphology of the defect. Inappropriate parameter combinations may result in weak defect signals or even failure to detect them, leading to missed detections.

[0006] 2. Outdated and inefficient methods for optimizing lighting parameters: Currently, determining lighting parameters mainly relies on two approaches:

[0007] The trial-and-error method involves experienced technicians manually adjusting and experimenting based on past experience. This method is highly subjective, has poor repeatability, is time-consuming and labor-intensive, and can hardly guarantee that it finds the globally optimal parameters, only "usable" ones.

[0008] Theoretical simulation exhaustive search method: Although wave optics simulation method can be used to accurately analyze the scattering characteristics of defects, the illumination parameters (angle, wavelength, polarization) and defect parameters (size, depth) constitute a complex space. Exhaustive simulation search requires a huge amount of computing resources and time, and is not practical for engineering. Summary of the Invention

[0009] The technical problem to be solved by this invention is:

[0010] To address the issues of low optimization efficiency caused by the reliance on human experience for illumination parameters in existing dark field detection technologies, and the huge computational burden and lack of engineering practicality of theoretical simulation exhaustive search methods.

[0011] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0012] This invention provides a machine learning-based method for dark-field imaging simulation and illumination parameter optimization, comprising the following steps:

[0013] S100. Generating a machine learning dataset through simulation, including: establishing a surface model of an optical element with specific geometric parameters using finite-difference time-domain (FDTD) wave optics simulation; defining a multidimensional input parameter space; generating multiple sets of parameter combinations in the multidimensional input parameter space using the Latin hypercube sampling uniform design method; post-processing the acquired near-field electromagnetic field data to simulate the physical processes of a real imaging system; and pairing the above parameter combinations with the calculated scattering intensity to form a training dataset containing multiple data points.

[0014] S200 uses the CatBoost machine learning algorithm for model training;

[0015] S300. Perform scattering prediction and application of lighting parameters, including: analyzing the importance of features affecting scattering intensity to obtain the adjustment priority of lighting parameters; predicting the trend of scattering intensity changes, and then obtaining the predicted value of scattering intensity.

[0016] Further, in step S100, the following is included:

[0017] S110. Using finite-difference time-domain (FDTD) wave optics simulation, a surface model of an optical element with a pit having specific geometric parameters is established; the total field scattering field (TFSF) light source is selected; the monitor is set in the upper half of the scattering field to record the electric field intensity; the perfectly matched layer (PML) is selected as the FDTD boundary condition to absorb electromagnetic waves; the simulation time is longer than the longest simulation convergence time.

[0018] S120. Define a multidimensional input parameter space, including defect radius R, defect depth H, light source incident angle θ, light source wavelength λ, and light source polarization state Pol.

[0019] S130. Using the Latin hypercube sampling uniform design method, 1000 sets of parameter combinations are generated in the multidimensional input parameter space. By dividing the probability space of the multidimensional variables into equal probability intervals and ensuring that the samples of each dimension are uniformly distributed, and performing electromagnetic simulation on each combination, the near-field scattering electromagnetic field distribution of the defect is obtained.

[0020] S140. Post-process the acquired near-field electromagnetic field data, including Fourier transform, NA filtering, and data point generation.

[0021] S150. The above parameters are combined as input features and the calculated scattering intensity. As output labels, they are paired to form a training dataset containing multiple data points.

[0022] Further, in step S140, the following is included:

[0023] S141. Fourier Transform: Perform a two-dimensional Fast Fourier Transform on the simulated near-field electric field to obtain its distribution in the spatial frequency domain. ;

[0024] S142, NA filtering and data point generation

[0025] Define the numerical aperture to be examined. According to the current Using the wavelength λ used in this finite-difference time-domain (FDTD) simulation, calculate the maximum spatial frequency that the imaging system can collect. Create a digital circular low-pass filter with a radius of... ;Will All spatial frequencies greater than The spectral components are set to zero to obtain the filtered spectrum. ; the filtered spectrum Performing a two-dimensional inverse fast Fourier transform yields the following result: Complex amplitude electric field distribution on the ideal image plane under simulated conditions ; Calculate the electric field intensity on the image plane ;

[0026] Integrating over the entire image plane yields the total scattered electric field intensity. .

[0027] Further, in step S300, the following are included:

[0028] S310. Feature importance analysis affecting scattering intensity: By using the feature importance scores of illumination parameters output by the machine learning model, the influence of each parameter on the final dark field imaging signal-to-noise ratio is quantified, thereby guiding the priority of adjusting illumination parameters in engineering practice.

[0029] S320, predicts the trend of scattering intensity. The Catboost model input parameters include incident angle, wavelength, polarization, defect depth, and defect radius, and the output is the predicted value of scattering intensity.

[0030] Furthermore, surface micro-defects were detected on the fused silica optical components. The surface micro-defects were triangular pyramidal pits. Using the dark field imaging simulation and illumination parameter optimization method, the optimal combination of engineering illumination parameters was obtained as follows: incident angle of 35° and 75°, S polarization state, and wavelength of 450nm.

[0031] A machine learning-based dark-field imaging simulation and illumination parameter optimization system is provided. The system has program modules corresponding to the above steps and executes the steps in the above-described machine learning-based dark-field imaging simulation and illumination parameter optimization method during runtime.

[0032] A computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement the steps of a machine learning-based dark-field imaging simulation and illumination parameter optimization method.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] High efficiency: This invention shortens the traditional simulation optimization process, which takes hours or even days, to a model prediction process in milliseconds, enabling rapid optimization of lighting parameters and possessing high engineering application value.

[0035] Performance advantages: This invention uses a systematic data-driven approach to find the globally optimal combination of illumination parameters, rather than relying on locally optimal solutions obtained through experience. This maximizes the improvement of defect scattering signals and enhances the signal-to-noise ratio and sensitivity of detection.

[0036] Highly adaptable: This method incorporates input variables such as defect size, incident angle, and wavelength into the model, thus enabling it to flexibly adapt to different detection tasks and hardware configurations, providing customized optimal lighting solutions for specific scenarios.

[0037] Technical simplification: This invention reduces the reliance on long-term simulations of computing clusters and also reduces the requirements for operators' optical engineering experience, making high-performance defect detection more widespread and economical. Attached Figure Description

[0038] Figure 1This is a schematic diagram of the dark field imaging principle in an embodiment of the present invention;

[0039] Figure 2 This is a simulation model diagram of defect scattering in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of near-field and far-field extrapolation within a finite aperture for defect scattering simulation in an embodiment of the present invention;

[0041] Figure 4 This is a flowchart of step S100 in an embodiment of the present invention, which involves generating a machine learning dataset through simulation.

[0042] Figure 5 This is a flowchart of the machine learning model training process in step S200 of this embodiment of the invention;

[0043] Figure 6 This is an example of the incident angle-scattering intensity diagram when R=2μm and H=1μm in this embodiment of the invention;

[0044] Figure 7 This is a flowchart illustrating the optimization of optimal light source parameters in an embodiment of the present invention.

[0045] Figure 8 The diagram shows the variation of scattering intensity with incident angle of light source with different polarization states in the embodiments of the present invention. Among them, (a) is the incident angle-scattering intensity diagram when R=4μm and H=1.5μm, and (b) is the incident angle-scattering intensity diagram when R=4μm and H=4μm. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Specific Implementation Plan 1: Combining Figures 1 to 7 As shown, this invention provides a machine learning-based method for dark-field imaging simulation and illumination parameter optimization, comprising the following steps:

[0048] S100, combined Figures 2 to 4 As shown, a machine learning dataset is generated through simulation;

[0049] Specifically, including,

[0050] S110. FDTD (Finite-Difference Time-Domain) wave optics simulation is used to establish a surface model of an optical element with specific geometric parameters featuring concave pits. The substrate material is fused silica, and the defects are triangular-section cone-shaped pits made of air. Since only scattering is considered, a TFSF (total-field scattered-field) light source is selected. This light source cannot use a broad spectrum when incident obliquely, and both ends of the TFSF have identical structures, refractive indices, and reflectivities. The monitor is placed in the upper half of the scattered field to record the electric field intensity. The FDTD boundary condition is PML (Perfectly Matched Layer) to absorb electromagnetic waves. The simulation time needs to be longer than the maximum convergence time. Other regions are made of air. Figure 2 The image shows a simulation model of defect scattering.

[0051] S120. Define a multidimensional input parameter space, including defect radius R, defect depth H, incident angle of light source θ, light source wavelength λ, and polarization state of light source Pol. The range of values ​​for each parameter is shown in Table 1.

[0052] Table 1. Value range of each input parameter

[0053]

[0054] S130. Using the Latin hypercube sampling uniform design method, 1000 sets of parameter combinations are generated in the multidimensional input parameter space.

[0055] Latin hypercube sampling is a stratified random sampling technique that divides the probability space of multidimensional variables into equally probable intervals and ensures that the samples in each dimension are uniformly distributed. It also performs electromagnetic simulation on each combination to obtain the near-field scattered electromagnetic field distribution of defects.

[0056] S140. Post-processing is performed on the acquired near-field electromagnetic field data. This post-processing simulates the physical processes of a real imaging system, including:

[0057] S141. Fourier Transform: Perform a two-dimensional Fast Fourier Transform (FFT) on the near-field electric field obtained from this simulation to obtain its distribution in the spatial frequency domain. ;

[0058] S142, NA filtering and data point generation

[0059] Define the numerical aperture to be examined. According to the current Using the wavelength λ used in this FDTD simulation, calculate the maximum spatial frequency that the imaging system can collect. Create a digital circular low-pass filter (simulating an objective aperture) with a radius of... ;Will All spatial frequencies greater than (Right now The spectral components of the filter are set to zero to obtain the filtered spectrum. ; the filtered spectrum Performing a two-dimensional inverse fast Fourier transform (IFFT) yields the result in the... Complex amplitude electric field distribution on the ideal image plane under simulated conditions ; Calculate the electric field intensity on the image plane ;

[0060] like Figure 3 The diagram shows a near-field extrapolation to the far-field within a finite aperture in a defect scattering simulation. Integrating over the entire image plane yields the total scattered electric field intensity. This will be denoted as "scattering intensity";

[0061] S150. Combine the above parameters (as input features) with the calculated scattering intensity. (As output labels) are paired to form a training dataset containing multiple data points;

[0062] S200, combined Figure 5 As shown, train the machine learning model and validate it;

[0063] Specifically, including,

[0064] During the model training and validation phases, the scattering intensities obtained from the simulation were preprocessed using logarithmic transformation, and the dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The CatBoost machine learning algorithm was used for model training. The model evaluation results after training showed that the model's coefficient of determination R on the training set was [value missing]. 2 The R-value reached 0.98 on both the validation and test sets. 2 The values ​​are 0.921 and 0.919 respectively; the high consistency between the test set and validation set metrics indicates that the surrogate model has excellent generalization ability, no obvious overfitting phenomenon, and can replace computationally expensive physical simulation to achieve high-precision and rapid prediction.

[0065] S300, combined Figure 6 and Figure 7 As shown, the scattering prediction and application of lighting parameters are carried out;

[0066] Specifically, including,

[0067] S310. Analysis of the importance of features affecting scattering intensity

[0068] By using the importance scores of illumination parameters output by the machine learning model, the influence of each parameter on the final signal-to-noise ratio of dark-field imaging can be quantified, thereby guiding the priority of adjusting illumination parameters in engineering practice. Model analysis shows that scattering intensity has the highest sensitivity to the incident angle of the light source (relative importance score of approximately 50.7), followed by defect depth (approximately 22.34) and defect radius (approximately 21.06), while the influence of light source wavelength (approximately 3.71) and polarization state (less than 1.5) is relatively small. Therefore, in the actual debugging of industrial dark-field detection equipment, the preferred optimization strategy is to first determine and lock the optimal incident angle of the light source, and then fine-tune the wavelength and polarization state, thereby significantly improving the efficiency of on-site debugging.

[0069] S320, predict the trend of scattering intensity changes.

[0070] The Catboost model takes the following input parameters: incident angle, wavelength, polarization, defect depth, and defect radius. The output is a predicted scattering intensity. By inputting specified parameters, the model can quickly provide the scattering intensity under given conditions, allowing for the calculation of how the scattering intensity changes with these parameters and the plotting of the curve. Figure 6 As shown.

[0071] Specific implementation scheme two: The present invention provides a machine learning-based dark field imaging simulation and illumination parameter optimization system. The system has program modules corresponding to the above steps, and executes the steps in the above-mentioned machine learning-based dark field imaging simulation and illumination parameter optimization method when running.

[0072] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0073] Specific Implementation Scheme 3: The present invention provides a computer-readable storage medium storing a computer program configured to implement, when called by a processor, the steps of a machine learning-based dark field imaging simulation and illumination parameter optimization method.

[0074] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0075] Example

[0076] For practical tasks involving the detection of micro-defects on the surface of optical components such as fused silica, the defect radius R is set to range from 1 μm to 5 μm, and the depth H to range from 0.1 μm to 5 μm. Due to the numerical aperture limitations of practical industrial camera imaging systems (e.g., NA=0.5 corresponds to a light collection solid angle of 30°), to avoid specular reflection light directly hitting the lens and reducing contrast, the optimal adjustment range for the light source incident angle θ is set between 35° and 90°. Using a trained machine learning model for full-parameter spatial prediction, the following rules guiding the configuration of practical lighting parameters are derived:

[0077] Incident Angle Selection Strategy: For defects with small radii (R < 3 μm) and some medium-sized defects (3 μm ≤ R ≤ 5 μm and 0.1 μm ≤ H ≤ 2 μm), the scattering intensity of the defects decreases significantly with increasing incident angle. For deep defects (2 μm < H ≤ 5 μm), when the incident angle is greater than 45°, the scattering intensity exhibits a non-linear change pattern of first increasing and then decreasing with increasing angle. To accommodate defects of different sizes and tolerate an installation error of approximately ±5° in actual equipment assembly, a dual-angle illumination combination is preferred: 35° is used as the primary incident angle to ensure a high signal-to-noise ratio for small defects, supplemented by 75° as the secondary incident angle to ensure effective detection of large, deep pit defects.

[0078] Polarization state selection strategy: Under specific incident angles of 35° and 75°, comparing the prediction results under P-polarization, S-polarization, and circular polarization (C) conditions, S-polarized light can excite the strongest defect scattering signal. Therefore, the preferred polarization state configuration for the light source is S-polarization.

[0079] Wavelength selection strategy: Theoretical predictions show that the shorter the wavelength of the illumination light, the stronger the defect scattering effect. However, in practical engineering applications, although 380nm violet light has the strongest theoretical scattering, considering that the quantum efficiency of most standard industrial cameras (CMOS / CCD) drops sharply in the violet light band, and that violet LEDs are expensive and have low electro-optical conversion efficiency, 450nm blue light is preferred as the detection light source wavelength, taking into account both hardware photosensitive characteristics and economic efficiency.

[0080] In summary, for the detection of surface micro-defects in this embodiment, the optimal combination of engineering lighting parameters selected by the machine learning proxy model is: an incident angle of 35° and 75°, S-polarization state, and a wavelength of 450nm.

[0081] It should be particularly noted that although this embodiment uses a triangular pyramidal pit on the substrate surface as an example for detailed description, this is only for clearly illustrating the technical solution of the present invention and is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that the machine learning-based dark field illumination parameter optimization method proposed in this invention is also fully applicable to other types, morphologies, and sizes of weak scatterers or defects on optical surfaces.

[0082] For example, when implementing the method of this invention for other common defect types on the surface of optical components, such as submicron scratches, micro-protrusions, polishing marks, and adhering impurities, the geometric model of the defect is simply replaced with the corresponding three-dimensional morphology in the FDTD wave optical simulation stage of step S100. This generates a scattering feature dataset for the corresponding defect type, and the machine learning proxy model described in this invention can then be used for feature training and parameter optimization. Therefore, any equivalent substitutions or model extensions made based on the core data-driven idea of ​​this invention for different morphological defects do not depart from the spirit and scope of this invention and should be included within the protection scope of this invention.

[0083] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A machine learning-based method for dark-field imaging simulation and illumination parameter optimization, characterized in that, Includes the following steps: S100. Generating a machine learning dataset through simulation, including: establishing a surface model of an optical element with specific geometric parameters using finite-difference time-domain (FDTD) wave optics simulation; defining a multidimensional input parameter space; generating multiple sets of parameter combinations in the multidimensional input parameter space using the Latin hypercube sampling uniform design method; post-processing the acquired near-field electromagnetic field data to simulate the physical processes of a real imaging system; and pairing the above parameter combinations with the calculated scattering intensity to form a training dataset containing multiple data points. S200 uses the CatBoost machine learning algorithm for model training; S300. Perform scattering prediction and application of lighting parameters, including: analyzing the importance of features affecting scattering intensity to obtain the adjustment priority of lighting parameters; predicting the trend of scattering intensity changes, and then obtaining the predicted value of scattering intensity.

2. The method for dark-field imaging simulation and illumination parameter optimization based on machine learning according to claim 1, characterized in that: In step S100, the following are included: S110. Using finite-difference time-domain (FDTD) wave optics simulation, a surface model of an optical element with a pit having specific geometric parameters is established; the total field scattering field (TFSF) light source is selected; the monitor is set in the upper half of the scattering field to record the electric field intensity; the perfectly matched layer (PML) is selected as the FDTD boundary condition to absorb electromagnetic waves; and the simulation time is longer than the longest simulation convergence time. S120. Define a multidimensional input parameter space, including defect radius R, defect depth H, light source incident angle θ, light source wavelength λ, and light source polarization state Pol. S130. Using the Latin hypercube sampling uniform design method, 1000 sets of parameter combinations are generated in the multidimensional input parameter space. By dividing the probability space of the multidimensional variables into equal probability intervals and ensuring that the samples of each dimension are uniformly distributed, and performing electromagnetic simulation on each combination, the near-field scattering electromagnetic field distribution of the defect is obtained. S140. Post-process the acquired near-field electromagnetic field data, including Fourier transform, NA filtering, and data point generation. S150. The above parameters are combined as input features and the calculated scattering intensity. As output labels, they are paired to form a training dataset containing multiple data points.

3. The method for dark-field imaging simulation and illumination parameter optimization based on machine learning according to claim 2, characterized in that: In step S140, the following are included: S141. Fourier Transform: Perform a two-dimensional Fast Fourier Transform on the simulated near-field electric field to obtain its distribution in the spatial frequency domain. ; S142, NA filtering and data point generation Define the numerical aperture to be examined. According to the current Using the wavelength λ used in this finite-difference time-domain (FDTD) simulation, calculate the maximum spatial frequency that the imaging system can collect. Create a digital circular low-pass filter with a radius of... ;Will All spatial frequencies greater than The spectral components are set to zero to obtain the filtered spectrum. ; Filtered spectrum Performing a two-dimensional inverse fast Fourier transform yields the following result: Complex amplitude electric field distribution on the ideal image plane under simulated conditions ; Calculate the electric field intensity on the image plane ; Integrating over the entire image plane yields the total scattered electric field intensity. .

4. The method for dark-field imaging simulation and illumination parameter optimization based on machine learning according to claim 3, characterized in that: Step S300 includes, S310. Feature importance analysis affecting scattering intensity: By using the feature importance scores of illumination parameters output by the machine learning model, the influence of each parameter on the final dark field imaging signal-to-noise ratio is quantified, thereby guiding the priority of adjusting illumination parameters in engineering practice. S320, predicts the trend of scattering intensity. The Catboost model input parameters include incident angle, wavelength, polarization, defect depth, and defect radius, and the output is the predicted value of scattering intensity.

5. The method for dark-field imaging simulation and illumination parameter optimization based on machine learning according to claim 4, characterized in that: Surface micro-defects were detected on fused silica optical components. The surface micro-defects were triangular pyramidal pits. Using the dark field imaging simulation and illumination parameter optimization method, the optimal combination of engineering illumination parameters was obtained as follows: incident angle of 35° and 75°, S polarization state, and wavelength of 450nm.

6. A machine learning-based dark-field imaging simulation and illumination parameter optimization system, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-5 above, and executes the steps in the above-described machine learning-based dark field imaging simulation and illumination parameter optimization method when running.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of any one of claims 1-5: a machine learning-based dark-field imaging simulation and illumination parameter optimization method.

Images