A tunable bidirectional metamaterial optical absorption device and methods of optimization and fabrication thereof

Abstract

The application discloses a tunable bidirectional metamaterial light absorption device and an optimization and preparation method thereof, and relates to the technical field of metamaterial optical absorption devices; the bidirectional absorption device is a six-layer structure of metal-dielectric-metal-dielectric-metal-dielectric, and the device units thereof are sequentially a silver metal layer, a liquid crystal dielectric layer, a silver metal layer, a liquid crystal dielectric layer, a chromium metal layer and a silicon nitride ring from bottom to top; the bidirectional light absorption device can realize wideband absorption on top incident plane light and narrowband absorption on bottom incident plane light in a wavelength range of 0.6-1.6 microns. By changing the liquid crystal molecule deflection angle in the liquid crystal dielectric layer, the absorption rate and the absorption bandwidth range of the device on the top or bottom incident light can be adjusted and controlled, so that tunability is realized. The application can realize the structural parameter optimization design of the device under multiple target constraints.

Description

A tunable bidirectional metamaterial optical absorption device and its optimization and fabrication method Technical Field

[0001] This invention relates to the field of metamaterial optical absorption device technology, specifically to a tunable bidirectional metamaterial optical absorption device and its optimization and fabrication method. Background Technology

[0002] Metamaterials are artificial materials with unique electromagnetic properties, exhibiting phenomena such as negative refractive index and perfect absorption. Compared with traditional electromagnetic absorbers, metamaterial absorbers have significant advantages such as flexible design and simple structure, and can be widely used in fields such as solar energy collection, stealth technology, and photoelectric detection.

[0003] Currently, existing metamaterial absorbers are typically designed to achieve broadband or narrowband absorption at specific frequencies and incident angles, but they generally only support unidirectional absorption, which greatly limits their practical applications.

[0004] Therefore, the development of bidirectional metamaterial optical absorbers has become a research hotspot in the field of optical absorption devices. The bidirectional switching capability between broadband and narrowband absorption greatly expands the application potential of bidirectional absorbers. However, the inherent coupling effect between the material parameters of each layer of the bidirectional absorber has a complex and mutually restrictive impact on the device performance. Compared with the design of unidirectional metasurface absorbers, the trade-off optimization to achieve the dual absorption target heavily relies on the expertise of researchers and requires more simulation time and computational resources. Summary of the Invention

[0005] In view of this, the present invention provides a tunable bidirectional metamaterial light-absorbing device that can absorb bidirectional incident light. By changing the deflection angle of liquid crystal molecules in the liquid crystal dielectric layer, the absorption rate and absorption bandwidth range of the light-absorbing device for bidirectional incident planar light can be controlled.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention proposes a tunable bidirectional metamaterial optical absorption device comprising a six-layer unit stacked structure of metal-dielectric-metal-dielectric-metal-dielectric, wherein the unit is arranged sequentially from the bottom layer to the top layer as follows:

[0008] The first layer is a silver metal layer with a thickness of H1, ranging from 30 to 120 nm.

[0009] The second layer is a liquid crystal dielectric layer with a thickness of H2 and a thickness range of 110-250 nm;

[0010] The third layer is a silver metal layer with a thickness of H3, ranging from 30 to 120 nm.

[0011] The fourth layer is a liquid crystal dielectric layer with a thickness of H4 and a thickness range of 110-250nm;

[0012] The fifth layer is a chromium metal layer with a thickness of H5 and a thickness range of 10-30 nm;

[0013] The silicon nitride ring located in the sixth layer has a thickness of H6 and a thickness range of 110-250 nm;

[0014] Preferably, the silver metal layer, the liquid crystal dielectric layer, the chromium metal layer and the silicon nitride ring have the same period, with period P being 0.6 μm.

[0015] Preferably, the inner radius R1 of the silicon nitride ring ranges from 0 to 100 nm, and the outer radius R2 ranges from 150 to 280 nm.

[0016] The tunable bidirectional metamaterial optical absorption device described above can achieve bidirectional optical absorption in the 0.6-1.6μm wavelength band, specifically by performing broadband absorption of the forward plane light incident from the top and narrowband absorption of the reverse plane light incident from the bottom.

[0017] The first silver metal layer, the second liquid crystal dielectric layer, and the third silver metal layer are used to regulate the absorption rate of the device for bottom incident light; the third silver metal layer, the fourth liquid crystal dielectric layer, the fifth chromium metal layer, and the sixth silicon nitride ring are used to regulate the absorption rate of the device for top incident light.

[0018] Furthermore, the deflection angle of the liquid crystal molecules in the second and fourth liquid crystal dielectric layers can be changed, with the deflection angle ranging from 0 to 90°. By changing the deflection angle of the liquid crystal molecules in the second and fourth liquid crystal dielectric layers, the absorption rate and absorption bandwidth of the device for bottom or top incident plane light can be adjusted respectively, thereby realizing the tunable characteristics of the bidirectional metamaterial light absorption device.

[0019] The first and third silver metal layers are made of silver; the second and fourth liquid crystal dielectric layers are made of liquid crystal; the fifth chromium metal layer is made of chromium; and the sixth silicon nitride ring is made of silicon nitride.

[0020] Secondly, this invention proposes an optimization method for a tunable bidirectional metamaterial optical absorption device, comprising the following steps:

[0021] S10: Constructing a prediction model for the absorption spectrum of bidirectional metamaterial optical absorption devices using artificial intelligence semi-supervised learning technology;

[0022] S20: Constructing reinforcement learning intelligent agents through artificial intelligence reinforcement learning technology;

[0023] S30: By combining the absorption spectrum prediction model of the bidirectional metamaterial optical absorption device with the artificial intelligence reinforcement learning agent, the bidirectional absorptivity trade-off and structural optimization of the bidirectional metamaterial optical absorption device under planar light incidence are realized.

[0024] Furthermore, step S10 includes the following sub-steps:

[0025] S11: Acquire electromagnetic simulation absorption spectrum data of the bidirectional metamaterial optical absorption device;

[0026] S12: Based on the electromagnetic simulation absorption spectrum data of the acquired bidirectional metamaterial optical absorption device, unlabeled data is generated using a noise reduction diffusion probability algorithm;

[0027] S13: Construct the absorption spectrum prediction model using the mean teacher algorithm framework and train the absorption spectrum prediction model based on the electromagnetic simulation absorption spectrum data and unlabeled data.

[0028] Furthermore, step S20 includes the following sub-steps:

[0029] S21: Configure a reinforcement learning pre-environment module including state space, reward function, and action space.

[0030] S22: Construct a proxy model by writing a superior action commenting algorithm framework.

[0031] Thirdly, this invention proposes a method for fabricating a tunable bidirectional metamaterial optical absorption device, comprising the following steps:

[0032] S100: Prepare the substrate by selecting a quartz material with high transmittance in the near-infrared band as the substrate and performing ultrasonic cleaning on the substrate.

[0033] S101: Deposit a silver metal layer on a quartz substrate using physical vapor deposition methods such as magnetron sputtering or electron beam evaporation.

[0034] S102: Coating alignment layer and rubbing alignment. A liquid crystal alignment film, such as a polyimide alignment layer, is spin-coated onto the surface of the silver metal layer. After being soft-dried, it is placed at a high temperature for thermal curing to form a uniform alignment film. Then, a wiping machine is used to rub the surface of the polymer film along a predetermined direction to induce the subsequent liquid crystal molecules to align in the set direction.

[0035] S103: Liquid crystal casting and polymerization. Micron-sized spherical spacers are placed on the alignment layer to limit the thickness of the liquid crystal layer. Nematic liquid crystal material mixed with photopolymerizable monomers is injected or spin-coated through capillary action. Then, the photopolymerizable monomers are polymerized and cross-linked by irradiation with ultraviolet light, thereby fixing the orientation of the liquid crystal.

[0036] S104: Deposit a silver metal layer again. After the liquid crystal layer has cured, deposit another silver metal layer on the liquid crystal layer using physical vapor deposition.

[0037] S105: Recoat the alignment layer and rub the alignment layer, spin-coat the liquid crystal alignment film on the surface of the deposited silver metal layer, heat cure, and rub the polymer film surface with a wiping machine.

[0038] S106: Perform liquid crystal casting and polymerization again, place spacers on the aligned layer, inject or spin-coate nematic liquid crystal material containing photopolymerizable monomers and perform ultraviolet polymerization.

[0039] S107: Deposit a chromium metal layer. After the liquid crystal layer has cured, a chromium metal layer is deposited on the liquid crystal layer using a physical vapor deposition method such as magnetron sputtering.

[0040] S108: Deposit a silicon nitride layer by depositing a silicon nitride layer on a chromium metal layer using a low-pressure chemical vapor deposition method;

[0041] S109: Patterned silicon nitride layer. Appropriate photoresist is spin-coated on the silicon nitride layer, and then a ring pattern is etched by photolithography, electron beam lithography or nanoimprinting. Then, the silicon nitride is etched by reactive ion etching. After etching, the residual photoresist is stripped off to obtain a ring-shaped silicon nitride structure array.

[0042] The beneficial effects of this invention are as follows:

[0043] This invention utilizes a six-layer stacked structure—a bottom-up silver metal layer, a liquid crystal dielectric layer, another silver metal layer, another liquid crystal dielectric layer, a chromium metal layer, and a silicon nitride ring—to achieve absorption of bidirectional incident light. Specifically, it exhibits broadband absorption of top-incident forward-facing planar light and narrow-band absorption of bottom-incident reverse-facing planar light within the 0.6-1.6 μm wavelength range. Compared to unidirectional absorbers that achieve broadband or narrow-band absorption in a single direction, bidirectional light absorbers offer superior applicability due to their ability to switch between broadband and narrow-band absorption. Furthermore, by altering the deflection angle of the liquid crystal molecules in the liquid crystal dielectric layer, the absorptivity and absorption bandwidth of the device for bottom or top-incident planar light can be tuned, thus achieving tunable characteristics.

[0044] This invention generates unlabeled data from electromagnetic simulation data using a denoising diffusion probability algorithm. A mean-teacher algorithm is used to build a prediction model for the absorption spectrum of a bidirectional metamaterial optical absorber. The model is trained using both the electromagnetic simulation data and the unlabeled data. This prediction model enables rapid and accurate prediction of the absorption response of the bidirectional metamaterial optical absorber with limited data samples, replacing the modeling and simulation process of traditional electromagnetic simulation software. An AI reinforcement learning agent is constructed using a superior action commenting algorithm to explore the trade-offs in the bidirectional light absorption rate and corresponding structural optimization. This replaces manual experience-based reasoning or extensive trial and error, improving the efficiency of device structure optimization. Furthermore, by combining the bidirectional metamaterial optical absorber absorption spectrum prediction model with the AI ​​reinforcement learning agent, rapid optimization of the absorption rate trade-offs and structure of the bidirectional metamaterial optical absorber under bidirectional planar light incidence is achieved, solving the problem of inherent coupling effects between material parameters of different layers of the bidirectional absorber that mutually restrict device performance. Attached Figure Description

[0045] Figure 1 is a schematic diagram of the structure of the tunable bidirectional metamaterial optical absorption device of the present invention;

[0046] Figure 2 is a flowchart of the optimization method for the tunable bidirectional metamaterial optical absorption device of the present invention;

[0047] Figure 3 shows a comparison between the predicted absorption spectra and the actual spectra of the bidirectional metamaterial optical absorption device of the present invention under three different structures.

[0048] Figure 4 shows the upward trend of the reward function of the artificial intelligence reinforcement learning agent of the present invention for a multi-objective optimization task;

[0049] Figure 5 shows the predicted and simulated absorption spectra of the tunable bidirectional metamaterial optical absorption device of the present invention after optimization;

[0050] Figure 6 shows the simulated absorption spectrum of the bidirectional absorptivity of the tunable bidirectional metamaterial optical absorption device of the present invention as a function of liquid crystal deflection angle.

[0051] Figure 7 is a flowchart of the fabrication method of the tunable bidirectional metamaterial optical absorption device of the present invention. Detailed Implementation

[0052] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings of the embodiments of the present invention.

[0053] As shown in Figure 1, this invention discloses a tunable bidirectional metamaterial light-absorbing device. Figure 1(a) is a front view of the device structure, Figure 1(b) is a left view of the device structure, and Figure 1(c) is a top view of the device structure. For planar light incident bidirectionally from the top and bottom of the light absorber in the wavelength range of 0.6-1.6 μm, the light absorber can absorb both incident light from the top and bottom, specifically by broadband absorption of the top-incident planar light and narrowband absorption of the bottom-incident planar light. The device has a six-layer stacked structure of metal-dielectric-metal-dielectric-metal-dielectric, consisting of a silver metal layer 10, a liquid crystal dielectric layer 20, a silver metal layer 30, a liquid crystal dielectric layer 40, a chromium metal layer 50, and a silicon nitride ring 60, from bottom to top. The silver metal layer 10, the liquid crystal dielectric layer 20, and the silver metal layer 30 are used to adjust the absorption rate of the device for light incident from the bottom plane; the silver metal layer 30, the liquid crystal dielectric layer 40, the chromium metal layer 50, and the silicon nitride ring 60 are used to adjust the absorption rate of the device for light incident from the top plane.

[0054] As shown in Figure 2, this invention discloses an optimization method for a tunable bidirectional metamaterial optical absorption device, comprising the following steps:

[0055] S10: Constructing a prediction model for the absorption spectrum of bidirectional metamaterial optical absorption devices using artificial intelligence semi-supervised learning technology;

[0056] S20: Constructing reinforcement learning intelligent agents through artificial intelligence reinforcement learning technology;

[0057] S30: By combining the absorption spectrum prediction model of the bidirectional metamaterial optical absorption device with the artificial intelligence reinforcement learning agent, the bidirectional absorptivity trade-off and structural optimization of the bidirectional metamaterial optical absorption device under planar light incidence are realized.

[0058] Furthermore, step S10 includes the following sub-steps:

[0059] S11: Acquire electromagnetic simulation absorption spectrum data of the bidirectional metamaterial optical absorption device;

[0060] S12: Based on the electromagnetic simulation absorption spectrum data of the acquired bidirectional metamaterial optical absorption device, unlabeled data is generated using a noise reduction diffusion probability algorithm;

[0061] S13: Construct the absorption spectrum prediction model using the mean teacher algorithm framework and train the absorption spectrum prediction model based on the electromagnetic simulation absorption spectrum data and unlabeled data.

[0062] In step S11, acquiring the electromagnetic simulation absorption spectrum data of the bidirectional metamaterial optical absorption device includes the following specific steps:

[0063] The bidirectional metamaterial optical absorption device was modeled and simulated using electromagnetic simulation software (such as MEEP or FDTD Solutions), and the absorption response dataset was obtained. The device structure was modeled from bottom to top using simulation software, consisting of a six-layer metal-dielectric-metal-dielectric-metal-dielectric structure: a silver metal layer 10, a liquid crystal dielectric layer 20, a silver metal layer 30, a liquid crystal dielectric layer 40, a chromium metal layer 50, and a silicon nitride ring 60. The corresponding layer thicknesses are H1, H2, H3, H4, H5, and H6, respectively, and the inner and outer radii of the top silicon nitride ring are R1 and R2, respectively. The refractive indices of silver and chromium were obtained from Palik's experimental data, the refractive index of silicon nitride from Luke et al.'s experimental data, and the liquid crystal layer material from Li J's experimental data. The silver metal layer, the liquid crystal dielectric layer, the chromium metal layer, and the silicon nitride ring have the same period, with period P set to 0.6 μm. Furthermore, periodic boundary conditions were used on the x and y axes, and the z-axis direction was set as a perfectly matched layer. The incident light was set to be incident from top to bottom and from bottom to top, respectively, with the light source wavelength range set to 0.6 to 1.6 μm, and a total of 301 sampling points were set. The parameter scanning ranges were set as follows: H1 and H3 ranged from 30 to 120 nm, H2, H4, and H6 ranged from 110 to 250 nm, H5 ranged from 10 to 30 nm, R1 ranged from 0 to 100 nm, and R2 ranged from 150 to 280 nm. 2916 sets of absorption response data were simulated at liquid crystal molecule angles of 0° and 90°, respectively, resulting in 5832 sets of absorption response data. The absorptivity was calculated using the following formula:

[0064] ;

[0065] in Transmittance, Reflectance.

[0066] In step S12, generating unlabeled data using the denoising diffusion probability algorithm includes the following specific steps:

[0067] A denoising diffusion probability algorithm was written using the Python programming language. This algorithm includes a forward diffusion process and a backward diffusion process. The forward diffusion process follows a Markov chain, as shown in the following formula:

[0068] ;

[0069] ;

[0070] Where variance scheduling is The mean is The variance is .

[0071] For the reverse process, the algorithm learns to reconstruct the original data step by step from Gaussian noise by building a neural network similar to the reverse transformation, as shown in the following formula:

[0072] ;

[0073] ;

[0074] Where the standard deviation is .

[0075] During the forward diffusion process, the total number of time steps was set to 500, and the noise variance parameter range was set to 0.001 to 0.02. The noise prediction network framework during the reverse diffusion process was set as a fully connected multilayer perceptron, including five hidden layers with 256 neurons each using the ReLU activation function. The noise prediction network used the Adam optimizer with a learning rate of 0.001. Furthermore, 10,000 unlabeled data points were generated using the denoising diffusion probability algorithm, and these unlabeled data points, along with the electromagnetic simulation data from step S11, were used to train the absorption spectrum prediction model for the bidirectional metamaterial optical absorption device.

[0076] In step S13, constructing the absorption spectrum prediction model using the mean teacher algorithm framework and training the absorption spectrum prediction model based on the electromagnetic simulation absorption spectrum data and unlabeled data includes the following specific steps:

[0077] (1) Construct the absorption spectrum prediction model using the mean teacher algorithm framework.

[0078] A semi-supervised learning framework based on the mean-teacher algorithm was constructed using the Python programming language to achieve accurate prediction of the spectral response of device structures with limited electromagnetic simulation labeled data. The semi-supervised learning framework consists of a student network and a teacher network, both with identical structures and built upon a multilayer perceptron. The student network model is optimized through backpropagation, while the teacher network model is updated using an exponential moving average mechanism of the student model parameters, with the update formula as follows:

[0079] ;

[0080] in This represents the momentum coefficient, and its value is set to 0.99.

[0081] The neural network takes a 9-dimensional structural parameter vector [H1, H2, H3, H4, H5, H6, R1, R2, liquid crystal molecule deflection angle] as input and outputs a 602-dimensional absorption response prediction spectrum. The first 301 dimensions represent the broadband absorption response prediction spectrum for light incident from the top plane, and the last 301 dimensions represent the narrowband absorption response prediction spectrum for light incident from the bottom plane. The neural network employs a 5-layer hidden layer design, with 128, 256, 512, 512, and 512 neurons in each layer. It uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 1000 training epochs. The loss function of the neural network is a weighted combination of mean absolute error and mean squared error, with the mean absolute error weight set to 0.2 and the mean squared error weight set to 0.8.

[0082] For electromagnetic simulation data samples, the supervision loss between the predicted value and the true value is calculated according to the following formula:

[0083] ;

[0084] in, and These represent the actual and predicted absorbance values ​​for supervised learning, and the loss weighting coefficient, respectively. Set it to 0.6.

[0085] For unlabeled data generated by the denoising diffusion probability algorithm, the consistency loss between the student network predictions and the teacher network pseudo-labels is calculated using the following formula:

[0086] ;

[0087] in and The predicted output values ​​of the student model and the teacher model, respectively, and the loss weighting coefficients. Set it to 0.6.

[0088] The total loss function is calculated using the following formula:

[0089] ;

[0090] Among them, the consistency loss coefficient Set it to 0.1.

[0091] (2) The absorption spectrum prediction model is trained based on the electromagnetic simulation absorption spectrum data and the unlabeled data.

[0092] The 5832 labeled data points obtained through numerical simulation were divided into training and testing sets in an 8:2 ratio. These, along with the 10000 unlabeled data points generated in step S12 using the denoising diffusion probability algorithm, were used to train the absorption spectrum prediction model based on the mean-teacher framework. Figure 3 shows the comparison between the predicted absorption spectra and the actual spectral values ​​of the trained bidirectional metamaterial optical absorption device absorption spectrum prediction model under three structural conditions. Figure 3(a) shows the size parameters of the three different structures; Figure 3(b) compares the predicted and simulated results for the broadband absorption spectrum of light incident on the top plane under the three structural conditions; and Figure 3(c) compares the predicted and simulated results for the narrowband absorption spectrum of light incident on the bottom plane under the three structural conditions. The absorption spectrum prediction results shown in Figure 3 are very close to the electromagnetic simulation results, verifying that the absorption spectrum prediction model based on the mean-teacher algorithm has high prediction accuracy.

[0093] Furthermore, step S20 includes the following sub-steps:

[0094] S21: Configure a reinforcement learning pre-environment module including state space, reward function, and action space.

[0095] S22: Construct a proxy model by writing a superior action commenting algorithm framework.

[0096] In step S21, configuring the reinforcement learning pre-environment module includes the following specific steps:

[0097] (1) Constructing the state space

[0098] To obtain a bidirectional metamaterial light absorber structure that satisfies wide-band absorption when the light source is incident from the top and narrow-band absorption with a small half-width when incident from the bottom, the structural parameters H1, H2, H3, H4, H5, H6, R1, R2 of the bidirectional metamaterial light absorber and the liquid crystal molecule deflection angle of the liquid crystal dielectric layer are selected as the state space. Furthermore, to ensure that the structural parameters are physically reasonable and fabricationally feasible, clear boundaries are set for the values ​​of each state variable. Specifically, H1 and H3 are in the range of 20-130 nm, H2, H4, and H6 are in the range of 100-260 nm, H5 is in the range of 10-30 nm, R1 is in the range of 0-125 nm, R2 is in the range of 125-300 nm, and the liquid crystal molecule deflection angle is in the range of 0-90°.

[0099] (2) Constructing the reward function

[0100] The multi-objective trade-off optimization requirements for the bidirectional metamaterial optical absorber are clearly defined. Specific objectives include: broadband absorption under top-plane light incidence, narrow-band absorption under bottom-plane light incidence, and minimizing the full width at half maximum (FWHM) during narrow-band absorption. Based on these objectives, a reward function is designed. The basic reward function consists of three parts: a broadband absorptivity weighted term, a narrow-band absorption peak intensity weighted term, and a narrow-band absorption FWHM inverse weighted term, which together constitute the basic reward. Furthermore, when the candidate structure simultaneously satisfies the conditions of an average broadband absorptivity greater than 97%, a narrow-band absorption peak intensity greater than 97%, and a narrow-band absorption FWHM less than 25 nm, an important incentive term is introduced to further enhance the reward function value. In summary, the final reward function is set as follows:

[0101] ;

[0102] in This represents the broadband absorptivity of light incident from the top plane within the wavelength range of 0.6–1.6 μm. This represents the peak value of the narrow-band absorption peak for light incident from the bottom plane. The term represents the full width at half maximum (FWHM) of the narrowband absorption peak, and is a penalty term. The definition is shown in the following formula:

[0103] ;

[0104] in This represents the number of absorption response points when the broadband absorption rate is below 90% after normalization, and is a normalized value between 0 and 1.

[0105] (3) Constructing the action space

[0106] Since the absorption rate of the absorption device can be adjusted by changing the structural parameters of the metasurface bidirectional light absorber and the liquid crystal molecule deflection angle of the liquid crystal medium layer, the following actions are set as the action space: liquid crystal molecule deflection angle ±45°, H1±4nm, H2±4nm, H3±4nm, H4±4nm, H5±4nm, H6±4nm, R1±4nm, and R2±4nm, for a total of 18 actions.

[0107] In step S22, constructing the proxy model by writing a superior action review algorithm framework includes the following specific steps:

[0108] The action network, evaluation network, and advantage function were written using the Python programming language. The action network was constructed using a six-layer fully connected hidden layer structure, with the number of neurons in each hidden layer set to 128, 256, 512, 512, 256, and 128, respectively. Each hidden layer used the ReLU activation function. The output layer of the action network used the Softmax function to output an 18-dimensional probability distribution, representing 18 actions that adjust the device structure in the action space. The evaluation network had the same structure as the action network, but it was configured to output only a single scalar value. The evaluation network was used to evaluate the advantage function of the actions taken by the action network, and the advantage function was defined as follows:

[0109] ;

[0110] in, Indicates the state Take action The expected return obtained at that time Indicates the state The average return that can be obtained based on the current strategy is as follows.

[0111] Since the dominance function is difficult to obtain directly and accurately, the time difference method is used to approximate its value according to the following formula:

[0112] ;

[0113] in, The direct reward for the current action. To carry out the operation The next state after that, The discount factor, used to weigh the impact of current and future rewards, is set to 0.99.

[0114] During network training, three loss functions—policy gradient loss (based on the advantage function), value loss, and entropy regularization—are used to simultaneously optimize the performance of both the action network and the comment network. The total loss function formed by these three functions is shown in the following formula:

[0115] ;

[0116] Among them, the weighting coefficient Set to 0.5. The initial value is set to 0.2 and gradually decays with a decay rate of 0.999 during the training process to control the degree of policy exploration.

[0117] In step S30, the joint absorption spectrum prediction model and reinforcement learning agent for device structure optimization includes the following specific steps:

[0118] The trained absorption spectrum prediction model based on the mean-teacher algorithm is integrated into a reinforcement learning agent based on the dominance-action comment algorithm to guide the agent in multi-objective optimization. The reinforcement learning agent can take structural parameters as state input and execute actions in the action space to adjust geometric parameters or liquid crystal molecule deflection angles to regulate the absorptivity of the absorption device. After each action, the agent inputs the updated structural parameters into the absorption spectrum prediction model to quickly obtain the corresponding broadband and narrowband absorption spectra and calculate the reward function value. As the reward function value increases, it gradually guides the agent to find the optimal structural parameters for multi-objective optimization.

[0119] Figure 4 shows the rising trend of the reward function of the reinforcement learning agent joint absorption spectrum prediction model for multi-objective trade-off device structure optimization. During the exploration process of the reinforcement learning agent, the reward function shows a continuous upward trend, eventually stabilizing and reaching a high level. In the early exploration phase before 1400 rounds, the agent gradually learns the relationship between the structural parameters of the bidirectional metamaterial optical absorber and its absorption response. Around 1700 rounds, the reward function value jumps to around 7000, indicating that the agent has learned a better structure adjustment strategy. After 2100 rounds, the reward function value stabilizes at around 9000, indicating that the model's exploration strategy has basically converged. The results in Figure 4 verify the effectiveness of the reinforcement learning agent joint absorption spectrum prediction model for multi-objective optimization of the bidirectional metamaterial optical absorber structure design, demonstrating its ability to explore and discover structural parameters that meet design objectives.

[0120] Figure 5 shows the structural parameters and absorption spectra of the bidirectional metamaterial optical absorber after optimization using the structural optimization method. Figure 5(a) shows the structural parameters of the bidirectional metamaterial optical absorber after optimization, Figure 5(b) shows the broadband absorption spectrum of the bidirectional metamaterial optical absorber for top-incident plane light after optimization, and Figure 5(c) shows the narrowband absorption spectrum of the bidirectional metamaterial optical absorber for bottom-incident plane light after optimization. The specific optimization results are as follows: After optimization, the bidirectional metamaterial optical absorber has an average broadband absorption rate of 97.20% for top-incident plane light in the 0.6-1.6 μm spectral range. When plane light is incident from the bottom, the bidirectional metamaterial optical absorber produces a narrowband absorption peak at 0.9 μm with an absorbance of 99.3% and a full width at half maximum (FWHM) of 22.5 nm.

[0121] Figure 6 shows the effect of different liquid crystal molecule deflection angles on the light absorption characteristics of the tunable bidirectional metamaterial optical absorption device. The deflection angles of the liquid crystal molecules in liquid crystal dielectric layers 40 and 20 were individually changed using methods such as electric field modulation and optical modulation to investigate the effect of these deflection angles on the top broadband absorption and the bottom narrowband absorption. As shown in Figure 6(a), when the deflection angle of the liquid crystal molecules in liquid crystal dielectric layer 20 changes from 0° to 90°, the tunable bidirectional metamaterial optical absorption device maintains a high average absorption efficiency for top-incident planar light. Furthermore, with the change in the deflection angle of the liquid crystal molecules, the spectral profile and peak position show systematic changes. Figure 6(b) shows the narrowband absorption spectrum when planar light is incident from the bottom. It can be seen that when the deflection angle of the liquid crystal molecules in liquid crystal dielectric layer 20 changes from 0° to 90°, the resonance wavelength of the narrowband absorption peak undergoes a significant redshift, achieving a tunable wavelength range of 107 nm. The results show that the tunable bidirectional metamaterial optical absorption device exhibits good tunability in both broadband and narrowband absorption modes, confirming its tunable characteristics in spectral control.

[0122] As shown in Figure 7, this invention discloses a method for fabricating a tunable bidirectional metamaterial optical absorption device. The detailed fabrication steps are as follows:

[0123] S100: Prepare the substrate, select a quartz material with strong near-infrared transmittance as the substrate, and perform ultrasonic cleaning on the substrate to remove contaminants.

[0124] S101: Deposit a silver metal layer. On a quartz substrate, a silver metal layer is deposited as the bottom metal mirror using physical vapor deposition methods such as magnetron sputtering or electron beam evaporation.

[0125] S102: Coating alignment layer and rubbing orientation, spin coating liquid crystal alignment film such as polyimide alignment layer on the surface of silver metal layer, soft drying and then heat curing at high temperature to form uniform alignment film, after curing, use a wiping machine to rub the surface of polymer film in a predetermined direction to make polymer chains align in the rubbing direction, thereby obtaining a consistent alignment field on the metal surface to induce liquid crystal molecules to align in a set direction.

[0126] S103: Liquid crystal casting and polymerization. Micron-sized spherical spacers are placed on the alignment layer to limit the thickness of the liquid crystal layer. Nematic liquid crystal material mixed with photopolymerizable monomers is injected or spin-coated through capillary action. Then, the photopolymerizable monomers are polymerized and cross-linked by irradiation with ultraviolet light, thereby fixing the orientation of the liquid crystal.

[0127] S104: Deposit a silver metal layer again. After the liquid crystal layer has cured, deposit another silver metal layer on the liquid crystal layer using physical vapor deposition.

[0128] S105: Recoat the alignment layer and rub the alignment layer, spin-coat the liquid crystal alignment film on the surface of the deposited silver metal layer, heat cure, and rub the polymer film surface with a wiping machine.

[0129] S106: Perform liquid crystal casting and polymerization again, place spacers on the aligned layer, inject or spin-coate nematic liquid crystal material containing photopolymerizable monomers and perform ultraviolet polymerization.

[0130] S107: Deposit a chromium metal layer. After the liquid crystal layer has cured, a chromium metal layer is deposited on the liquid crystal layer using a physical vapor deposition method such as magnetron sputtering.

[0131] S108: Deposit a silicon nitride layer by depositing a silicon nitride layer on a chromium metal layer using a low-pressure chemical vapor deposition method;

[0132] S109: Patterned silicon nitride layer. Appropriate photoresist is spin-coated on the silicon nitride layer, and then a ring pattern is etched by photolithography, electron beam lithography or nanoimprinting. Then, the silicon nitride is etched by reactive ion etching. After etching, the residual photoresist is stripped off to obtain a ring-shaped silicon nitride structure array.

[0133] The embodiments described above are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A tunable bidirectional metamaterial optical absorption device, characterized in that, The structure comprises a six-layer unit stack structure consisting of a metal-dielectric-metal-dielectric-metal-dielectric layer. The unit is arranged sequentially from bottom to top as follows: a silver metal layer (10) in the first layer, a liquid crystal dielectric layer (20) in the second layer, a silver metal layer (30) in the third layer, a liquid crystal dielectric layer (40) in the fourth layer, a chromium metal layer (50) in the fifth layer, and a silicon nitride ring (60) in the sixth layer. All the silver metal layers, all the liquid crystal dielectric layers, the chromium metal layer, and the silicon nitride ring have the same period. The deflection angle of the liquid crystal molecules in the second liquid crystal dielectric layer (20) and the fourth liquid crystal dielectric layer (40) can be changed. The deflection angle of the liquid crystal molecules in the second liquid crystal dielectric layer (20) and the fourth liquid crystal dielectric layer (40) can be changed to adjust the absorption rate and absorption bandwidth range of the device for the bottom or top incident plane light, thereby realizing the tunable characteristics of the bidirectional metamaterial light absorption device.

2. The tunable bidirectional metamaterial optical absorption device according to claim 1, characterized in that, The tunable bidirectional metamaterial optical absorption device described above can achieve bidirectional optical absorption in the 0.6-1.6μm wavelength band, performing broadband absorption of the forward plane light incident from the top and narrowband absorption of the reverse plane light incident from the bottom.

3. The tunable bidirectional metamaterial optical absorption device according to claim 1, characterized in that, The first silver metal layer (10), the second liquid crystal dielectric layer (20), and the third silver metal layer (30) are used to regulate the absorption rate of the device for bottom incident light; the third silver metal layer (30), the fourth liquid crystal dielectric layer (40), the fifth chromium metal layer (50), and the sixth silicon nitride ring (60) are used to regulate the absorption rate of the device for top incident light.

4. A tunable bidirectional metamaterial optical absorption device according to claim 1, characterized in that, The first layer is a silver metal layer (10) with a thickness of H1 and a thickness range of 30-120 nm; the second layer is a liquid crystal dielectric layer (20) with a thickness of H2 and a thickness range of 110-250 nm; the third layer is a silver metal layer (30) with a thickness of H3 and a thickness range of 30-120 nm; the fourth layer is a liquid crystal dielectric layer (40) with a thickness of H4 and a thickness range of 110-250 nm; the fifth layer is a chromium metal layer (50) with a thickness of H5 and a thickness range of 10-30 nm; the sixth layer is a silicon nitride ring (60) with a thickness of H6 and a thickness range of 110-250 nm; the period P of all the silver metal layers, all the liquid crystal dielectric layers, the chromium metal layers and the silicon nitride ring is 0.6 μm; the inner radius R1 of all the silicon nitride rings (60) ranges from 0-100 nm and the outer radius R2 ranges from 150-280 nm.

5. An optimization method for a tunable bidirectional metamaterial optical absorption device as described in any one of claims 1 to 4, characterized in that, The process includes the following steps: S10: Constructing an absorption spectrum prediction model for a bidirectional metamaterial optical absorbing device using semi-supervised learning technology, including the following sub-steps: S11: Collecting electromagnetic simulation absorption spectrum data of the bidirectional metamaterial optical absorbing device; S12: Generating unlabeled data using a denoising diffusion probability algorithm based on the collected electromagnetic simulation absorption spectrum data of the bidirectional metamaterial optical absorbing device; S13: Constructing the absorption spectrum prediction model using a mean teacher algorithm framework and training the absorption spectrum prediction model based on the electromagnetic simulation absorption spectrum data and the unlabeled data; S20: Constructing a reinforcement learning agent using reinforcement learning technology, including the following sub-steps: S21: Configuring a reinforcement learning pre-environment module including a state space, reward function, and action space; S22: Constructing a proxy model by writing a superior action commenting algorithm framework; S30: Combining the bidirectional metamaterial optical absorbing device absorption spectrum prediction model with the artificial intelligence reinforcement learning agent to achieve the bidirectional absorptivity trade-off and structural optimization of the bidirectional metamaterial optical absorbing device under planar light incidence.

6. A method for fabricating a tunable bidirectional metamaterial optical absorption device as described in any one of claims 1 to 4, characterized in that, The steps include: S100: Selecting a quartz material with strong near-infrared transmittance as a substrate and ultrasonically cleaning the substrate; S101: A silver metal layer is deposited on a quartz substrate using physical vapor deposition (PVD). S102: A liquid crystal alignment film is spin-coated onto the silver metal layer surface, soft-dried, and then thermally cured at high temperature to form a uniform alignment film. The polymer film surface is rubbed along a predetermined direction using a wiping machine to induce subsequent liquid crystal molecules to align in the set direction. S103: Micron-sized spherical spacers are placed on the alignment layer to limit the liquid crystal layer thickness. Nematic liquid crystal material mixed with photopolymerizable monomers is injected or spin-coated via capillary action. The photopolymerizable monomers are then polymerized and cross-linked using ultraviolet light, thereby fixing the liquid crystal alignment. S104: After the liquid crystal layer has cured, another silver metal layer is deposited on the liquid crystal layer using physical vapor deposition (PVD). S105: Spin-coat a liquid crystal alignment film onto the deposited silver metal layer, thermally cure it, and rub the polymer film surface with a wiping machine; S106: Place spacers on the aligned layer, inject or spin-coat a nematic liquid crystal material containing photopolymerizable monomers, and perform ultraviolet polymerization; S107: After the liquid crystal layer has cured, deposit a chromium metal layer using physical vapor deposition; S108: Deposit a silicon nitride layer on the chromium metal layer using low-pressure chemical vapor deposition; S109: Spin-coat an appropriate photoresist onto the silicon nitride layer, then use photolithography, electron beam lithography, or nanoimprint lithography to depict a ring pattern, then use reactive ion etching to etch the silicon nitride, and after etching, remove the residual photoresist to obtain a ring-shaped silicon nitride structure array.

Images