A terahertz biosensor and pesticide detection method based on the HIT effect

Abstract

The present invention discloses a terahertz biosensor and a pesticide detection method based on the HIT effect. The lower flexible material of the sensor is polyimide, and the upper metasurface is composed of a number of periodically arranged unit structures. Among them, the periodically arranged unit structures are composed of N×N random units; the random units are coding matrices composed of metal or air. The implementation of the present invention not only improves the application potential of the terahertz biosensor, but also provides a new and effective way for the detection of pesticide residues.

Description

Technical Field

[0001] The present invention relates to the field of biosensing technology, and particularly relates to a terahertz biosensor based on the HIT effect and a pesticide detection method. Background Art

[0002] Terahertz (THz) waves are located between the microwave and infrared bands. Their unique electromagnetic properties have made them a research hotspot in the current scientific research and industrial circles. Terahertz waves have specific absorption characteristics for a variety of chemical substances and biological molecules. Therefore, they show great application potential in non-destructive detection, biomedical imaging, and environmental monitoring. Especially in the field of pesticide residue detection, because they can provide the "spectral fingerprint" of substances, rapid and accurate detection can be achieved.

[0003] However, although terahertz technology theoretically has many advantages, its practical application still faces a series of challenges. The design of terahertz biosensors is complex, extremely sensitive to environmental noise and sample impurities, and is easily interfered by external conditions, which limits its application effect in complex environments. In addition, traditional terahertz biosensors rely on a single resonance shift mode for detection. Quantitative detection of simple substances can be achieved through frequency shift, but it limits the detection sensitivity and affects the resolution ability for mixed substances, and there are deficiencies in qualitative discrimination detection. In addition, the design and implementation of metamaterial biosensors also face challenges because the performance of metamaterials depends on the precise configuration of their microstructures. Developing a metamaterial structure with specific electromagnetic responses requires delicate design and a long optimization process, which usually involves complex numerical simulations and experimental verifications. Each step of adjustment may consume a large amount of time and resources. The manufacturing process of metamaterials also needs to be precisely controlled to ensure that the prepared materials can accurately match the designed parameters. To sum up, the current terahertz biosensors and pesticide detection technologies mainly face problems such as complex design, lack of efficient design methods, environmental interference, and qualitative discrimination detection. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a terahertz biosensor based on the HIT effect and a pesticide detection method, which uses the vibration mode generated by pesticide molecules to couple with the plasma mode generated by the metamaterial structure, and qualitatively analyzes pesticides through the transparent window formed at the resonance frequency, solving the problem that traditional terahertz metamaterial sensors sense through the resonance frequency. In actual sample detection, a complex environmental background and potential noise or impurities will interfere with the normal response of the sensor, resulting in frequency and amplitude offsets.

[0005] Technical solution: A terahertz biosensor based on the HIT effect according to the present invention, the lower flexible material is polyimide, and the upper metasurface is composed of a plurality of periodically arranged unit structures. Among them, the periodically arranged unit structures are composed of N×N random units; the random units are coding matrices composed of metal or air, where 1 represents metal and 0 represents air.

[0006] Further, the period p of each unit is 50-150 μm

[0007] Further, the polyimide has a dielectric constant of 3.1-3.5, a loss tangent of 0.01-0.1, and a thickness of 10-50 μm.

[0008] Further, the dimensions of the random unit in length, width, and thickness are 5-15 μm, 5-15 μm, and 0.1-0.3 μm respectively.

[0009] A design method of a terahertz biosensor based on the HIT effect according to the present invention includes the following steps:

[0010] (1) Establish a dual-channel inverse design model of the metamaterial structure. Each channel of the model is composed of 4 sets of convolutional layers, pooling layers, and RESTNET modules, and then outputs through a connection layer and a fully connected layer to learn and predict the electromagnetic response of the metamaterial covered with different pesticides.

[0011] (2) Use a deep learning algorithm to train the model. The input data is the transmission spectrum of the metasurface covered with different pesticides with a thickness of 0-500 μm and the absorption coefficient of the covered pesticide, including the absorption coefficients of triadimefon, 2,4-D, fipronil, and the absorption coefficients of different ratios of triadimefon and 2,4-D. The output data is the corresponding coding matrix of the metamaterial structure.

[0012] (3) Verify and optimize the model to ensure that the designed metamaterial structure can achieve the expected detection effect in actual applications.

[0013] Further, in step (1), the convolutional layer is used to fully extract features, the pooling layer is used to reduce the number of extracted features, and the ResNet block avoids the problem of gradient disappearance by constructing a shortcut connection between the input and the output.

[0014] Further, in step (2), the LeakRelu function is used as the activation function for each layer, the Adam optimization algorithm is used to minimize the loss function, and the mean square error MSE is used as the loss function to characterize the performance of the model.

[0015] A preparation method of a terahertz biosensor based on the HIT effect according to the present invention includes the following steps:

[0016] (S1) Clean the silicon wafer. Clean the high-resistance silicon with acetone. The surface of the substrate is required to be dry. Attach a polyimide with a thickness of 10 - 50 μm to the high-resistance silicon.

[0017] (S2) Spin-coat the photoresist. Drop the photoresist at the middle position of the polyimide, and make the photoresist evenly adhere to the entire surface through the high-speed rotation of the spin coater. Spin the first layer of LOR resist at a speed of 600 / 4000 rpm for 10 / 60 seconds, bake at a temperature of 150 °C for 5 minutes. Spin the second layer of AZ1500 resist at a speed of 600 / 6000 rpm for 6 / 40 seconds, bake for 10 minutes at a temperature of 90 °C.

[0018] (S3) UV exposure and development. Align the sample with the mask plate on the lithography machine, the exposure time is 7 seconds. Immediately after exposure, develop with a positive photoresist developer for 12 seconds.

[0019] (S4) Evaporate metal and lift-off. Place the developed sample in a magnetron sputtering system to complete metal evaporation. Immerse the sample with evaporated metal in an acetone solution for 10 minutes, and use ultrasonic cleaning for metal lift-off, with an intensity of 75 Hz for 1 minute. Remove the polyimide from the high-resistance silicon to complete the preparation of the biosensor.

[0020] A method for detecting pesticides using a terahertz biosensor based on the HIT effect according to the present invention includes the following steps:

[0021] (A1) Place the pesticides triadimefon, 2,4-D, fipronil, and the mixture of triadimefon and 2,4-D in an oven to dry the moisture. All pesticides are solid powders and are made into samples by pressing tablets. Measure using a terahertz time-domain spectrometer.

[0022] (A2) Import the desired electromagnetic response and the corresponding pesticide absorption coefficient as inputs into the trained dual-channel inverse design model.

[0023] (A3) Import the dielectric constant and dielectric loss of the pesticide into CST for simulation to form a specific "fingerprint" transparent window.

[0024] (A4) Cover the pesticides triadimefon and 2,4-D on the terahertz biosensor based on the HIT effect, and change their thickness to simulate the actual change in pesticide content to form a specific "fingerprint" transparent window.

[0025] (A4) Compare the simulation results with the experimental results.

[0026] Further, in step (A1), when making the sample tablet, use a mold with a diameter of 13 mm, press at a pressure of 12 Mpa, and take it out after pressing for 30 s.

[0027] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: By constructing a dual-channel reverse design model, it can quickly and accurately predict and design metamaterial structures, significantly improving the design efficiency and accuracy of biosensors. At the same time, based on the HIT effect, highly sensitive and selective detection of pesticide residues is achieved. The implementation of the present invention not only improves the application potential of terahertz biosensors but also provides a new effective way for the detection of pesticide residues. Brief Description of the Drawings

[0028] Figure 1 It is a schematic diagram of the dual-channel reverse design model and structure of the present invention;

[0029] Figure 2 It is the training process of the prediction model of the present invention;

[0030] Figure 3 They are the absorption coefficient, dielectric constant, and dielectric loss of the pesticides measured in the present invention;

[0031] Figure 4 They are the single-resonant and double-resonant biosensors designed by the present invention;

[0032] Figure 5 It is the simulation result of the single-resonant biosensor designed by the present invention covering a single pesticide;

[0033] Figure 6 It is the simulation result of the double-resonant biosensor designed by the present invention covering mixed pesticides;

[0034] Figure 7 It is the preparation and test result of the single-resonant biosensor designed by the present invention;

[0035] Figure 8 It is the experimental result of the single-resonant biosensor designed by the present invention covering pesticides. Detailed Embodiments

[0036] The technical solution of the present invention will be further described below with reference to the drawings.

[0037] Example 1

[0038] A terahertz biosensor based on the HIT effect, with a flexible polyimide material on the lower layer and a metasurface on the upper layer, which is composed of several periodically arranged unit structures. Among them, the periodically arranged unit structures are composed of N×N random units; the random units are a coding matrix composed of metal or air, where 1 represents metal and 0 represents air.

[0039] Among them, the period p of each unit is 100μm, the dielectric constant of polyimide is 3.1, the loss tangent is 0.1, and the thickness is 50μm. The length, width, and thickness of the random units are 15μm, 15μm, and 0.3μm respectively.

[0040] Example 2 Deep Learning-based Dual-channel Inverse Design Model

[0041] This example provides a rapid design of a terahertz biosensor based on the HIT effect, which is composed of a flexible material, polyimide, and metal patches. The polyimide has a dielectric constant of 3.1, a loss tangent of 0.05, and a thickness of 30 μm. The metal is aluminum, and the metal patches are composed of 8×8 random units, with each metal unit having a size of 10 μm×10 μm×0.2 μm. The designed metasurface is composed of several periodically arranged unit structures, and the period p of each unit is 100 μm. A total of 50,000 groups of data are collected to establish an inverse design model of the metasurface structure. Each group of data includes the encoding matrix of the random patch structure ('1' represents metal, '0' represents air) and the S-parameters corresponding to different pesticides covering 0 - 500 μm.

[0042] Furthermore, on this structure basis, a dual-channel inverse design model as shown in Figure 1 is established, which consists of a convolutional layer and a linear layer. One of the inputs in the dual channel is the transmission spectrum of the metasurface covering different pesticides with a thickness of 0 - 500 μm, and the other input is the absorption coefficient covering this kind of pesticide, including the absorption coefficients of triadimefon, 2,4-D, fipronil, and the absorption coefficients of different ratios of triadimefon and 2,4-D. The output of the network is the encoding matrix of the structure. ResNet blocks are added to the inverse design network. By constructing a shortcut connection between the input and the output, the problem of gradient disappearance encountered in traditional convolutional neural networks is avoided. Each channel consists of 8 ResNet blocks, as well as convolutional layers, pooling layers, and fully connected layers. The input of the inverse design is two one-dimensional vectors, and one-dimensional convolution is used to construct the inverse network. In addition, the number of features propagated in each network layer in the dual channel is different. Different numbers of features mean that each branch of the network can learn and extract features at different levels. Since the two channels may learn different feature representations, such diversity can improve the robustness of the model to input changes and enhance the model's representation ability. The LeakRelu function is used as the activation function for each layer. The results of the dual channel are connected by a connection layer with 256 nodes, followed by a linear layer, and the last layer is the output layer. Among them, the number of features extracted by the first layer channel is 64, 128, 256, 512. The number of features extracted by the second layer is 512, 256, 128, 64.

[0043] The Adam optimization algorithm is used to minimize the loss function, and the mean square error (MSE) is used as the loss function to characterize the performance of the model. It is defined as:

[0044]

[0045] Where Yi represents the true label, Pi represents the prediction of the network, and n represents the number of samples.

[0046] During the model training process, the initial learning rate was set to 0.001, and the training batch size was set to 64. The training results after 2000 iterations are as Figure 2 shown in a. As the number of iterations increases, the MSE fluctuation decreases and tends to be stable. The mean square errors of the training set and validation set of the prediction model are 0.078 and 0.081 respectively. Figure 2 b shows the network design result. From the prediction results, it can be seen that the inverse design network can accurately predict the corresponding coding matrix according to the target spectrum, and the waveform offset is within an acceptable range. At the same time, the prediction time for each structure is 2 ms, which is 30,000 times faster than the simulation software.

[0047] Example 3 Qualitative Discrimination Detection of Single and Mixed Pesticides of Triadimefon, 2,4-D and Fipronil by Single Resonance and Multiple Resonances

[0048] The pesticides triadimefon, 2,4-D, fipronil, and the mixture of triadimefon and 2,4-D were placed in a drying oven to dry the moisture. All pesticides are solid powders and are made into samples by pressing tablets. A die with a diameter of 13 mm was used for sample preparation by pressing. After pressing for 30 s under a pressure of 12 Mpa, it was taken out and measured using a terahertz time-domain spectrometer (THz-TDS) of model TAS7500SP. The absorption coefficient as shown in Figure 3 a and the dielectric constant and dielectric loss as shown in Figure 3 b were measured and calculated.

[0049] Single-resonance and double-resonance biosensors were designed according to the characteristic absorption peak positions of the pesticides triadimefon and 2,4-D. The desired electromagnetic response and the corresponding pesticide absorption coefficient were used as inputs and imported into the trained dual-channel inverse design model. The output results of the model are as Figure 4 shown in a and 4b. It can be seen from the two figures that the desired electromagnetic response is basically consistent with the results simulated by the inverse-designed structure.

[0050] Furthermore, the dielectric constant and dielectric loss of the pesticides were imported into CST. The pesticides triadimefon and 2,4-D were covered on the single-resonance biosensor, and their thickness was changed to simulate the actual change in pesticide content. Figure 5Figures a and 5b show the changes in the transmission spectra after triadimefon and 2,4-D with different thicknesses are coated on the surfaces of two single-resonant biosensors. At the characteristic absorption peak of triadimefon (0.957 THz), the hybrid-induced transparency (HIT) effect generated after adding triadimefon produces narrowband transmission peaks at the wide transmission valleys of the metamaterial spectrum. This is because the characteristic frequency of the pesticide molecules resonantly couples with the plasmon resonance frequency of the metamaterial structure, forming a transparent window at the original resonance frequency, which can be regarded as the "fingerprint" of the detected molecules. After the pesticide 2,4-D is coated on the surface of the metamaterial, narrow transmission peaks appear at the characteristic absorption peak of 2,4-D (1.36 THz). The designed single-resonant biosensor can achieve qualitative detection of pesticides.

[0051] Furthermore, the pesticide mixture of triadimefon and 2,4-D is coated on the multi-resonant biosensor. The simulation results are as Figure 6 shown in Figure a. Narrow transmission peaks appear at both 0.957 THz and 1.36 THz, corresponding to the characteristic absorption peaks of the two pesticides at these positions. Specific detection of the mixed pesticide can be achieved by this method. Fipronil, which has no obvious absorption peak, is coated on the multi-resonant biosensor. The simulation results are as Figure 7 shown in Figure b. Since fipronil has no obvious characteristic absorption peak in the range of 0 - 2 THz, and at the same time, the resonance characteristics of the designed biosensor cannot resonantly couple with this pesticide, only resonance shift occurs. The results show that the single-resonant and double-resonant biosensors designed by reverse design achieve qualitative discrimination detection of triadimefon, 2,4-D, fipronil, and the mixed pesticide.

[0052] Example 4 Preparation, testing of the single-resonant biosensor and qualitative discrimination detection of 2,4-D and fipronil

[0053] The preparation of the metamaterial biosensor includes the following steps:

[0054] (1) Clean the silicon wafer. Clean the high-resistance silicon with acetone. The surface of the substrate is required to be dry, and polyimide with a thickness of 30 μm is pasted on the high-resistance silicon;

[0055] (2) Spin-coat photoresist. Drop the photoresist at the middle position of the polyimide, and make the photoresist evenly adhere to the entire surface through the high-speed rotation of the spin coater. Spin the first layer of LOR resist at a speed of 600 / 4000 rpm for 10 / 60 seconds, bake at a temperature of 150 °C for 5 minutes. Spin the second layer of AZ1500 resist at a speed of 600 / 6000 rpm for 6 / 40 seconds, bake for 10 minutes at a temperature of 90 °C;

[0056] (3) UV exposure and development. Align the sample with the mask plate on the lithography machine, the exposure time is 7 seconds, and immediately use a positive photoresist developer for development after exposure, with a time of 12 seconds;

[0057] (4) Metal evaporation and stripping: Place the developed sample in a magnetron sputtering system to complete metal evaporation. Immerse the sample with evaporated metal in an acetone solution for 10 minutes, and perform metal stripping using ultrasonic cleaning with an intensity of 75 Hz for 1 minute. Remove the polyimide from the high-resistance silicon to complete the preparation of the biosensor as Figure 7 shown in Fig. a.

[0058] Furthermore, use a terahertz time-domain spectroscopy system to test the prepared biosensor. The test results are as Figure 7 shown in Fig. b. The difference in transmittance between the simulation and experimental results is due to the different dielectric losses of polyimide during the actual preparation process.

[0059] Furthermore, drop 2,4-D solution on the surface of the prepared biosensor, 2 mg each time. The results are as Figure 8 shown in Fig. a. Due to the change in dielectric properties on the surface of the biosensor, the resonant frequency will also change due to different dielectric properties. At the same time, after dropping 2,4-D solution, a transparent window appears at 1.36 THz and a transmission peak appears in the original transmission valley. The frequency position of the appeared transmission peak corresponds to the characteristic absorption peak of 2,4-D. This transparent window can be regarded as the "fingerprint" feature of the analyte 2,4-D. Drop fipronil solution with no obvious absorption peak on the surface of the prepared biosensor. The results are as Figure 8 shown in Fig. b. Since fipronil has no obvious absorption peak in the range of 0 - 2 THz, the resonance frequency of the added pesticide molecules does not match the resonance frequency of the structure, and only a frequency shift appears in the transmission spectrum, which is significantly different from the detection of 2,4-D. The method can quickly design a metamaterial biosensor according to the position of the characteristic absorption peak of pesticide molecules, and based on the specific "fingerprint" transparent window that appears in the sensing detection, it realizes the qualitative discrimination of 2,4-D in the trace state of 2,4-D and fipronil, and can be further extended to the qualitative and quantitative detection research of other analytes.

Claims

1. A design method of a terahertz biosensor based on the HIT effect, characterized in that: The following steps are involved: (1) A dual-channel reverse design model of metamaterial structure is established. Each channel of the model consists of 4 groups of convolutional layers, pooling layers, and ResNet blocks. Then, the electromagnetic response of the metamaterial covered with different pesticides is learned and predicted through the output of the connection layer and the fully connected layer. The metamaterial structure is composed of a lower layer of flexible material polyimide and an upper layer of metasurface composed of several periodically arranged unit structures. The periodically arranged unit structure is composed of N×N random units. The random unit is a coding matrix composed of metal or air, where 1 represents metal and 0 represents air. (2) The model was trained using a deep learning algorithm. The input data included the transmission spectra of different pesticides covered on the metasurface with a thickness of 0-500 μm, the absorption coefficients of triadimefon, 2,4-D, fipronil, and different ratios of triadimefon and 2,4-D. The output data was the corresponding metamaterial structure encoding matrix. (3) Verify and optimize the model to ensure that the designed metamaterial structure can achieve the expected detection effect in practical applications.

2. The design method of a terahertz biosensor based on the HIT effect according to claim 1, characterized in that: In step (1), the convolution layer is used to fully extract features, the pooling layer is used to reduce the number of extracted features, and the ResNet block avoids the gradient vanishing problem by building shortcut connections between input and output.

3. The design method of a terahertz biosensor based on the HIT effect according to claim 1, characterized in that: In step (2), the LeakRelu function is used as the activation function of each layer, the Adam optimization algorithm is used to minimize the loss function, and the mean square error MSE is used as the loss function to characterize the performance of the model.

Images