A full-spectrum detection method for whispering gallery microcavity based on machine learning

Abstract

The present invention discloses a full-spectrum detection method of a whispering gallery microcavity based on machine learning, which belongs to the field of optical microcavity sensing / detection. First, an experimental device for the full spectrum of a whispering gallery microcavity is built, and the whispering gallery mode transmission spectrum under known different target parameters is collected in real time as a sensor signal training set for training a multi-layer perceptron neural network, and the weight is adjusted according to the gradient descent method to establish a nonlinear mapping relationship between the input whispering gallery mode transmission spectrum and the output target parameter. Then, a new whispering gallery mode transmission spectrum is collected under different target parameters to be measured, and input into the trained multi-layer perceptron neural network, and the target parameter measurement result corresponding to the transmission spectrum is directly output with high precision. The present invention extracts the overall characteristics of the spectrum, makes full use of the sensing information contained in the spectrum, provides more degrees of freedom, and provides a new method for realizing high-precision reading of the absolute value of the target to be measured and multi-parameter parallel detection based on a single microcavity sensing platform.

Description

Technical Field

[0001] The present invention belongs to the field of optical microcavity sensing / detection, and specifically relates to a whispering gallery microcavity full-spectrum detection method based on machine learning. Background Art

[0002] In recent years, optical microcavities have attracted the attention of researchers at home and abroad due to their ultrahigh sensitivity, rapid response, and tiny size, and have been widely used in biomedical research, health care, environmental monitoring, and homeland security.

[0003] Commonly used optical microcavities include whispering gallery microcavities, photonic crystal microcavities, and Fabry-Perot cavities. Among them, the whispering gallery microcavity has an ultra-high quality factor (Q) and a tiny mode volume (V), which can constrain the energy of the light field in time and space dimensions, greatly enhance the interaction between light and matter, and thus significantly improve the sensing sensitivity. It is an ideal platform for achieving ultra-high sensitivity optical detection. At present, the whispering gallery microcavity combined with localized surface plasmon enhancement, laser mode locking, optical spring effect and other technologies or mechanisms has reached the level of single molecule and even single ion detection, such as reference 1. Yu, D., Humar, M., Meserve, K. et al. "Whispering-gallery-mode sensors for biological and physical sensing," Nature Reviews Methods Primers 1, 83 (2021).

[0004] Although the whispering gallery microcavity has excellent detection sensitivity, the traditional detection mechanism relies on the sensing performance of a single specific mode (such as the shift of a single mode resonant wavelength), which is greatly limited by the sweep frequency range of the laser. In contrast, the multiple resonant modes in the same spectrum of the whispering gallery microcavity have different responses to the target to be measured, which provides rich sensing information for the measurement of the parameter to be measured. By combining these sensing modes, a larger detection range and smaller uncertainty can be achieved. For example, in 2020, the research group of Professor Dong Chunhua of the University of Science and Technology of China realized voltage detection by extracting the transmission depth of multiple resonant modes, and the detection limit was 6.7% lower than that of the single-mode sensing mechanism, as shown in reference 2. Lu J, Niu R, Wan S, et al. "Experimental demonstration of multimode microresonator sensing by ML," IEEE Sensors Journal, 2021, 21(7): 9046-9053. In 2021, the research group of Professor Lan Yang of Washington University in St. Louis proposed a barcode technology based on multiple whispering gallery modes, which achieved high-precision measurement of actual temperature within a wide dynamic range, as shown in reference 3. Liao, J., Yang, L. "Optical whispering-gallery mode barcodesfor high-precision and wide-range temperature measurements," Light Science&Applications 10, 32(2021).

[0005] In fact, the global features of the transmission spectrum, such as resonant wavelength, number of resonant modes, transmission depth, mode spacing and mode linewidth, also contain effective sensing information. Therefore, the extraction of the overall characteristics of the spectrum can further make full use of this information and provide more and more opportunities for high-precision detection. However, traditional microcavity transmission spectra often exhibit a high spectral density. For example, within a laser sweep range of 1nm, the number of resonant modes of the whispering gallery microbubble cavity is as high as 305; such as reference 3. Liao, J., Yang, L. "Optical whispering-gallery mode barcodes for high-precision and wide-range temperature measurements," Light Science & Applications 10, 32 (2021). Therefore, it is a great challenge to manually analyze a large number of optical resonance modes and establish a mapping relationship between different optical resonance mode responses and the target to be measured.

[0006] As a data-driven method, machine learning can automatically search for effective information from a large number of data sets, revealing the inherent mechanisms hidden behind the data, thereby establishing a mapping relationship between the input spectrum and the target to be measured, providing an alternative method for the analysis of complex optical data; for example, reference 4. Y.LeCun, Y.Bengio, and G.Hinton., “Deeplearning,” Nature 521, 436-444 (2015).

[0007] In recent years, machine learning has attracted widespread attention from researchers at home and abroad in the field of optical inspection. For example, in 2019, Jennifer Dionne's group at Stanford University used convolutional neural networks to analyze Raman spectra and achieved accurate identification of 30 common bacterial pathogens at low signal-to-noise ratio (~4.1dB); such as reference 5.CSHo, N.Jean, and CAHogan, et al., "Rapid identification of pathogenic bacteria using Raman spectroscopyand deep learning," Nature Communications 10, 4927 (2019); in 2021, Professor Hao Yilong of Peking University and his collaborative team used a multilayer perceptron to analyze the real-time offset of the resonant mode based on a label-free microring resonator, and achieved the identification of the concentration of a mixture of glucose, polyethylene glycol and bovine serum albumin, with a root mean square error of 0.13 to 2.28 mg / mL; such as reference 6.Z.Li, H.Zhang, BTT Nguyen, et al. "Smart ring resonator-based sensorfor multicomponent chemical analysis via machine learning," Photonics Research 9, B38-B44 (2021). Summary of the invention

[0008] The present invention proposes a full-spectrum detection method for a whispering gallery microcavity based on machine learning, which can extract the overall characteristics of the spectrum, make full use of the sensing information contained in the transmission spectrum, and mine the complex nonlinear mapping relationship between the target to be measured and the spectrum, providing a new technology for high-precision measurement of the actual values ​​of the target parameters to be measured.

[0009] The whispering gallery microcavity full spectrum detection method based on machine learning has the following specific steps:

[0010] Step 1: Build an experimental device for the full spectrum of the whispering gallery microcavity, and collect the whispering gallery mode transmission spectrum under different known target parameters as the sensing signal in real time;

[0011] Different target parameters include pressure, temperature, solution concentration, and magnetic field;

[0012] The experimental device comprises: a laser and an attenuator connected thereto in sequence, a polarization controller, a whispering gallery microcavity, a photodetector, an oscilloscope, a data acquisition card, a computer and a signal generator;

[0013] The light emitted by the laser passes through the variable attenuator and the fiber polarization controller in turn. After adjusting the polarization state and power, it enters the whispering gallery microcavity through the fiber cone coupling to excite the whispering gallery mode. The light signal at the output end is converted into an electrical signal after passing through the photodetector, one path enters the oscilloscope to monitor the sensor signal in real time, and the other path enters the data acquisition card to collect the sensor signal in real time.

[0014] Step 2: Collect the corresponding whispering gallery mode transmission spectra under different target parameters as training sets.

[0015] Step 3: The training set is input into the multi-layer perceptron neural network for training through the back propagation algorithm, and the weights are adjusted according to the gradient descent method to establish a nonlinear mapping relationship between the input whispering gallery mode transmission spectrum and the output target parameters;

[0016] The multilayer perceptron neural network comprises a single input layer, a single hidden layer and a single output layer;

[0017] The input layer includes n neurons, and the transmission spectrum composed of n neurons of each sample is used as input to pass the sensing information to the hidden layer, which contains m neurons. The output layer contains 1 neuron, which is the target parameter to be measured;

[0018] The specific implementation process is as follows:

[0019] First, the input value of the i-th neuron in the input layer is: x(i) = [x1, x2, ... x n ]; i = 1, 2, ... n;

[0020] Subsequently, the input value of the sample is input into the fully connected multilayer perceptron neural network, and the jth neuron in the hidden layer is expressed as follows:

[0021]

[0022] Among them, w′0 is the bias, w′ ij is the weight factor of the connection between the i-th neuron in the input layer and the j-th neuron in the hidden layer, F(·) is the nonlinear activation function, j = 1, 2, … m;

[0023] Finally, the information is passed from the hidden layer to the output layer, and the value of the output layer neuron is determined by the following formula:

[0024]

[0025] Among them, w0 is the bias, w j is the weight factor connecting the jth neuron in the hidden layer and the neurons in the output layer, and F′(·) is the nonlinear activation function.

[0026] Step 4: Collect new whispering gallery mode transmission spectra under different target parameters and input them into the trained multi-layer perceptron neural network to directly output the target parameter measurement results corresponding to the transmission spectrum.

[0027] The present invention has the following advantages:

[0028] 1. The present invention proposes the cross-application of optical microcavity detection and machine learning, which can effectively extract the overall characteristics of the spectrum, reveal the inherent mechanism behind the data, explore the complex relationship between the input spectrum and the measured parameters, and achieve high-precision measurement of the absolute value of the target to be measured within a wide dynamic range.

[0029] 2. Different from the traditional sensing mechanism that tracks the changes of a single specific resonance mode, the present invention realizes the detection of the measured parameter by extracting the full spectrum information, which can make full use of the effective information sensed by the resonance mode and improve the detection accuracy, with an accuracy of up to 99.97%.

[0030] 3. The present invention realizes the real-time reading of the absolute value of the parameter to be measured outside the training data set, with an accuracy rate of 98.94%, and has excellent generalization ability.

[0031] 4. The present invention can provide more degrees of freedom, providing an opportunity for realizing independent decoupling and precise measurement of multiple parameters based on a single microcavity sensing platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of a full spectrum detection method of a whispering gallery microcavity based on machine learning of the present invention;

[0033] Figure 2 (a) is a schematic diagram of the experimental device of the present invention, and (b) is a schematic diagram of the characterization of the quality factor of the whispering gallery microbubble cavity;

[0034] Figure 3 Schematic diagram of the characterization of the coupling stability of the whispering gallery mode microcavity of the present invention;

[0035] Figure 4 It is a schematic diagram of the principle of realizing microcavity full-spectrum sensing based on machine learning method of the present invention;

[0036] Figure 5 (a) is the result of the machine learning method of the present invention for measuring the absolute value of the parameter to be measured, and (b) is the measurement result of the unknown target parameter outside the training set. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0038] The present invention proposes a full-spectrum detection method for whispering gallery microcavity based on machine learning, which can extract the overall characteristics of the spectrum, make full use of the contained sensing information, realize high-precision measurement of the absolute value of the parameter to be measured, and provide more degrees of freedom, thus providing an opportunity for a single microcavity sensing platform to realize independent decoupling and precise measurement of multiple parameters.

[0039] The whispering gallery microcavity full spectrum detection method based on machine learning, such as Figure 1 As shown, the specific steps are as follows:

[0040] Step 1: Build an experimental device for the full spectrum of the whispering gallery microcavity, and collect the whispering gallery mode transmission spectrum under different known target parameters as the sensing signal in real time;

[0041] like Figure 2 As shown in (a), the experimental device includes: a laser and an attenuator connected thereto in sequence, a polarization controller, a whispering gallery microcavity, a photodetector, an oscilloscope, a data acquisition card, a computer and a signal generator;

[0042] A tunable laser is used as the light source at the input end. The light emitted by the laser is adjusted in polarization state by a fiber polarization controller in turn. After the power is adjusted by a variable attenuator, the light enters the whispering gallery microcavity through fiber cone coupling to stimulate the whispering gallery mode. The optical coupling efficiency of the fiber cone can exceed 95%. The signal at the output end is converted into an electrical signal after passing through a photodetector. One path enters an oscilloscope to monitor the sensor signal in real time; the other path enters a data acquisition card to collect sensor signals in real time as training data sets and test data sets for the multi-layer perceptron neural network model.

[0043] In addition to pressure, target parameters can also be temperature, magnetic field, solution concentration, etc.

[0044] The target parameter selected in this embodiment is pressure. One end of the whispering gallery microcavity is sealed with ultraviolet glue, and the other end is connected to a pressure valve and a pressure gauge through a microfluidic tube for controlling and recording the pressure.

[0045] Step 2: Collect the corresponding whispering gallery mode transmission spectra under different target parameters as training sets.

[0046] In this embodiment, the pressure is measured, and a syringe is selected to pressurize the microcavity through a microfluidic tube, and a pressure gauge is used to record the applied pressure. The pressure valve is equivalent to a switch.

[0047] Step 3: The training set is input into the multi-layer perceptron neural network for training through the back propagation algorithm, and the weights are adjusted according to the gradient descent method to establish a nonlinear mapping relationship between the input whispering gallery mode transmission spectrum and the output target parameters to be measured;

[0048] The multilayer perceptron neural network comprises a single input layer, a single hidden layer and a single output layer;

[0049] The input layer includes n neurons, k is the number of transmission spectrum samples in each model training, and the transmission spectrum composed of n neurons of each sample is used as input, and the target parameters corresponding to each transmission spectrum are used to label and convert into a matrix; that is, the independent variable is composed of a transmission spectrum matrix of k×1×n;

[0050]

[0051] The matrix containing the sensor information is input into the network model and passed to the hidden layer. The hidden layer contains m neurons and the output layer contains 1 neuron, which is the target parameter to be detected.

[0052] The specific implementation process is as follows:

[0053] The input value of the i-th neuron in the input layer is: x(i) = [x1, x2, ... x n ]; i = 1, 2, ... n;

[0054] Subsequently, the input value of the sample is input into the fully connected three-layer perceptron neural network, and the jth neuron in the hidden layer is expressed as follows:

[0055]

[0056] Among them, w′0 is the bias, w′ ij is the weight factor of the link between the i-th neuron in the link input layer and the j-th neuron in the hidden layer, F(·) is the nonlinear activation function, j = 1, 2, … m;

[0057] Finally, the information is passed from the hidden layer to the output layer, and the value of the output layer neuron is determined by the following formula:

[0058]

[0059] Among them, w0 is the bias, w j is the weight factor linking the jth neuron in the link hidden layer to the neurons in the output layer, and F′(·) is the nonlinear activation function.

[0060] After reaching the training goal (training accuracy reaches more than 95%), a high-quality multi-layer perceptron neural network is obtained.

[0061] Step 4: Start the actual measurement, collect additional whispering gallery mode transmission spectra under different target parameters as test sets, and input them into the trained multi-layer perceptron neural network to directly output the target parameter measurement results corresponding to the transmission spectrum.

[0062] It should be noted that the training dataset and the test dataset are recorded independently without any overlap, which ensures that the trained network does not contain the information of the test dataset.

[0063] Example:

[0064] In this example, the wavelength of the detection light used is in the 780nm band; the single-mode fiber cone used is mainly composed of silicon dioxide, with a diameter of 1-3μm, and is used to efficiently couple the incident light into and out of the whispering gallery microcavity; the whispering gallery microcavity used is a microbubble cavity with a natural microfluidic channel, mainly composed of silicon dioxide, with a diameter of 60-300μm, a wall thickness of 1-5μm, and a thermal conductivity of more than 10 7 The optical Q value, such as Figure 2 As shown in (b); the relative position between the two is precisely controlled using a micro-nano translation stage. In order to ensure the stability of the coupling system, the microbubble cavity and the optical fiber cone are tightly attached together.

[0065] In this example, the transmission spectra of whispering gallery modes with known pressures are first collected as training sets for training the neural network model; then, actual measurements are started, and additional transmission spectra of whispering gallery modes are collected at different pressures as test sets. Next, the training set is input into the network for training through the back propagation algorithm, which adjusts the weights according to the gradient descent method and establishes a nonlinear mapping relationship between the input variables (transmission spectra of whispering gallery modes) and the output variables (corresponding pressures); finally, the test set is input into the trained network, and the final pressure measurement results are directly output by the network.

[0066] 1. The long-term stability of the whispering gallery microbubble cavity used in this experiment was tested in real time, such as Figure 3 As shown, the maximum shift of the resonant wavelength obtained by the test is ~1.242pm, and the maximum change in transmittance is ~0.115679. The microcavity used has good stability, which lays the foundation for the collection of high-quality data sets.

[0067] 2. The whispering gallery microcavity transmission spectrum used in the present invention has multiple resonance modes. The machine learning algorithm can extract the overall characteristic information of the spectrum, establish a nonlinear mapping relationship between the input transmission spectrum and the pressure to be measured, and realize the accurate measurement of the actual pressure. The machine learning algorithm used in this example is a multilayer perceptron (MLP), which includes a single input layer, a single hidden layer, and a single output layer. Figure 4As shown. The learning rate of the network model used is 10 -3 , Epoch is 30, the number of hidden layer neurons is 100, the activation function is linear rectification function (Linear rectification function, ReLU), the loss function is Mean Square Error (MSE), the number of training sets under each pressure is 600 groups, and the number of test sets is 150 groups.

[0068] 3. The present invention measures the transmission spectrum of pressure in the range of 0-100 kPa with a step of 5 kPa, and the prediction accuracy of the pressure to be measured is as high as 99.97%. Figure 5 In addition, the transmission spectra in the pressure range of 30-80 kPa and 35-75 kPa were collected as training data sets and test data sets respectively with a step of 10 kPa, and the calculated prediction accuracy was 98.94%, as shown in Figure 5 As shown in (b), this method has excellent generalization ability.

Claims

1. A full spectrum detection method for whispering gallery microcavity based on machine learning, characterized in that: The specific steps are as follows: Firstly, an experimental device for full spectrum of whispering gallery microcavity was built to collect the transmission spectrum of whispering gallery mode under different known target parameters as sensing signals in real time. The experimental device comprises: a laser and an attenuator connected thereto in sequence, a polarization controller, a whispering gallery microcavity, a photodetector, an oscilloscope, a data acquisition card, a computer and a signal generator; The light emitted by the laser passes through the variable attenuator and the fiber polarization controller in turn. After adjusting the polarization state and power, it enters the whispering gallery microcavity through the fiber cone coupling to stimulate the whispering gallery mode. The light signal at the output end is converted into an electrical signal after passing through the photodetector. One path enters the oscilloscope to monitor the sensor signal in real time; the other path enters the data acquisition card to collect the sensor signal in real time. Then, the optical microcavity detection and machine learning methods are cross-applied. The corresponding whispering gallery mode transmission spectra are collected under different target parameters as training sets, which are input into the multi-layer perceptron neural network to extract the overall characteristics of the spectrum. The training is carried out through the back-propagation algorithm, and the weights are adjusted according to the gradient descent method. All the sensor information contained in the spectrum is fully utilized to establish a nonlinear mapping relationship between the input whispering gallery mode transmission spectrum and the output target parameters. The multilayer perceptron neural network comprises a single input layer, a single hidden layer and a single output layer; The input layer includes n neurons, and the transmission spectrum of n neurons of each sample is used as input to pass the sensing information to the hidden layer, which contains m neurons. The output layer contains 1 neuron, which is the target parameter to be detected; The specific implementation process is as follows: First, the input value of the i-th neuron in the input layer is: x(i) = [x i1 ,x i2 ,…x in ]; i = 1, 2, ... n; Subsequently, the input value of the sample is input into the fully connected multilayer perceptron neural network, and the jth neuron in the hidden layer is expressed as follows: Among them, w′0 is the bias, w′ ij is the weight factor of the link between the i-th neuron in the input layer and the j-th neuron in the hidden layer, F(·) is the nonlinear activation function, j = 1, 2, … m; Finally, the information is passed from the hidden layer to the output layer, and the value of the output layer neuron is determined by the following formula: Among them, w0 is the bias, w j is the weight factor of the link between the jth neuron in the link hidden layer and the neuron in the output layer, and F′(·) is the nonlinear activation function; Finally, new whispering gallery mode transmission spectra are collected under different target parameters and input into the trained multi-layer perceptron neural network to directly output the target parameter measurement results corresponding to the new whispering gallery mode transmission spectra with high precision.

2. The full spectrum detection method of whispering gallery microcavity based on machine learning as claimed in claim 1, characterized in that: The different target parameters include physical parameters and biochemical substance information measurements.

3. The full spectrum detection method of whispering gallery microcavity based on machine learning as claimed in claim 2, characterized in that: When collecting the corresponding whispering gallery mode transmission spectrum for pressure, one end of the whispering gallery microbubble cavity in the experimental device is sealed with UV glue, and the other end is connected to a pressure valve and a pressure gauge through a microfluidic tube to control and record the pressure.

Images