Design method for highly sensitive biochiral molecule detection devices based on reinforcement learning

Abstract

The present invention discloses a design method for a highly sensitive biochiral molecule detection device based on reinforcement learning. The detection device processes a metal nanostructure at the bottom of a microfluidic channel, and realizes the sensing of biochiral molecules by characterizing the coupling signal between the metal nanostructure and the biochiral molecule. In order to obtain a metal nanostructure that matches the target chiral molecule, the present invention uses reinforcement learning technology to search the parameter space of the geometric configuration of the metal nanostructure to obtain a metal nanostructure with high chirality. This design method fully realizes the intelligent optimization of nanophotonic devices, and has the advantages of high degree of freedom, strong robustness, wide range of applications, and fast calculation speed, and has important guiding significance for the design of other multifunctional optical devices. The biochiral molecule detection device based on a microfluidic chip designed by the present invention has multiple advantages such as real-time monitoring, reusability, and high sensitivity, and has extremely high use value.

Description

Technical Field

[0001] The present invention belongs to the technical field of biochiral molecule detection, and relates to a method for sensing and detecting biochiral molecules using nanophotonic devices. Specifically, a special nanophotonic device is designed based on the emerging artificial intelligence technology of reinforcement learning to enhance the coupling between biomolecules and metal nanostructures, thereby achieving convenient and highly sensitive sensing of biochiral molecules, especially with obvious advantages in the discrimination of enantiomers. Background Art

[0002] Chirality is one of the most fundamental properties of natural objects, describing the property that an object and its mirror image cannot overlap. It is ubiquitous and plays a vital role in numerous fields. In particular, in medical biochemistry, molecules with different chirality within the same component often exhibit distinct pharmacological properties and serve different physiological processes, yet distinguishing between these enantiomers is challenging. Consequently, the detection and identification of optical chiral signals in biomolecules has garnered significant attention. However, the chiral optical response of most biomolecules in nature is very weak, often in the ultraviolet (UV) region, significantly complicating practical applications.

[0003] With the rise of micro- and nanophotonics, surface plasmons have been discovered to have the ability to amplify optical chiral signals. Collective oscillations of electrons on the surface of metal nanostructures can generate extremely strong local electric fields. Superchiral fields can appear on the surface of chiral metal nanostructures, leading to significant chiral signals in the far-field spectrum. When metal nanostructures couple with chiral biomolecules, the chiral spectrum of the system undergoes a significant shift. Compared to the chiral signal of the biomolecule itself, this shift is not only more pronounced but can be controlled within the visible and near-infrared regions, significantly improving the sensitivity of molecular chirality detection while reducing detection costs. Therefore, surface plasmon-based chiral detection has long been considered a powerful tool for future biomolecule detection. The key to this detection approach lies in finding the right metal nanostructure for the target biomolecule to maximize the coupling signal. However, surface plasmon resonance is highly sensitive to the shape, size, and environment of the metal nanostructure, resulting in strong tunability. Its coupling with biomolecules is a complex physical model, which greatly complicates the design of metal nanostructures.

[0004] In recent years, the rapid development of artificial intelligence has spawned a number of remarkable self-learning algorithms, including reinforcement learning. Reinforcement learning algorithms learn through trial and error, adjusting their decision-making patterns based on the results of trying different decision-making options, gradually arriving at an optimal strategy. Unlike conventional supervised learning, reinforcement learning uses the results obtained from interacting with the environment to guide behavioral changes, aiming for optimal results, similar to the process by which an organism continuously adapts to its environment. With advances in artificial neural network design and the growth of computer hardware computing power, reinforcement learning has made significant contributions to numerous fields, including autonomous driving, financial trading, natural language processing, robotic automation, and healthcare, ushering in a new generation of technological innovation. Its excellent performance in solving complex process problems has made it a key area of machine learning.

[0005] Recent scientific research has demonstrated the remarkable success of artificial intelligence algorithms in micro- and nanophotonics, with significant potential in the design of photonic crystals, metamaterials, and integrated silicon photonic devices. These algorithms can enable both forward prediction and reverse design based on design parameters (such as geometry, materials, topology, and spatial arrangement) and target optical properties (such as polarization, phase, wavelength, and orbital angular momentum). Summary of the Invention

[0006] The present invention aims to provide a method for designing and fabricating a biochiral molecule detection device. This device utilizes surface plasmons to amplify the chiral signal of biomolecules, enabling highly sensitive and convenient detection of specific biomolecules. Reinforcement learning played a crucial role in the design of this device.

[0007] The fundamental principle of this biodetection method is to exploit the coupling effect between metal surface plasmons and biochiral molecules. This coupling causes changes in the far-field chiral spectrum of the metal nanostructure before and after the addition of the biomolecule. By capturing these spectral changes, the biochiral molecules can be sensed. This method applies reinforcement learning to the design of metal nanostructures, leveraging the unique advantages of reinforcement learning algorithms to solve the complex problem of biochiral molecule detection, resulting in a highly efficient and convenient method for biomolecular chirality detection.

[0008] The biochiral molecule detection device of the present invention (see Figure 1 and Figure 2) includes a microfluidic chip and an optical detector, wherein a special metal nanostructure is processed and manufactured at the bottom of the microfluidic channel of the microfluidic chip. The structure of the microfluidic chip is substrate, metal nanostructure array, microfluidic structure from bottom to top. The optical detector is used to collect the optical chirality signal of the biomolecule solution flowing through the microfluidic chip. The substrate material can be a pure dielectric system composed of SiO2, Si, Si3N4, GaAs, etc., or a mixed multilayer system composed of metal and dielectric. The metal nanostructure can be made of common metals such as gold, silver, and aluminum, and the material of the microfluidic structure is usually a high molecular polymer, such as polydimethylsiloxane (PDMS). The optical detector needs to have the ability to accurately observe the metal nanostructure array (hundred-micron pattern) in the microfluidic channel and measure its circular dichroism spectrum. A microscope equipped with a polarizer and a quarter glass slide is usually used as an optical detector. The biomolecule solution is introduced into the microfluidic chip by syringe injection or other drainage methods, so as to fully contact and couple with the metal nanostructure. A microfluidic chip has two sets of flow channels, which can be used for comparative experiments. Each set of flow channels has two inlets and one outlet. The two inlets are used to introduce distilled water and the solution to be analyzed, respectively. The introduction of distilled water can clean the microfluidic channel, enabling rapid multiple measurements of different solutions. Figure 3 .

[0009] The device is used as follows: First, distilled water is passed through the microfluidic chip, and the circular dichroism spectrum of the metal nanostructure array is measured using a microscope. Then, the solution to be analyzed is passed through the microfluidic chip, and the circular dichroism spectrum of the metal nanostructure array is also measured. By comparing the two circular dichroism spectra, the chirality of the biomolecules in the solution can be determined based on the difference between them. The circular dichroism spectra are calculated by measuring the light intensity reflectance of the structure when incident with left-handed and right-handed circularly polarized light.

[0010] The detection effect of the present invention is reflected in the fact that the circular dichroism spectra obtained by the two measurements mentioned above can show obvious peak shifts. This spectral change is due to the coupling between the chiral molecules and the metal nanostructures, rather than the change in the refractive index of the solution (see Figure 7 A simple qualitative theoretical explanation for this phenomenon is that different chiral solutions have different effective refractive indices for left-handed and right-handed circularly polarized light, resulting in distinct near-field electromagnetic modes. Enantiomers of opposite chirality induce opposite effective refractive index changes, resulting in different frequency shifts in the chiral resonance peaks of metallic nanostructures. The direction of this chiral peak shift can then be used to determine the chirality of the biomolecule being tested.

[0011] The key to this invention lies in designing metal nanostructures specifically for target biomolecules. Any metal nanostructure represents an extremely large parameter space, while those with superior chirality only account for a tiny fraction. The core concept of the algorithm employed in this invention leverages the advantages of reinforcement learning, using a neural network model to guide exploration of this parameter space while continuously updating the parameters of the neural network model.

[0012] Here we provide a complete set of metal nanostructure design process based on reinforcement learning (see Figure 4 ):

[0013] 1) Randomly generate metal nanostructures and calculate the light intensity reflectivity of the structure when left-handed (or right-handed) circularly polarized light is incident (hereinafter referred to as left-handed (right-handed) reflectivity) through electromagnetic field simulation as the initial data set;

[0014] 2) Using the data in the dataset, we train convolutional neural networks with different structures, enabling them to acquire the ability to preliminarily predict the left-handed (right-handed) reflectivity of metal nanostructures based on their structural parameters;

[0015] 3) Using Bayesian optimization to automatically sample the optical chirality of metal nanostructures, calculate their circular dichroism with the left-handed (right-handed) reflectivity predicted by the neural network, and repeatedly optimize to obtain a metal nanostructure that may have been optimized;

[0016] 4) Compare the predictions given by different neural networks to test metal nanostructures that may have been optimized: If different neural networks give the same prediction, it means that the neural network's prediction is correct and the metal nanostructure has been optimized. If multiple neural networks give completely different predictions, it means that the metal nanostructure is a new structure that the neural network has not learned and needs to be added to the dataset.

[0017] 5) Perform electromagnetic field simulation calculations on the new structure in step 4) to obtain its left-hand (right-hand) reflectivity and add them to the data set;

[0018] 6) Repeating steps 2) to 5) to achieve intelligent automatic exploration of the geometric parameter space of the metal nanostructure, and outputting a metal nanostructure with excellent chirality.

[0019] The left-handed (right-handed) reflectivity data of the metal nanostructures in steps 1) and 5) are obtained by numerical simulation methods such as finite difference time domain (FDTD).

[0020] The target optical parameter calculated by the electromagnetic field simulation in steps 1) and 5) above can be the light intensity reflectivity of the metal nanostructure when incident on left-handed circularly polarized light, or the light intensity reflectivity of the metal nanostructure when incident on right-handed circularly polarized light. Both methods are acceptable and do not affect subsequent design, but only one method needs to be calculated.

[0021] The above step 2) trains the neural network based on the data in the data set by passing the structural parameters of the input metal nanostructure to the neural network and adjusting the weights in the neural network nodes using the gradient descent method so that the output of the neural network approaches the corresponding optical response (left-handed or right-handed reflectivity).

[0022] The convolutional neural network described in step 2) above is mainly composed of convolutional layers, pooling layers, activation layers, and fully connected layers. It realizes the extraction of global information from local features by stacking multiple layers of small convolution kernels (see Figure 5 The front end of the convolutional neural network is composed of a stack of convolutional layers, each of which contains multiple channels, and each channel contains a data matrix. The input layer of the neural network has only one channel, which represents the matrix of the geometric configuration of the metal nanostructure (see Figure 1 As the network progresses, the size of the convolutional layer matrix decreases while the number of channels increases. Each channel extracts certain characteristic information, and in conjunction with the fully connected layers at the back end of the network, it gradually transforms the two-dimensional image into a one-dimensional vector representing a certain characteristic. This output, representing the left-handed (right-handed) reflectivity, serves as the final network output. Pooling and activation layers are interspersed between the convolutional and fully connected layers to increase the network's nonlinearity and enhance fitting capabilities.

[0023] The method for calculating circular dichroism in step 3) above primarily utilizes specular reflection transformation. When calculating the circular dichroism of a metal nanostructure, a neural network is first used to predict its left-handed (right-handed) reflectivity. Then, the left-handed (right-handed) reflectivity of the metal nanostructure after the specular reflection transformation is predicted. Due to the symmetry of the system, the latter is the right-handed (left-handed) reflectivity of the original metal nanostructure. The circular dichroism spectrum can then be calculated based on these two reflectivities.

[0024] The Bayesian optimization described in step 3) above is an optimization algorithm that finds the value of the optimized objective function by establishing a probability model based on the past calculation results of the objective function. The biggest difference between the Bayesian optimization method and random search or grid enumeration is that it refers to the previous evaluation results before trying the next set of parameters, so it can save a lot of computing resources. Bayesian optimization sets the structural parameters of the metal nanostructure as independent variables and the optical chirality of the metal nanostructure as the dependent variable. It uses the Gaussian model to analyze the mapping relationship constructed by the neural network, evaluates the objective function corresponding to the new structural parameters, and uses the structural parameter with the largest objective function as the structural parameter information that should be explored in the next round of sampling. Repeated Bayesian optimization is performed multiple times in step 3).

[0025] The optimization algorithm in step 3) above is used to optimize the metal nanostructure to obtain a circular dichroism spectrum with a sharp peak shape. Therefore, the objective function needs to reflect the peak shape characteristics of the circular dichroism spectrum. The detection effect of the present invention is reflected in that the introduction of biomolecules will cause the peak of the circular dichroism spectrum of the metal nanostructure to shift. By characterizing this peak shape shift, the detection of biochiral molecules can be completed. (See Figure 7 ) In order to improve the detection sensitivity, the resonance peak of the circular dichroism spectrum of the metal nanostructure needs to have a very narrow half-maximum width and a high peak value. Therefore, the objective function of the optimization algorithm comprehensively considers the half-maximum width (FWHM) of the resonance peak and the maximum value of the circular dichroism (CD max ). This objective function is negatively correlated with FWHM and CD max Positively correlated functions, such as the linear superposition of the two.

[0026] In step 4), the results of different neural networks are compared to verify the metal nanostructures that may have been optimized. The principle is as follows: under a limited data set, the prediction results of the neural network are not necessarily accurate. For completely new targets with no similar structures in the data set, the predictions of the neural network will be highly random. However, for structures with a considerable number of similar configurations in the data set, the neural network can often give accurate predictions. Therefore, the accuracy of the neural network prediction can be determined by comparing the prediction results of different neural networks. If the metal nanostructure is a new configuration, the probability of multiple neural networks giving the same result is extremely low. On the contrary, if the neural network has completed learning the metal nanostructure, multiple neural networks will give consistent and accurate prediction results.

[0027] The reinforcement learning system proposed in this paper is the first nanophotonic device design system built using reinforcement learning algorithms. It exhibits strong robustness and significantly reduces computational resources. It not only enables circular dichroism spectroscopy analysis but can also be extended to other optical parameters. In this paper, highly chiral metallic nanostructures with sharp resonance peaks in various geometric configurations are designed for chirality detection in biomolecules.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] Unlike traditional optimization algorithms, the reinforcement learning system of the present invention uses neural networks to learn the design of metal nanostructures and their corresponding optical properties, avoiding the shortcomings of traditional optimization algorithms in their insufficient data fitting ability and providing new ideas for the analysis and design of nanophotonic devices.

[0030] Compared to traditional supervised machine learning, this algorithm reduces a large number of unnecessary simulation calculations. In classic supervised learning, a large number of electromagnetic simulation calculations are first required to obtain the optical responses of various metal nanostructure configurations, and then the optimal metal nanostructures are found based on the neural network's predictions. However, accurate predictions by the neural network require training data for all configurations, which includes a large number of weakly chiral structures, wasting a lot of computing resources and time. The reinforcement learning model selects parameter exploration simultaneously with model training. After several rounds of exploration, the exploration range can be basically locked to the range of high-chirality structures, thereby significantly reducing the number of simulation calculations.

[0031] In the present invention, the neural network only plays an auxiliary role. The joint operation of multiple neural networks can greatly reduce the structural design deviation caused by the prediction error of the neural network.

[0032] Compared with traditional biochemical methods, the biomolecule detection method of the present invention is extremely convenient, and can achieve chirality discrimination without the need for chemical reactions. It is non-destructive to the test samples and can perform real-time monitoring. Only a small amount of sample is required to complete the detection. It has excellent promotional value and has the potential to achieve highly sensitive chiral detection of various biological macromolecules. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The figure shows the metal nanostructure in the microfluidic chip described in an embodiment of the present invention and its 01 coding method. In the figure, 1 is the silicon substrate, 2 is the silicon dioxide spacer layer, 3 is the metal nanostructure, and 5 is the solution flowing through the microfluidic channel.

[0034] Figure 2Schematic diagram of the structure of the highly sensitive biochiral molecule detection device according to an embodiment of the present invention, wherein: 1 is a silicon substrate, 2 is a silicon dioxide spacer layer, 3 is a metal nanostructure, 4 is a microfluidic material, and 6 is a microscope objective lens.

[0035] Figure 3 These are actual photos of the biochiral molecule detection device prepared in an embodiment of the present invention, wherein the upper photo includes a microfluidic chip, a catheter, and an infusion device, and the lower photo is the microfluidic chip.

[0036] Figure 4 Schematic diagram of the cyclic process of the reinforcement learning system of the present invention.

[0037] Figure 5 This is an architectural diagram of the convolutional neural network described in the present invention.

[0038] Figure 6 Scanning electron microscope characterization images of three metal nanostructures designed using a reinforcement learning system in a specific embodiment of the present invention.

[0039] Figure 7 For the use in the specific implementation of the present invention Figure 6 The chirality detection effects of the three metal nanostructures shown.

[0040] Figure 8 In Example 1 of the present invention, Figure 6 A comparison of experimental results and simulated data for circular dichroism spectra of three metal nanostructures obtained by the reinforcement learning system shown. DETAILED DESCRIPTION

[0041] The present invention will be described in further detail below through specific embodiments in conjunction with the accompanying drawings so that those skilled in the art can understand the present invention more clearly.

[0042] See also Figure 1The biochiral molecule detection device of the present invention adopts the method of a microfluidic chip, and a special metal nanostructure array is processed and manufactured at the bottom of the microfluidic channel. From bottom to top, there is a silicon substrate 1, a silicon dioxide spacer layer 2 (with a thickness of 100nm), and a metal nanostructure 3. The solution 5 flows through the microfluidic channel. The metal nanostructure used in this embodiment is a metasurface composed of multiple gold cuboids with a height of 30nm and a hundred nanometers. The side length of the cuboid is kept as an integer multiple of 10nm. Such a structure is easy to parameterize and convenient to process. This embodiment adopts a periodic arrangement method, and each structural unit is 400nm×400nm in size. Each period is divided into 40×40 units with a unit of 10nm×10nm, and each unit is encoded with 01. "1" represents the presence of a metal nanostructure at that position, and "0" represents the absence of a metal nanostructure at that position. This "01" binary matrix representing the geometric configuration of the metal nanostructure is the input of the neural network.

[0043] Optical circular dichroism is defined as follows:

[0044] CD=(R LCP –R RCP ) / (R LCP +R RCP )

[0045] In the above formula, CD represents optical circular dichroism, R LCP Represents the reflectivity under left-handed light incidence, R RCP represents the reflectivity under right-handed incident light. Therefore, the circular dichroism spectrum measurement method in the present invention is to first apply left-handed circularly polarized light to a standard silver mirror to obtain the incident light spectrum; then, apply left-handed circularly polarized light to the microfluidic chip to measure the reflected light spectrum, and then divide the two to obtain the left-handed reflectance spectrum. A similar method can be used to obtain the right-handed reflectance spectrum, and the circular dichroism spectrum can be calculated using the above formula.

[0046] In the present invention, the output of the neural network is the reflectance spectrum of the metal nanostructure under left-handed circularly polarized light. The reason for choosing to let the neural network learn to calculate the reflectance spectrum of left-handed circularly polarized light is to avoid the problem of being unable to optimize the structure. The neural network can be understood as a fitter with extremely strong fitting capabilities. Its innovation ability is very weak. When it encounters a new structure, it can only analyze this structure based on previous data. Therefore, its output result will inevitably have a high similarity with the previous data output, which brings great difficulties to the optimization task. Assuming that the neural network directly fits the circular dichroism spectrum, it will encounter difficulties in the subsequent process of optimizing the chiral signal. The neural network only learns the circular dichroism of some structures. When the neural network encounters a new structure with a particularly strong chiral signal, even if the circular dichroism value of this structure is greater than all existing structures, the neural network generally will not produce a result that exceeds the maximum circular dichroism in the existing data. Similarly, if the neural network produces an output that exceeds the maximum circular dichroism of all existing structures, this result is also questionable, because giving such an output means that it has exceeded the range that the neural network can theoretically fit.

[0047] However, if the learning objective is changed to the reflectance spectrum of left-handed circularly polarized light, the reflectance spectra of left-handed and right-handed circularly polarized light of a strongly chiral structure may be mediocre. A considerable number of similar spectra can be found in the dataset, but when they belong to the same structure, it indicates that the structure has strong chirality. In this way, the optimization algorithm combined with the neural network truly has the ability to optimize the optical chirality of metal nanostructures.

[0048] When predicting the reflectance spectrum for right-handed circularly polarized light, we can perform a specular transformation on the system. This transformation is then fed into a neural network, which then outputs the reflectance spectrum for right-handed circularly polarized light. Once the spectra for left-handed and right-handed circularly polarized light are determined, the circular dichroism spectrum can be calculated.

[0049] In the convolutional neural network of this embodiment, all convolution kernels are of size 3×3, and the pooling layers all use 2×2 maximum pooling. This network architecture of stacking small convolution kernels with maximum pooling can significantly enhance the nonlinearity of the network, achieve feature invariance, improve the network's fitting ability, and prevent overfitting to a certain extent.

[0050] Except for the last layer, all activation functions adopt the relu function. Its main advantage is that when using the optimization algorithm based on gradient calculation for parameter learning, the gradient of the relu function is very simple and clear, which can achieve faster convergence. Compared with other activation functions, the relu function solves the problem of slow convergence or even non-convergence due to the disappearance of the gradient. At the same time, its linear characteristics make neural network operations simpler and faster. Practical experience in computer science shows that the relu function has better performance as an activation function in processing image problems. The activation function of the last layer is the sigmoid function. This is because the output of the neural network in the present invention is the reflectivity when left-handed circularly polarized light is incident, which is a value between 0 and 1, and the sigmoid function is just a function with a value range of [0,1].

[0051] In addition to the above-mentioned network main framework, the present invention uses batch normalization technology (BatchNormalization), the core of which is to normalize the data in the middle layer of the network and introduce reconstruction parameters for "restoration" transformation. The main function is to improve the convergence efficiency of the model and alleviate the overfitting problem to a certain extent. For deep neural networks, shallow parameter updates will cause changes in the input data distribution of subsequent network layers. Since it involves the superposition of many layers, the input data distribution of the network layer at the end of the neural network will change very drastically, which forces the end network layer to constantly relearn to adapt to the parameter changes of the shallow network, resulting in a serious decrease in the convergence efficiency of the model. The normalization of batch normalization technology can make the input of each layer in the entire network structure have a very similar distribution, so that the change in the input distribution of the end network layer caused by the update of shallow parameters is very small, thereby effectively training the parameters of the deep network and achieving rapid convergence.

[0052] The following describes the workflow of the Bayesian algorithm. Assume that the objective function to be optimized is f(x), where x represents a parameter vector. This function is very complex, and the Bayesian algorithm needs to optimize the value of f(x) as much as possible within limited computing resources. First, several parameters x are randomly selected, the corresponding f(x) is calculated, and an initial dataset is constructed. A Gaussian distribution is fitted to this dataset to determine its mean and variance. Following this Gaussian distribution, the optimal value of f(x) and the corresponding parameter x are found. This set of parameters is substituted into the objective function, the corresponding f(x) is calculated, and the dataset is updated. The Gaussian distribution is then again fitted to the new dataset, and this cycle repeats. As the dataset is continuously updated, the value of f(x) is gradually optimized.

[0053] The objective function of the optimization algorithm in this invention describes the spectral characteristics of metal nanostructures that may have sensitive detection capabilities. Studies have shown that the circular dichroism spectra of such metal nanostructures exhibit high peaks and sharp peaks. Therefore, the objective function of the optimization algorithm needs to balance the resonance peak half-width (FHWM) and the circular dichroism maximum (CD max )The relationship between the two. max represents the height of the circular dichroism spectrum peak. If CD max If it is less than 0.35, the detection results of the corresponding device are significantly affected by the processing error and measurement error. max As the value of , the device's tolerance to experimental errors will be significantly enhanced. However, perfect chirality (CD max Approaching 1) has no special significance for the detection accuracy of biomolecules, so there is no need to pursue a particularly high circular dichroism. FHWM is one of the decisive factors for detection sensitivity. The smaller the FHWM, the more obvious the shift of the peak position. In summary, the objective function should be set as a function of FHWM and CD. max is a binary function of the independent variable. max When CD is relatively large, the structure has met the characteristics of strong chirality. At this time, the objective function is mainly affected by FHWM, guiding the optimization algorithm to find a structure with a sharp peak on this basis. max When the relative smallness is small, the objective function should follow CD max The smaller the value of , the smaller the value of , which makes the optimization algorithm avoid the weak chiral structure. For example:

[0054]

[0055] The unit of FHWM in the above formula is nanometers. When the optimization algorithm, guided by the above objective function, obtains metallic nanostructures with maximum function values, these structures typically have sufficiently large optical circular dichroism and very sharp circular dichroism spectral peaks.

[0056] In the present invention, the frequency shift amplitudes brought about by a pair of enantiomers at the same concentration may not be equal. Figure 7 , a group of left-handed glucose and right-handed glucose showed unequal peak shifts. When the biochiral molecule is located at the hot spot, the local field enhancement effect of the metal nanostructure causes the circular dichroism of the molecule in the ultraviolet band to extend to the visible light and near-infrared bands. The left-handed circularly polarized and right-handed circularly polarized modes of the high-chirality metal nanostructure have different resonance frequencies, so the enhanced chiral signals from the enantiomers appear at different wavelengths. The change in the circular dichroism spectrum depends on the near-field electromagnetic mode of the metal nanostructure. When the left-handed circularly polarized and right-handed circularly polarized electromagnetic modes show asymmetric deviation from the resonance frequency of the circular dichroism peak, the peak position shift caused by the biochiral molecule may be completely different.

[0057] The following further provides a method for preparing the microfluidic chip designed based on the reinforcement learning system, which includes the following steps:

[0058] Step 1: Use plasma enhanced chemical vapor deposition (PECVD) to deposit silicon dioxide with a thickness of 100 nm on the silicon substrate to obtain a SiO2 / Si substrate.

[0059] Step 2: Ultrasonic cleaning of the SiO2 / Si substrate was performed in the order of acetone (cleaning time 5-10 min) → ethanol (cleaning time 5-10 min) → deionized water (cleaning time 5-10 min). Finally, the deionized water remaining on the substrate was blown dry with a nitrogen gun to obtain a clean SiO2 / Si substrate.

[0060] Step three, spin-coat MMA glue and PMMA A2 glue (3000rad / s, 40s) on the upper surface of the SiO2 / Si substrate in turn, and dry at 180°C for 5min. Then use the electron beam exposure (EBL) system to etch out the designed metal nanostructure shape, and place it in the developer (MIBK) for development (about 50s). After development, immediately place it in an isopropanol solution for fixing (about 5min), then take out the sample and use a nitrogen gun to blow dry the residual isopropanol solution. Next, use electron beam evaporation to deposit titanium with a thickness of 2nm and gold with a thickness of 30nm. Titanium mainly plays an adhesive role to ensure that the gold structure is firmly fixed on the silicon dioxide surface. Finally, the entire sample is placed in an acetone solution for about 5 hours, and the metal nanostructure array is finally peeled off by rinsing with acetone solution, as shown in FIG. Figure 6 shown.

[0061] Step 4: Use PDMS material to make a microfluidic structure, align it with the prepared metal nanostructure under a microscope, and fix the microfluidic structure.

[0062] Step 5: Install the catheter and infusion device to complete the entire microfluidic chip (see Figure 3 ).

[0063] Example 1

[0064] The method of the present invention enables the discrimination and detection of L-glucose and D-glucose. L-glucose and D-glucose are a pair of enantiomers, possessing opposite chirality but essentially no optical chirality in the visible light range. Their original chiral spectra are primarily limited to the ultraviolet band, making direct detection inconvenient.

[0065] First, the reinforcement learning design system described above was used to design metal nanostructures, and a variety of high-chirality structures were obtained. They have different geometric configurations and different resonance wavelengths, which are suitable for the recognition of different chiral molecules. The numerical simulation calculations in the design process used the time-domain finite difference (FDTD) technology. Using electron beam exposure technology on a SiO2 / Si substrate, multiple 100μm×100μm metal nanostructure arrays were made with the obtained high-chirality structure as the basic unit. Then, the coupling effect of different metal nanostructures with glucose molecules was tested, specifically by observing the peak changes in the circular dichroism spectrum of the metal nanostructure array before and after the introduction of glucose molecules. After experimental screening, three structures were finally selected that showed the ability to distinguish L-glucose and D-glucose. The scanning electron microscope characterization images of these three structures are shown in Figure 6 .

[0066] By comparing the experimental results with the numerical simulation results (see Figure 8 ), we can see that the theoretical spectrum and the experimental test results are in good agreement, the peak positions are basically consistent, and there is an obvious peak shape, which shows that this theoretical design system based on reinforcement learning has a good design effect.

[0067] The experiment selected 0.5mL L-glucose, D-glucose and sodium chloride solutions and measured their circular dichroism spectra. After the L-glucose and D-glucose solutions were introduced, the experiment observed a significant shift in the circular dichroism spectrum peak position (see Figure 7 In this experiment, sodium chloride solution served as a control to eliminate peak shifts caused by changes in refractive index due to changes in solution concentration. However, the introduction of L-glucose and D-glucose caused peak shifts in opposite directions, which is consistent with theoretical expectations.

[0068] Compared with the traditional chemical method for distinguishing L-glucose from D-glucose, the present invention requires a very small amount of sample, does not require chemical reactions, and will not destroy the sample. The same instrument can perform repeated measurements, and the maintenance cost and measurement cost are extremely low. Moreover, the detection method of the present invention can be monitored in real time. It is only necessary to continuously pass the solution to be tested to monitor the spectral changes to capture the changes in the solution composition. This feature allows the present invention to be used for real-time detection of the concentration of biological molecules in patients, as well as monitoring of processing processes in factory production processes and other aspects. The present invention is not limited to the discrimination of enantiomers of glucose, but can also be used for the identification of other biological chiral molecules. This method can be extended to the identification of different biological molecules.

[0069] This invention utilizes reinforcement learning for the first time to achieve the design of high-chirality nanostructures. Compared to traditional micro-nanophotonic device design methods, the reinforcement learning system of this invention efficiently explores the parameter space, combining the strong fitting capabilities of neural networks under large data sets with the scientific and efficient data sampling of Bayesian optimization. It avoids the drawback of classical supervised learning, which requires excessive amounts of training data, significantly improves design efficiency, and achieves highly robust results. The entire design process is fully intelligent, providing valuable guidance and participation for the future design of micro-nanophotonic devices.

[0070] Finally, it should be noted that the purpose of disclosing the embodiments is to facilitate a further understanding of the present invention. Those skilled in the art should understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments; the scope of protection claimed by the present invention shall be determined by the scope defined in the claims.

Claims

1. A method for designing a biochiral molecule detection device, wherein the biochiral molecule detection device comprises a microfluidic chip and an optical detector, wherein the optical detector is used to collect optical chiral signals of a biomolecule solution flowing through the microfluidic chip; characterized in that: The microfluidic chip has a metal nanostructure processed at the bottom of the microfluidic channel. The required metal nanostructure is designed according to the following steps: 1) Randomly generate metal nanostructures and calculate the light intensity reflectivity of the metal nanostructures when left-handed or right-handed circularly polarized light is incident through electromagnetic field simulation, referred to as left-handed or right-handed reflectivity, as the initial data set; 2) Using the data in the dataset, we train convolutional neural networks with different structures, enabling them to initially predict the left-handed or right-handed reflectivity of metal nanostructures based on their structural parameters; 3) Using Bayesian optimization to automatically sample the optical chirality of metal nanostructures, calculate their circular dichroism with the left-handed or right-handed reflectivity predicted by the neural network, and repeatedly optimize to obtain a potentially optimized metal nanostructure; 4) Compare the predictions given by different neural networks to test the structures that may have been optimized: If different neural networks give the same prediction, it means that the neural network's prediction is correct and the metal nanostructure has been optimized. If multiple neural networks give completely different predictions, it means that the metal nanostructure is a new structure that the neural network has not learned and needs to be added to the dataset. 5) Perform electromagnetic field simulation calculations on the new structure in step 4) to obtain its left-handed or right-handed reflectivity and add them to the data set; 6) Repeat steps 2) to 5) to realize intelligent automatic exploration of the geometric parameter space of the metal nanostructure and output the metal nanostructure with excellent chirality.

2. The design method according to claim 1, wherein: In steps 1) and 5), the left-handed or right-handed reflectivity of the metal nanostructure is obtained by a time-domain finite difference numerical simulation method.

3. The design method according to claim 1, wherein: In step 2), the training of the neural network is to pass the structural parameters of the input metal nanostructure to the neural network, and use the gradient descent method to adjust the weights in the neural network nodes so that the output of the neural network approaches the corresponding left-handed or right-handed reflectivity.

4. The design method according to claim 1, wherein: Step 2) The neural network is a convolutional neural network composed of convolutional layers, pooling layers, activation layers and fully connected layers, which realizes the extraction of global information from local features by stacking multiple layers of small convolution kernels; the front end of the convolutional neural network is stacked by convolutional layers, each convolutional layer contains multiple channels, and each channel contains a data matrix; the input layer of the neural network has only one channel, which represents the matrix of the geometric configuration of the metal nanostructure; as the network goes deeper, the size of the convolutional layer matrix continues to shrink and the number of channels gradually increases. Each channel extracts certain feature information, and cooperates with the fully connected layer at the back end of the network to gradually convert the two-dimensional image into a one-dimensional vector representing a certain feature as the final network output, which represents the left-handed or right-handed reflectivity.

5. The design method according to claim 4, wherein: In step 2), the geometric configuration of the metal nanostructure is encoded as a "01" binary matrix as the input of the neural network, where "1" represents the presence of the metal nanostructure and "0" represents the absence of the metal nanostructure.

6. The design method according to claim 1, wherein: Step 3) When calculating the circular dichroism of a metal nanostructure, first use a neural network to predict its left-handed or right-handed reflectivity, and then predict the left-handed or right-handed reflectivity of the metal nanostructure after the specular reflection transformation. Due to the symmetry of the system, the latter is the right-handed or left-handed reflectivity of the original metal nanostructure. The circular dichroism is then calculated based on these two reflectivities: CD = (R LCP – R RCP ) / (R LCP + R RCP ) Where CD stands for optical circular dichroism, R LCP Represents the reflectivity under left-handed light incidence, R RCP Represents the reflectivity under right-handed light incidence.

7. The design method according to claim 1, wherein: The optimization objective function of step 3) comprehensively considers the half-width of the circular dichroism spectrum resonance peak and the maximum value of the circular dichroism.

Images