*2.2. Composite Micro Raman Detection Method for Microfluidic Chips*

The microfluidic chip composite micro Raman detection method proposed in this study has the characteristics of rapidity, accuracy, and simple operation. As shown in Figure 1, this method mainly consists of three parts: spore capture, separation and enrichment, micro Raman collection, and Raman data processing and modeling. In this study, nanospheres (3–8 μm) were mixed with four spore suspensions; the four kinds of spores were *U. virens*, *R. blast*, *A. niger*, and *A. carbonarius*. The spore concentration in the real environment is very low, and there will be disease spores in the air unless the disease is about to occur. Therefore, in order to simulate the real experimental environment, the spores were mixed into an aerosol generator and an aerosol was generated, and the aerosol was released into the air. In order to simulate the real use environment, the spore concentration in the air was consistent with the concentration before the outbreak of a disease, 200 spore /m3 [21]. An air pump was used to pump air into the microfluidic chip at the exit of the chip for about 3 min per acquisition. After the aerosol entered the microfluidic chip, the enrichment of 3–4 μm and 6–8 μm spores and impurities was completed through the two enrichment areas of the chip. The particles larger than 8 μm remained in the pretreatment channel due to their large inertia. The target particles entered the corresponding two enrichment areas, while the particles smaller than 3 μm flowed out of the chip through the outlet.

**Figure 1.** Detection of crop disease spores by microfluidic chip combined with micro Raman spectroscopy: (**A**) spore capture, separation, and enrichment; (**B**) micro Raman collection; (**C**) Raman data processing and modeling.

In this study, a total of 50 sets of Raman spectra of 4 species of spores were collected, and the collected 200 sets of Raman spectra were randomly divided into training set and test set at a ratio of 3:1. The samples were collected with a frequency shift of 200–2000 cm−<sup>1</sup>

using an XploRA PLUS Raman microscope (HORIBA, France), and the collected samples had 1053 features. The spectral acquisition parameters were set as follows: the excitation power was set to 365 mW, the 50 times objective lens was selected, spot size was 5 μm. Before the Raman spectrum was collected, the Raman spectrometer needed to be calibrated with wave number to eliminate the significant difference between the instrument response and the measured Raman spectrum value on the wave number axis and the true value. Wavenumber calibration requires measurement of the reference Raman spectrum from a standard material with well-defined Raman bands. In this study, the Raman spectrum peak of silicon wafer was selected to calibrate the spectrometer. When the first-order Raman spectrum peak of silicon wafer was located at the 520.7 cm−<sup>1</sup> frequency shift, the instrument was calibrated. In the laser illumination channel, the narrowband single-mode continuous light with the wavelength of 785 nm was selected as the excitation light of Raman scattering. This is because the 785 nm excitation light can effectively reduce the background spontaneous fluorescence noise and improve the signal-to-noise ratio of the collected Raman signal.

In this study, spectral data were randomly divided into training set (140) and test set (60). Spore identification model was established according to the test set. Before data analysis, the collected Raman spectra were preprocessed by SG smoothing and SNV, and the iterative polynomial fitting method was used for baseline correction to eliminate the interference of baseline drift and spectrum noise of Raman spectra.

The principle of polynomial replacement fitting is to continuously compare and adjust the original spectral data during polynomial replacement fitting, and directly compare the adjusted spectral data with the points on the fitting curve. The advantage of baseline correction with this method is to gradually adjust the coefficients of the polynomial so as to gradually approach the actual baseline shape, and the calculated baseline function form is closer to the actual baseline [22]. Standard normalized variate (SNV) algorithm refers to a deviation method to standardize variable values. Through the numerical standardization and transformation of the original variables, the transformation results will eventually fall within the range of [0, 1]. The premise of SNV algorithm is that all wavelength variables present normal distribution, and then they are standardized. The main way to remove noise is to remove light scattering [23].

Principal component analysis (PCA) is an algorithm for dimension reduction of data features. Spectral data were transformed from high to low dimensions by linear variation. Low-order principal components were retained and high-dimensional and invalid information was removed, reducing the data dimension. Using the most relevant low-dimensional data for classification identification can effectively reduce the difficulty and complexity of data analysis [24,25]. Usually, PCA needs to retain the principal components to make the variance contribution rate reach more than 85%. In this study, all PCAs with cumulative contribution rate greater than 95% were selected. The stability compatible reweighted sampling (SCARS) algorithm measures the magnitude of the stability of a variable, and the larger the stability value, the more likely the variable is to be selected, and the more consistent the bands selected at each iteration. This enables guaranteeing stable and rapid variable selection. The principles of the SCARS algorithm are to take each wavelength as one individual, use the adaptive reweighted sampling and exponentially revealing function to remove regression coefficients, take band points with small weights from the partial least squares model, pick out band points with large stable values, and retain the subset with the lowest RMSECV for interaction validation to find the optimal combination of variables with high efficiency [26].

SVM is a supervised machine learning method based on finite sample statistical learning theory. According to the structural risk minimization (SRM) principle, small samples and nonlinearity can be solved by constructing an optimal classification hyperplane in high-dimensional space. BPANN is a powerful learning algorithm that enables highly nonlinear mapping between inputs and outputs by training sample data, constantly modifying network weights and thresholds to minimize the error function in the direction of a negative gradient, ultimately approaching the expected output [27].

#### *2.3. Chip Design and Simulation*

To realize the purification and enrichment of fungal disease spores, a two-stage separation and enrichment microfluidic chip with arcuate pretreatment channel was designed. When a particle enters the inertial impactor with accompanying air, its trajectory is related to the size of the particle. Some small mass particles can cross stream lines and be separated. However, other small mass particles can flow away with the deflection of airflow. This behavior of particles in the curved channel can be characterized by Stokes number [17,28].

$$stk = \frac{\rho\_p d\_p^2 \mathbb{C}\_\varepsilon Q}{9\mu\mathcal{W}}\tag{1}$$

where *dp* is the particle size (m), *ρ<sup>p</sup>* is the particle density (1000 kg m<sup>−</sup>3), *μ* is the air viscosity (1.81 × <sup>10</sup>−<sup>5</sup> <sup>N</sup> · <sup>s</sup> · <sup>m</sup><sup>−</sup>2), *<sup>Q</sup>* is the air velocity at the inlet of the microfluidic device (m · <sup>s</sup><sup>−</sup>1), and *<sup>W</sup>* is the nozzle width (m). *Cc* is the Cunningham sliding correction coefficient based on particle size, which can be obtained by Equation (2) [16]:

$$\mathbb{C}\_{\mathbb{C}} = 1 + \frac{2A\lambda}{d} + \frac{2Q\lambda}{d}e^{-\frac{\Delta\xi}{2\lambda}} \tag{2}$$

where *A* = 1.234, *Q* = 0.413, b = 0.904, and λ is the average free path of an air molecule with a value of 6.95 × <sup>10</sup>−<sup>8</sup> m. Thus, according to Equation (3) it can be reduced to

$$\mathcal{C}\_{\mathbb{C}} \approx \begin{cases} 1 + 2.52 \frac{\lambda}{d'}, d > 2\lambda \\ 1 + 3.29 \frac{\lambda}{d'}, d < 2\lambda \end{cases} \tag{3}$$

Furthermore, *stk*<sup>50</sup> represents a Stokes number corresponding to a particle collection efficiency of 50%, which can be rearranged by Equation (1):

$$d\_{50} = \sqrt{\frac{9\mu \mathcal{W}\_{stk\_{50}}}{\rho\_p \mathbb{C}\_c V}}\tag{4}$$

The *d*<sup>50</sup> is defined as the cut-off size of the particles producing 50% collection efficiency at each impact stage, and for this study, *U. virens*, *R. blast*, *A. niger*, and *A. carbonarius* collected at *d*<sup>50</sup> were set as 3–5 μm and 4–6 μm to obtain the up-to-size *W* for the two separated enrichment stages of the chip.

The microfluidic chip consists of an arc-shaped preprocessing channel and a two-stage separation and enrichment structure, and the particles enter the chip inlet jointly with air, then enter the preprocessing channel; the particles bonded together will be scattered and then enter the first enrichment area, and the remaining particles enter the second enrichment area due to the constant velocity at the inlet, while the particles that have a wide width channel enter the narrow width channel, so the airflow velocity is elevated, using velocity variation to manipulate particle enrichment versus rounding, which greatly increases collection efficiency. The structure of the microfluidic chip is schematically shown in Figure 2.

The preprocessing channel includes particle inlet, channel 1, and arc channel. The first separation structure includes channel 1, collection area 1, and channel 2. The second separation structure includes channel 3, enrichment zone 2, channel 4, and particle outlet. R1 is the radius of the pretreatment channel, and D1 and D2 are the diameters of the collection area, which are 5000 μm and 3700 μm, respectively. The length of channel 0 is 7000 μm, and the width is set to W0 = 1300 μm. Channel 1 has a length of 3500 μm and a width of 800 μm. Channel 2 has a length of 5650 μm and a width of 1100 μm. The length of

channel 3 was set to 4400 μm, and the width was set to 400 μm. The length of channel 4 was set to 3600 μm, and the width was set to 600 μm.

**Figure 2.** A 2D diagram of the microfluidic chip.
