CN-122021269-A - Narrow-band filter device design method and device based on machine learning

CN122021269ACN 122021269 ACN122021269 ACN 122021269ACN-122021269-A

Abstract

The invention relates to the fields of computer science and technology and micro-nano optics, in particular to a narrow-band filter device design method based on machine learning. And fitting the simulated spectrum to obtain excitation peak parameters, so as to establish a corresponding relation table of simulated spectrum data and the excitation peak parameters. And setting an optimization target based on the corresponding relation table, obtaining a candidate parameter set by utilizing machine learning, and obtaining a representative parameter set for guiding the production of the device by utilizing multi-target optimization. The invention also includes devices produced by the above design method. The invention integrates the machine learning intelligent algorithm into the design flow of the photonic device to form a data driving and automatizing method, solves the problems of low efficiency, poor precision and poor expandability in the design of the high-performance narrow-band filter device, and provides a powerful technical tool for the rapid development of the integrated and miniaturized optical device.

Inventors

LI GUANHAI
ZHOU YIRAN
ZHANG YUKANG
JIN RONG
YU FEILONG
CHEN Jin
GUO JIAOYANG
Fu Zhenchu

Assignees

中国科学院上海技术物理研究所

Dates

Publication Date: 20260512
Application Date: 20260113

Claims (10)

1. The design method of the narrow-band filter device based on machine learning is characterized by comprising the following steps: step S100, selecting a design carrier of a narrow-band filter, wherein the design carrier defines the selectable range of the structural parameters of the narrow-band filter; step S200, selecting different structural parameters in the optional range, and obtaining a first simulation spectrum data set corresponding to the structural parameters through simulation calculation; step S300, based on the first simulation spectrum data set, obtaining excitation peak parameters of a spectrum by fitting, wherein the excitation peak parameters comprise line width parameters and peak position parameters; Step S400, combining the structural parameters with the excitation peak parameters obtained from the first simulation spectrum data set to establish a corresponding relation table; step S500, setting an optimization target according to the excitation peak parameters, performing machine learning based on the corresponding relation table to obtain a candidate parameter set composed of the structure parameters, wherein a spectrum formed by the values of the structure parameters in the candidate parameter set corresponds to the excitation peak parameters, and a spectrum formed by the values of the structure parameters is locally superior to the excitation peak parameters; And step S600, performing multi-objective optimization based on the candidate parameter set to form a representative parameter set, wherein the values of the structural parameters in the representative parameter set are corresponding to the excitation peak parameters, and the formed spectrum is globally superior to the spectra formed by the values of other structural parameters and corresponds to the excitation peak parameters.
2. The method according to claim 1, wherein in step S100, the design carrier is a square lattice super surface resonator, and the design carrier is composed of a nano-pillar structure in a periodic lattice, and the structural parameters include a lattice period and a structural dimension of the nano-pillar structure.
3. The machine learning based narrow band filter design method according to claim 1, comprising the sub-steps of: step S301, using an adaptive background fitting function, fitting a first background set based on the first simulation spectrum data set; step S302, based on the first background set, screening out points where significant peaks are located from the first simulation spectrum data set, and recording the points as a first significant peak data point set; step S303, replacing data points corresponding to the significant peaks and N adjacent data points with null values for each significant peak in the first simulation spectrum data set to form a second simulation spectrum data set; step S304, fitting out a second background set of the marked background based on the second simulation spectrum data set based on a smooth spline background fitting function; Step S305, subtracting the first simulated spectrum data set from the second background set, detecting the point where the significant peak is located again, and recording the point as a second significant peak data point set; Step S306, selecting a data point corresponding to the significant peak and N adjacent data points aiming at each significant peak in the second significant peak data point set, and performing peak fitting on the data points to obtain a peak fitting result corresponding to each significant peak; Step S307, traversing each wave crest fitting result to obtain the line width parameter and the peak position parameter of the remarkable peak, and recording the line width parameter and the peak position parameter as the excitation peak parameter; Step S308, traversing each peak fitting result, and reserving the peak fitting result with the largest peak and the corresponding excitation peak parameter for a plurality of peak fitting results with overlapping degrees larger than an overlapping threshold.
4. The method for designing a narrow-band filter based on machine learning according to claim 3, wherein in step S301, the adaptive background fitting function is polynomial fitting, and the polynomial fitting is increased from 0 th order until the polynomial fitting residual meets fitting conditions, and finally a final residual, a best polynomial coefficient and a best polynomial order are output.
5. The method for designing a narrow-band filter based on machine learning according to claim 3, wherein in step S304, the smooth spline background fitting function includes a smoothing factor, the fitted curve passes through all data points when the smoothing factor is 0, the fitted curve balances the fitting degree and smoothness when the smoothing factor is greater than 0, and the smooth spline background fitting function generates new independent variables and new dependent variables based on effective data points and finally outputs a fitting value.
6. The machine learning based narrow band filter design method of claim 1, wherein in step S306, the peak fitting uses a Fano linear fitting formula expressed as: 。
7. The method for designing a narrow-band filter based on machine learning according to claim 1, wherein in step S500, a sub-step is included, Step S501, setting optimization targets as excitation peak positions and excitation peak line widths, wherein the excitation peak positions are required to be as close to preset peak positions as possible, and the excitation peak structure line widths are required to be as small as possible; Step S502, partitioning a parameter space in which the structural parameters in the corresponding relation table are located to form a plurality of parameter subspaces; Step S503, traversing each parameter subspace, and carrying out global search in each parameter subspace by applying a Gaussian process regression algorithm to find subspace candidate solutions of each parameter subspace; in step S504, all the subspace candidate sets form the candidate parameter set.
8. The machine learning based narrow band filter design method of claim 1, wherein in step S600, the multi-objective optimization is performed using a pessary pareto optimization algorithm, comprising the sub-steps of: Step S601, grouping the candidate parameter sets to obtain candidate parameter sets; Step S602, selecting, from each candidate parameter subset, one or more values of the structural parameters that perform well for a plurality of factors in the optimization objective, as an optimal parameter combination of the candidate parameter subset; Step S603, based on the obtaining of the optimal parameter combinations by all the candidate parameter subsets, using pessary pareto optimization algorithm to find out the optimal parameter combinations which are not completely surpassed by other optimal parameter combinations on all factors in the optimization target, and using the optimal parameter combinations as leading edge output; Step S604, selecting the optimal parameter combination from the leading edge outputs by using a sample selection algorithm, to form the representative parameter set.
9. A narrow-band filter device implemented using the machine learning based narrow-band filter device design method of claim 1, characterized in that it is a square lattice super-surface resonator using a multilayer stacked system, the substrate layer of the resonator is located at the bottom layer, the periodic lattice layer is located above the substrate layer and is arranged along the vertical direction, and each lattice layer contains nano-pillar structures and is arranged in three-dimensional periodicity.
10. The machine learning based narrow band filter device design method and device of claim 9, wherein the substrate layer is made of aluminum oxide and the lattice layer is made of silicon.

Description

Narrow-band filter device design method and device based on machine learning Technical Field The invention relates to the fields of computer science and technology and micro-nano optics, in particular to a narrow-band filter device design method and device based on machine learning. Background The narrow-band filter has the effects of improving the communication capacity, performing high-precision detection, analyzing microscopic information, guaranteeing information safety and other key fields of modern technology. In recent years, narrow-band filter devices designed by using metamaterials have appeared. The continuous Constraint State (BICs) can realize extremely high quality factor (Q factor) and strong light field localization in an open system, and provides a brand new physical mechanism for breakthrough of the performance of the photonic device. The optical super-surface periodic unit structure can be designed, so that the optical super-surface periodic unit structure becomes a big research hot spot of micro-nano photonics in recent years, and the function of narrow-band filtering can be easily realized. However, the traditional quasi-BIC design relies on intuitive parameter scanning with symmetrical breakage, has the bottlenecks of high resource consumption, low optimization precision and the like, and is difficult to meet the directional requirements of industrial scenes on multidimensional indexes such as excitation peak positions, line widths, peak values and the like. In the existing research work, the continuous constraint state structure optimization mostly uses a large-scale parameter scanning and experience, and the traditional optimization method, such as gradient descent and genetic algorithm, faces two major dilemmas when dealing with a high-dimensional parameter space, namely, the time complexity of electromagnetic simulation calculation grows exponentially along with the parameter dimension, and the time cost is estimated:, Two, the topology protection characteristics of BIC result in a large number of local extrema of the objective function, about And each. This severely constrains the development efficiency of high performance subsurface devices, and therefore, how to achieve rapid optimization of the structure, and thus achieving high-throughput optical output, becomes a difficulty. Disclosure of Invention The invention aims to provide a narrow-band filter device design method and device based on machine learning, which mainly solve the problems in the prior art. In order to achieve the above object, the present invention provides a method for designing a narrow-band filter based on machine learning, which is characterized by comprising the steps of: step S100, selecting a design carrier of a narrow-band filter, wherein the design carrier defines the selectable range of the structural parameters of the narrow-band filter; step S200, selecting different structural parameters in the optional range, and obtaining a first simulation spectrum data set corresponding to the structural parameters through simulation calculation; step S300, based on the first simulation spectrum data set, obtaining excitation peak parameters of a spectrum by fitting, wherein the excitation peak parameters comprise line width parameters and peak position parameters; Step S400, combining the structural parameters with the excitation peak parameters obtained from the first simulation spectrum data set to establish a corresponding relation table; step S500, setting an optimization target according to the excitation peak parameters, performing machine learning based on the corresponding relation table to obtain a candidate parameter set composed of the structure parameters, wherein a spectrum formed by the values of the structure parameters in the candidate parameter set corresponds to the excitation peak parameters, and a spectrum formed by the values of the structure parameters is locally superior to the excitation peak parameters; And step S600, performing multi-objective optimization based on the candidate parameter set to form a representative parameter set, wherein the values of the structural parameters in the representative parameter set are corresponding to the excitation peak parameters, and the formed spectrum is globally superior to the spectra formed by the values of other structural parameters and corresponds to the excitation peak parameters. Further, in step S100, the design carrier is a square lattice super-surface resonator, and is composed of nano-pillar structures in a periodic lattice, and the structural parameters include lattice period and structural dimensions of the nano-pillar structures. Further, in step S300, the sub-steps are included: step S301, using an adaptive background fitting function, fitting a first background set based on the first simulation spectrum data set; step S302, based on the first background set, screening out points where significant peaks are located from the first