Prediction and preparation method of an antibacterial surface of polyetheretherketone based on machine learning and femtosecond laser

Abstract

The present invention discloses a method for predicting and preparing an antibacterial surface of polyetheretherketone based on machine learning and femtosecond laser, including processes such as the establishment of machine learning tools, the preparation of surface structures, the acquisition and analysis of surface topography information, the prediction of antibacterial ability, and experimental verification. The original machine learning method based on surface images can significantly shorten the time required for scientific research to explore high antibacterial structures on the surface of polyetheretherketone, save experimental costs, and lay a foundation for the efficient preparation of green, safe, and reliable implants.

Description

Technical Field

[0001] The present invention belongs to the technical field of bone implant preparation, and particularly relates to a method for predicting and preparing an antibacterial surface of polyetheretherketone based on machine learning and femtosecond laser. Background Art

[0002] Polyetheretherketone has characteristics such as an elastic modulus similar to that of human bone, high elongation at break, corrosion resistance, high temperature resistance, and good biocompatibility, and is an excellent bone implant material. However, the surface of untreated polyetheretherketone material is easily wrapped by fibrous tissue or colonized by bacteria. The formation of a biofilm on the polyetheretherketone surface can cause exponential growth of bacterial colonies such as Staphylococcus epidermidis (S. epidermidis), Staphylococcus aureus (S. aureus), Escherichia coli (E. coli), etc., and linear growth of bacterial colonies of Enterococcus faecalis (E. faecalis). Compared with the surface of Ti or Si3N4, the biofilm affinity of the polyetheretherketone surface is 6.7 times higher than that of Ti and 16 times higher than that of Si3N4. At the same time, the number of live bacteria on the polyetheretherketone surface is also the highest, up to 30 times that of the Si3N4 surface. This will cause the polyetheretherketone implant to fail to form a good bond with bone tissue, and may even cause bacterial infection and lead to implant failure. Therefore, the research on the antibacterial surface of polyetheretherketone has attracted much attention in recent years.

[0003] The attachment of bacteria to the implant surface is a common phenomenon. Initially, planktonic bacterial cells will attach to the implant surface, then they will proliferate and secrete extracellular polymeric substances, and finally multiple layers of cells will aggregate and form a biofilm. Compared with planktonic bacterial cells, bacterial cells living in the biofilm have stronger resistance to the host immune system, external physical stress, chemical bactericides, and antibiotics, often leading to chronic infection and bone resorption around the implant, which may cause the implant to loosen and fall off, ultimately resulting in implant failure. Therefore, preventing the occurrence of inflammation around the implant and blocking the attachment of bacteria and the formation of biofilm are of great significance for the success or failure of the implant surgery and the durability of the implant.

[0004] Since the attachment of bacteria is the first step in the formation of biofilm, preventing the initial attachment of bacteria is crucial for preventing the formation of biofilm. Changing the chemical composition of the material surface, including doping with toxic bactericides and grafting functional groups, is considered an effective method to prevent bacterial attachment. In addition, it has also been found that the surface topography of the material has an important impact on the interaction between bacteria and the surface, and the attachment of bacteria can also be blocked by physically modifying the material surface. The antibacterial ability generated by physical methods is usually effective for a long time and environmentally friendly, so it is considered a very promising alternative to chemical modification.

[0005] The surface roughness and surface morphology of materials are the main physical factors that affect the adhesion of bacteria on the surface. First of all, for surface roughness, some scholars believe that since the adhesion force increases with the increase of roughness, a surface with low roughness is more conducive to inhibiting the initial adhesion of bacteria. However, some scholars hold the opposite view, believing that surface roughness has almost no effect on the initial adhesion of bacteria, and even higher surface roughness is sometimes conducive to inhibiting the initial adhesion of bacteria. Therefore, there is still a lack of consensus on the relationship between surface roughness and bacterial adhesion in this field. This may be because most studies consider the surface roughness of the material while not deeply controlling the surface morphology of the material, thereby ignoring the impact of different surface morphologies. But on the other hand, changes in surface roughness are usually accompanied by changes in surface morphology. In this type of study, it is very difficult to control one variable in isolation to discuss another variable.

[0006] Therefore, for the problem of how to obtain a surface that can inhibit the initial adhesion of bacteria, a good solution is to use a type of analysis tool that can comprehensively consider multiple factors at the same time, and the "black box algorithm" of machine learning happens to provide such an optional solution. Summary of the invention

[0007] To this end, the present invention provides a prediction and preparation method for the antibacterial surface of polyetheretherketone based on machine learning and femtosecond laser: a femtosecond laser is used to prepare a variety of micrometer / nanoscale structures with different roughness and morphology on the surface of polyetheretherketone, and the antibacterial properties of the surface are predicted by machine learning, and finally verified by actual antibacterial experiments. The results of the antibacterial experiment also prove that the trained machine learning model has high accuracy, thus opening up a new path for the rapid, accurate and efficient preparation of antibacterial surfaces, solving the problems of inaccurate preparation process, low preparation efficiency and uncertain preparation results caused by the lack of unified and standardized preparation means in the prior art, making it possible to normalize and standardize the preparation process of the antibacterial surface.

[0008] The purpose of the present invention is achieved through the following technical solutions:

[0009] The present invention discloses a method for predicting and preparing the antibacterial surface of polyetheretherketone based on machine learning and femtosecond laser:

[0010] S1. Surface preparation: obtaining polyetheretherketone surfaces with different morphologies and roughness by linearly adjusting the energy density of the laser pulse;

[0011] S2. Obtaining structural information of the polyetheretherketone surfaces of different morphologies, including:

[0012] 1) The quasi-three-dimensional morphology information of the surface is obtained by using a scanning electron microscope for emission, and the third-dimensional information, that is, the protrusion height or depression depth of the material surface, is reflected by the gray value of the generated pixel points;

[0013] 2) The true three-dimensional morphology information of the surface is obtained by using an atomic force microscope, and the third-dimensional information, that is, the protrusion height or depression depth information of the material surface, is reflected by the real height value or depth value;

[0014] S3. Prediction and analysis: According to the three-dimensional structure information of the material surface obtained in step 2, a machine learning tool is used for classification and prediction to predict the antibacterial ability of the polyetheretherketone surface with different morphologies, and the best antibacterial structure is selected;

[0015] S4. Through actual antibacterial experiments, verify the antibacterial performance of the polyetheretherketone surface with different morphologies, and test the accuracy of the prediction by the machine learning method.

[0016] Furthermore, it also includes the selection and training of a machine learning model. The training steps include: obtaining the surface structure information and corresponding antibacterial performance data of known antibacterial materials from published literature and experimental data, selecting and training a machine learning model for classification and prediction. The surface structure information of the antibacterial material includes the statistics extracted based on the quasi-three-dimensional and true three-dimensional morphology information, and the statistics describing the surface texture information provided by the gray-level co-occurrence matrix generated therefrom, and these statistics form a high-dimensional antibacterial data set.

[0017] Furthermore, the process of obtaining the quasi-three-dimensional and true three-dimensional morphology information of the material surface includes reading each SEM image to generate a two-dimensional gray matrix, and the matrix elements take discrete values; or reading each AFM image to generate a two-dimensional matrix describing the surface undulation information, and the matrix elements take continuous values. Based on these two types of two-dimensional matrices, the variance and one-dimensional information entropy can be calculated, and their specific expressions are as follows:

[0018] 1) Variance: The variance of all matrix element values in the matrix, reflecting the contrast information of the material surface, characterizing the unevenness of the surface undulation of the material,

[0019]

[0020] where m and n are the dimensions of the matrix; Q(x,y) is the matrix element value in the x-th row and y-th column of the matrix, corresponding to the gray value of the pixel at the corresponding position in the SEM image, or the actual height value of the corresponding point in the AFM image; is the average value of all matrix elements, reflecting the average brightness of each point in the SEM image, or the average height of each point in the AFM image;

[0021] 2) One-dimensional information entropy: Reflecting the richness and complexity of the information contained in an image;

[0022] The one-dimensional information entropy calculated based on the SEM grayscale image information of the quasi-three-dimensional topography information is as follows:

[0023]

[0024] In the formula, f(i) represents the proportion of pixels with grayscale value i: #(i) represents the number of pixels with grayscale value i, m and n are the dimensions of the grayscale matrix. In an 8-bit grayscale image, the value range of i is an integer between 0 and 255.

[0025] When calculating the one-dimensional information entropy based on the AFM image of the true three-dimensional topography information, it is necessary to first discretize the values of the matrix elements. By dividing the grid intervals, the number of matrix elements within a given range can be counted, and then the one-dimensional information entropy can be calculated using the above formula. When calculating the one-dimensional information entropy based on the true three-dimensional topography information, it is not affected by the number of bits of the image, and the number of intervals divided can be flexibly selected according to actual needs.

[0026] Furthermore, the statistic for describing the surface texture information of the material further includes generating a gray-level co-occurrence matrix based on the quasi-three-dimensional and true three-dimensional information of the surface, and obtaining four statistics of energy, contrast, correlation, and uniformity through the gray-level co-occurrence matrix. The specific expressions are as follows:

[0027] 1) Gray-level co-occurrence matrix: The gray-level co-occurrence matrix generated based on the SEM grayscale image of the quasi-three-dimensional topography information is defined as follows

[0028]

[0029] where

[0030] In an 8-bit grayscale image, the maximum dimension of the obtained gray-level co-occurrence matrix is 256×256.

[0031] To generate a gray-level co-occurrence matrix based on the AFM image of the true three-dimensional information, it is necessary to first discretize the values of the matrix elements. As mentioned above, when dividing the number of intervals based on the true three-dimensional topography information, it can be carried out according to actual needs and is not limited by the number of bits of the image.

[0032] Considering that the characteristic scale of the surface of materials related to antibacterial is usually between 0.1μm and 2μm, when performing the above calculations, regardless of whether it is based on the SEM image or the AFM image, we respectively take all possible pixel point spacings between 0.1μm and 2μm and calculate the gray-level co-occurrence matrices at different spatial scales. At the same spatial scale, we respectively take 4 different spatial directions of horizontal, vertical, 45 degrees to the left, and 45 degrees to the right to calculate 4 different gray-level co-occurrence matrices, and the obtained statistics are respectively averaged to obtain the most general results.

[0033] 2) Statistics for describing surface texture information: Here, taking the gray-level co-occurrence matrix generated from the SEM image based on quasi-three-dimensional topography information as an example, the gray-level co-occurrence matrix generated from the AFM image based on true three-dimensional information is the same algorithmically, only the value range of pixel gray levels is more flexible.

[0034] ① Energy:

[0035]

[0036] ② Contrast

[0037]

[0038] ③ Correlation

[0039]

[0040] where μ i , μ j , σ i 2 , σ j 2 are the mean and variance of the elements in the i-th row (or j-th column) of the gray-level co-occurrence matrix, respectively.

[0041] ④ Homogeneity

[0042]

[0043] Furthermore, the machine learning models used include support vector machine classifier, naive Bayes classifier, gradient boosting algorithm, adaptive boosting algorithm, random forest classifier, and multi-layer perceptron;

[0044] Testing steps: For each machine learning model, in 50 independent processes respectively, using the train_test_split tool of the Scikit-learn library, the data set is randomly divided into a training set and a test set; the training set is used to train the model, and the test set is used to evaluate the performance and generalization ability of the trained model; according to the scale of the collected data set, 25% of the test set data is selected from the entire data set; in each independent process, first through 10-fold cross-validation, using the GridSearchCV tool to perform grid search on the hyperparameter space to determine the most suitable hyperparameter combination for the current task for each model; then use the selected optimal hyperparameter combination to train the final model and make predictions on the test set data; finally, evaluate the performance of the model on the test set according to the existing classification labels, including: accuracy, precision, and recall; each performance metric is averaged over 50 independent processes

[0045] During the test, the classification threshold increases by 0.1, and the classification threshold range is 0.1 to 0.9.

[0046] Furthermore, in step 1, the change range of the laser energy density is 0 to 4 J / cm 2 .

[0047] Furthermore, when preparing the surface, a multi-level micro-nano composite structure was established by using a linear scanning method. The scanning line spacing was fixed at 13 - 18 μm, the scanning speed was fixed at 12 - 16 mm / s, and the spot size was fixed at 26 - 35 μm to achieve approximately 2 - 5 pulses in the irradiation area of each light spot.

[0048] Furthermore, the surface area for obtaining three-dimensional information is 20×20 μm 2 .

[0049] Compared with the prior art, the present invention has at least the following advantages and beneficial effects:

[0050] The present invention creatively uses machine learning tools to achieve the prediction of the antibacterial ability of the material surface, and the prediction accuracy is about between 70% and 80%, proving the feasibility of predicting the antibacterial rate through the morphological information of the material surface;

[0051] The present invention provides a green and pollution-free method for the one-step preparation of polyetheretherketone antibacterial surfaces, which has application potential in both the medical and industrial fields;

[0052] The present invention provides a new idea for the preparation of the surface of polyetheretherketone antibacterial structures, which is expected to greatly shorten the scientific research exploration time of the surface structure of polyetheretherketone antibacterial materials and has application potential in the preparation of anti-fouling surfaces in the industrial field and the preparation of medical polyetheretherketone antibacterial surfaces;

[0053] The present invention opens up a new path for the rapid, accurate and efficient preparation of antibacterial surfaces, making it possible to standardize and standardize the process of preparing antibacterial surfaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0055] Figure 1 is the basic principle diagram of the present invention;

[0056] Figure 2 is the flow chart of the present invention;

[0057] Figure 3 Accuracy of various machine learning algorithms for classifying antibacterial surfaces;

[0058] Figure 4 Morphology, roughness and wettability of the samples after laser treatment;

[0059] (A) Morphology and roughness of the FS-A surface; (B) Morphology and roughness of the FS-B surface; (C) Morphology and roughness of the FS-C surface; (D) Plot of the change in wettability with increasing laser energy density; (E) Change in surface morphology with increasing laser energy density;

[0060] Plot of the change in wettability with increasing laser energy density; (E) Change in surface morphology with increasing laser energy density;

[0061] Figure 5 Antibacterial effect of FS-A against Escherichia coli and machine learning prediction diagram;

[0062] Figure 6 Antibacterial effect of FS-A against Staphylococcus aureus and machine learning prediction diagram;

[0063] Figure 7 Machine learning flow chart; Specific implementation manners

[0064] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all of the embodiments. Generally, the components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations.

[0065] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0066] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0067] It should be noted that like reference numerals and letters denote like items in the following drawings, and thus, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0068] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the inventive product is usually placed during use, or the orientation or positional relationship commonly understood by those skilled in the art. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation on the present invention. In addition, the terms "first", "second", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0069] In the description of the present invention, it should also be noted that unless otherwise clearly specified and defined, the terms "set", "install", "connect", and "couple" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0070] The present invention discloses a method for predicting and preparing an antibacterial surface of polyetheretherketone based on machine learning and femtosecond laser. The basic principle is as Figure 1 shown, and the main steps are as Figure 2 shown, including the following steps:

[0071] I. Establishment of machine learning tools

[0072] 1. Collection of antibacterial data

[0073] In this stage, antibacterial experiment data is mainly extracted from published literature. A total of nearly 100 published literatures are investigated before and after, and available antibacterial data is obtained through 22 of them, including 101 antibacterial surfaces with different materials and structures. Each surface provides at least one photo under scanning electron microscopy (SEM), and antibacterial experiment data for at least one of two bacteria, Staphylococcus aureus (S. aureus, a representative of Gram-positive bacteria) or Escherichia coli (E. coli, a representative of Gram-negative bacteria). Among them, there are 82 antibacterial experiments for Staphylococcus aureus and 76 antibacterial experiments for Escherichia coli.

[0074] 2. Statistical description of material surface morphology

[0075] The micro-nano structures on the material surface can be detected by a variety of instruments and methods, such as scanning electron microscopy (SEM) and atomic force microscopy (AFM). Among them, AFM can provide true three-dimensional information of the material surface topography. In theory, it can provide higher-dimensional and more detailed data for the study of antibacterial properties. However, the AFM test requires a relatively small surface roughness, and it has not been widely applied to the study of the surface topography of antibacterial materials. In contrast, SEM can only provide quasi-three-dimensional information of the material surface. For the height / depth information of the surface undulations, it can only be approximately represented by the gray scale of the picture. However, the current popularity rate of SEM is relatively high, and in the surveyed literature, the microscopic topography of the material surface is generally given through SEM photos.

[0076] Read the SEM pictures through the Python image processing library PIL (Python Image Library), and convert each SEM picture into a gray scale matrix. The gray scale information and its variation law of the matrix elements reflect the undulation condition and texture information of the material surface. Based on the gray scale matrix, the following several statistical quantities are obtained:

[0077] Variance: That is, the variance of the gray scale values of all pixels in an SEM picture, which reflects the contrast information of the gray scale image and can characterize the uneven degree of the surface undulations of the material.

[0078]

[0079] where m and n are the dimensions of the gray scale matrix, and Q(x, y) is the matrix element in the x-th row and y-th column (corresponding to the gray scale value of the corresponding pixel). is the average gray scale value of all pixels, which reflects the average brightness of the gray scale image.

[0080] One-dimensional information entropy: In information theory, "self-information" reflects the amount of information contained when an event occurs in a certain way, and information entropy is the mathematical expectation of the self-information when an event occurs in all possible forms, that is, the average amount of information contained in the event. In image processing, the one-dimensional information entropy of a gray scale image is defined as

[0081]

[0082] Also known as image entropy, it reflects the richness and complexity of the amount of information contained in an image. In the formula, f(i) represents the proportion of pixels with gray scale value i: #(i) represents the number of pixels with gray scale value i, m and n are the dimensions of the gray scale matrix, and the value range of i is 0 to 255.

[0083] The above statistics are global statistical descriptions for a grayscale image and do not contain the relative position information between pixel points. Therefore, they cannot accurately reflect the texture topography of the material surface. Usually, it is also necessary to use the Grey Level Co-occurrence Matrix (GLCM) to obtain statistics that can contain more texture information. The definition of GLCM is as follows: The value of the matrix element in the i-th row and j-th column of it is equal to the total number of times a pixel point with a gray value of j is found starting from any pixel point with a gray value of i in the image, along a certain direction (angle θ) and separated by a fixed distance (denoted as D). By traversing all possible values of i and j (the value ranges of i and j are both 0 to 255), the obtained result can be expressed in the form of a matrix, that is, the grey level co-occurrence matrix GLCM. Specifically,

[0084]

[0085] Among them

[0086] In practical applications, polar coordinates (D, θ) are often used instead of rectangular coordinates (Δx, Δy), and the dimension of the obtained GLCM matrix is usually 256×256.

[0087] Based on GLCM, the statistics describing texture information include the following:

[0088] Energy, also known as Angular Second Moment (ASM), is a measure of the uniformity of the image gray level distribution and the fineness of the texture.

[0089]

[0090] Contrast reflects the gray level difference between local pixel pairs in the image and affects the visual clarity of the image.

[0091]

[0092] Correlation reflects the proportion of elements with different gray values in the image and the uniformity of their distribution.

[0093]

[0094] Among them are the average value and variance of the elements in the i-th row (or j-th column) of GCLM respectively.

[0095] Homogeneity reflects the clarity and regularity of the texture.

[0096]

[0097] For each grayscale matrix converted from an SEM grayscale image, the Python image processing library scikit-image is imported, and the gray-level co-occurrence matrix GLCM is calculated through the graycomatrix function. Then, the above four statistical quantities are calculated respectively through the graycoprops function. During the calculation process, the polar coordinate parameters (D, θ) are used to replace the rectangular coordinate parameters (Δx, Δy). Since the typical microscale involved in surface structure antibacterial is generally between a few hundred nanometers and a few micrometers. For each SEM image, all possible D values with the corresponding pixel point spacing between 0.1 μm and 2 μm are taken to calculate the GLCM matrix at different spatial scales. And at the same D value, 4 different spatial directions (corresponding to θ = 0°, 45°, 90° and 180° respectively) are taken to calculate 4 different GLCMs, and the obtained statistical quantities are averaged respectively. In this way, about a dozen GLCM statistical quantities at different spatial scales are calculated for each SEM image (since the sizes and resolutions of different images are different, the GLCM at some scales for some images may not be calculable). Adding the two grayscale matrix statistical quantities described above and using the antibacterial rate values of the corresponding surface against the two bacteria as labels, a high-dimensional data set required for the machine learning model is constructed.

[0098] 3. Selection and Training of Machine Learning Tools

[0099] First, five commonly used machine learning models were selected to attempt to classify the antibacterial surfaces in the dataset, including Support Vector Classifier (SVC), Naive Bayes classifier, Gradient Boosting (GBoost), Adaptive Boosting (AdaBoost), and Random Forest classifier (RF). Among them, SVC and Bayes belong to classical single-classification learning algorithms, which use only one basic learner for modeling and prediction at a time. While GBoost, AdaBoost, and RF belong to ensemble learning algorithms, which train multiple basic learners simultaneously and then combine them into a more powerful ensemble model in different ways. In addition, a deep learning algorithm, namely Multilayer Perceptron (MLP), was preliminarily attempted. It uses a structure similar to a biological neural network to simulate the complex associations between data and completes prediction and training through the feedforward of data and the feedback of errors. The above six machine learning models are respectively suitable for different datasets and application scenarios. First, the most suitable classification model for the antibacterial surface dataset needs to be selected through testing.

[0100] The test was carried out as follows: for each machine learning model, in 50 independent processes respectively, using the train_test_split tool of the Scikit-learn library, the dataset was randomly divided into a training set and a test set; the training set was used to train the model, and the test set was used to evaluate the performance and generalization ability of the trained model. According to the dataset size, 25% of the test set data was selected from the entire dataset. Then, in each independent process, through 10-fold cross-validation, the GridSearchCV tool was used to perform a grid search on the hyperparameter space to determine the most suitable hyperparameter combination for the current task for each model. Then, the model was trained with the optimal hyperparameter combination and the test set data was predicted. Finally, according to the existing classification labels, the generalization performance of the model on the test set was evaluated, including: accuracy, precision, and recall, etc. The average value of each performance metric was taken in 50 independent processes.

[0101] In addition, the performance of model classification is directly related to the classification threshold; for each classification task, a definite antibacterial rate threshold must be given, and then the model can divide the dataset into two groups: the antibacterial rate is greater than or equal to the given threshold, and the antibacterial rate is less than the given threshold. Since there are different requirements for the antibacterial rate of materials in different actual application scenarios and there is no unified standard. Therefore, during the testing process, 9 different classification thresholds of 0.1, 0.2, 0.3, …… 0.9 were taken respectively to test the performance of each machine learning model at each threshold, so as to better evaluate the application value of machine learning methods in different scenarios.

[0102] Figure 3 The classification results of the above 6 machine learning models on the antibacterial dataset are shown; the antibacterial data of two kinds of bacteria, Staphylococcus aureus (S.aureus) and Escherichia coli (E.coli), were classified respectively. Each classification task was to take the average classification accuracy after 50 independent processes at different antibacterial thresholds. As can be seen from the figure, at most classification thresholds, the average accuracy of the three ensemble learning algorithms (GBoost, AdaBoost and RF) is significantly higher than that of the single learning algorithms (SVC, Bayes and MLP).

[0103] As shown in the above figure, when the single learning algorithm fits the complex relationship between the surface structure and antibacterial performance, especially when the quantity and quality of the data are limited, the simulation ability is slightly insufficient; while the ensemble learning algorithm has obvious advantages in dealing with the complex association problems of small datasets. Among them, the Random Forest (RF) algorithm shows better performance than the other two ensemble learning algorithms.

[0104] Therefore, in this embodiment, the Random Forest (RF) algorithm is preferentially selected, and the final determined machine learning process is as Figure 7 shown.

[0105] II. Prediction Analysis

[0106] 1. Obtaining of polyetheretherketone surfaces with different morphologies: The samples were ultrasonically cleaned with ethanol and ultrapure water for 5 minutes each time; then the samples were placed in a vacuum dryer and dried at 60 °C for 20 minutes; the laser beam (104 fs, 1 kHz, 800 nm) of a regenerative amplified Ti:sapphire femtosecond laser system (Solstice Ace, spectrum-physics) was used to fabricate hierarchical structures through a scanning galvanometer (IntelliSCAN III 14, SCANLAB).

[0107] In this embodiment, a linear scanning method is adopted to establish a multi-level micro-nano composite structure:

[0108] The laser scanning speed is 10 - 15 mm / s, the line spacing is 10 - 15 μm, and the number of laser scans is 1 - 2 times. By changing the laser energy density (0 - 4 J / cm 2 ), different hierarchical structures are prepared on the surface of polyetheretherketone and their properties are evaluated, as Figure 4 shown:

[0109] The ablation threshold of PEEK is about 0.5 J / cm 2 . Gradually increasing the laser energy density from 0 to 1 J / cm 2 , shallow micro-pits begin to form irregularly. When the laser energy density increases from 1 J / cm 2 to 2 J / cm 2 , the formation of LIPSS gradually dominates the surface. Subsequently, when the laser energy density increases to 3 J / cm 2 , the LIPSS structure transforms into a dominant pore-like structure with a diameter range from 100 nm to 2 μm. Further increasing the laser energy density to 4 J / cm 2 results in the formation of large-sized melted pits, accompanied by large-sized pore clusters (≥5 μm) as the main structure. With the increase of the laser energy density, the surface roughness of PEEK gradually increases. For convenience, we classified the surfaces according to the observed dominant structures. Surfaces mainly with LIPSS, while having fewer pits and micropore clusters, are labeled as FS-A; surfaces mainly with micro-pits and pore clusters are labeled as FS-B; surfaces mainly with slightly larger pits (2 - 4 μm) are labeled as FS-C. It is worth noting that in FS-A, the depth of the LIPSS structure is about 100 - 200 nm, and the average surface roughness ranges from 30 - 45 nm, which is significantly higher than the surface roughness of untreated PEEK (6 nm). Compared with FS-A (≈45 nm), the surface roughness of FS-B (≈120 nm) and FS-C (≈170 nm) further increases, indicating a positive correlation between the surface roughness and the laser energy density within a certain range. After femtosecond laser modification, no new elements were observed to be introduced. However, the carbon-oxygen ratio changed at different energy densities. With the increase of the laser energy density, the carbon atom ratio initially increases and then decreases within a small range. The initial increase in the carbon atom ratio may be attributed to the carbonization during the laser irradiation process, resulting in the breaking and rearrangement of the molecular structure on the polymer surface, releasing carbon atoms and increasing the surface carbon content. However, with the increase of the laser energy density, the oxidation reaction gradually dominates, leading to an increase in the O / C ratio in the irradiated area. Wettability can affect various processes related to bacterial infection, such as adsorption, wetting, adhesion, friction, and lubrication. The wettability was tested using an optical contact angle measurement system.

[0110] The results show that with the increase of the laser energy density, the water contact angle of the treated area gradually decreases. Compared with the contact angle (≥85°) of the untreated PEEK surface, the contact angle of the PEEK surface treated by femtosecond laser decreases to 50 - 75°. To explore the reason for the change in hydrophobicity, XPS analysis was further carried out and it was found that femtosecond laser modification introduced hydrophilic functional groups of C-OH to the surface of polyetheretherketone. This may be the main reason for the hydrophilic characteristics presented by the PEEK surface after femtosecond laser treatment.

[0111] In the Wenzel model, it is assumed that the liquid completely penetrates into the rough surface, and the apparent contact angle θ* is given by the following formula:

[0112] cosθ * =rcosθ

[0113] where r is the roughness ratio (1 for an ideally smooth surface), and θ is the Young's contact angle. Since r is always greater than 1, this model predicts that for a surface with a liquid contact angle θ < 90°, as the surface roughness increases, the surface contact angle decreases (θ* < θ), and the hydrophilicity increases. For a non-wetting surface (θ > 90°) when it becomes rough, the surface hydrophobicity increases with the increase of roughness (θ* > θ).

[0114] Conversely, in the Cassie and Baxter model, it is assumed that the liquid cannot completely penetrate the rough surface. Therefore, air bubbles are trapped under the liquid, and the liquid is located above the composite surface composed of solid and air. In this case, the contact angle θ*

[0115] is the average value between the value on the air (i.e., 180°) and the value on the plane (i.e., θ), and is given by the following formula:

[0116] cosθ * =-1 + f(1 + cosθ)

[0117] where f is the proportion of the solid surface in contact with the liquid in the projected area. Since f is always less than 1, this model always predicts an increase in hydrophobicity, regardless of the value of the Young's contact angle θ. The lower the f value, the smaller the solid-liquid contact area and the larger the measured contact angle.

[0118] Through femtosecond laser treatment, hydrophilic functional groups were introduced to the surface, increasing the surface energy of the PEEK surface, which is manifested as the wettability gradually improving with the increase of the laser energy density. There is no obvious groove structure on the surface, so no gas-liquid interface sufficient to change the wettability is formed, thus ensuring hydrophilicity.

[0119] 2. Acquisition of surface images of polyetheretherketone with different morphologies: The surface morphology was obtained by a field emission scanning electron microscope (SEM) (Ultra 55, Carl Zeiss NTS GmbH, Germany), and then the surface morphology of the sample was further analyzed by a white light interference confocal laser scanning microscope (CLSM, Olympus Lext OLS4000). Finally, the surface roughness (Sa), kurtosis (Sku), and height-width ratio (Str) of the surface features of the material were characterized by an atomic force microscope (Seiko (Japan) model: SPI3800N) on a surface area of 20×20 μm. 2 was measured.

[0120] 3. Machine learning analysis of surface images of polyetheretherketone with different morphologies: Given the scanning electron microscope (SEM) images of the material surface, it was predicted by machine learning tools that FS-A had good antibacterial effects.

[0121] 4. Antibacterial performance experimental evaluation: Antibacterial performance was evaluated by the plate coating technique.

[0122] Culture medium preparation:

[0123] LB liquid medium - Measure 100 mL of distilled water with a measuring cylinder and pour it into a 250 mL reagent bottle. Weigh 2.5 g of LB broth medium with an analytical electronic balance and add it to the mixture. After mixing, sterilize it in a high-temperature and high-pressure steam sterilizer at 121 °C for 15 min and set aside for use. LB solid medium - Measure 100 mL of distilled water with a measuring cylinder and pour it into a 250 mL reagent bottle. Weigh 2.5 g of LB broth medium and 1.5 g of agar powder with an analytical electronic balance. Add the weighed reagents to the mixture and mix well. Then sterilize it in a high-temperature and high-pressure steam sterilizer at 121 °C for 15 min. Wait for the medium to cool to about 40 - 50 °C, and use an electric pipette to suck 15 mL of the medium and pour it into a disposable sterile petri dish.

[0124] 1 / 500 NB aqueous solution - Measure 100 mL of distilled water with a measuring cylinder and pour it into a 250 mL reagent bottle. Weigh 1.8 g of nutrient broth medium with an analytical electronic balance and add it to the mixture to dissolve. Dilute the nutrient broth 500 times with deionized water and sterilize it in a high-temperature and high-pressure steam sterilizer at 115 °C for 30 min for standby.

[0125] Bacterial suspension preparation:

[0126] Take three 12 mL bacterial culture tubes, each add 3 mL of LB liquid medium, pick single colonies from the solid media of Staphylococcus aureus and Escherichia coli and add them to the liquid medium respectively, and use another tube as a blank control. Place them in a constant temperature oscillator (37 °C, 200 rpm) and shake and culture overnight (15 h).

[0127] Treatment of the sample to be tested:

[0128] Put the samples into 50 mL centrifuge tubes according to the grouping, and sterilize them at 121 °C for 20 min in a high-temperature and high-pressure steam sterilizer for later use.

[0129] Co-culture + plate coating counting test:

[0130] Dilute the bacterial solution to 10 5 CFU / mL with 1 / 500 NB aqueous solution, and evenly drop 23 μL of the bacterial solution on the surface of the sample. Place it in a constant temperature incubator at 37 °C and let it stand for 24 h. After the incubation is completed, elute the bacterial solution with 2.3 mL of sterile PBS solution, and make serial 10-fold dilutions of the eluate with sterile PBS solution. Take 100 μL of the diluted solution and evenly coat it on the LB solid medium. Place it in a constant temperature incubator at 37 °C for 18 h, take it out, take pictures and record the number of colonies.

[0131] Through the above operations, the results as shown in Figure 5 , Figure 6 are obtained:

[0132] In the figure, the abscissa is the predicted antibacterial rate value r, and the ordinate is predicted by the random forest classification algorithm, which is the probability that the antibacterial rate of the given surface is greater than this value. The ideal classification result should produce a sharp drop in probability near the actual antibacterial rate value. Therefore, select the abscissa value corresponding to the intersection of the prediction curve and the horizontal dotted line with a probability equal to 0.5 as the predicted antibacterial rate value.

[0133] In the figure, for the same sample, three pictures with magnifications of 1000x, 5000x, and 25000x are taken at different magnifications and are respectively used for prediction.

[0134] It is shown by Figure 5 that the antibacterial rate of FS-A against Escherichia coli reaches 92.3%;

[0135] It is shown by Figure 6 that the antibacterial rate of FS-A against Staphylococcus aureus reaches 77.8%.

[0136] Through multiple repeated verifications, the accuracy rate of the model prediction is about between 70% and 80%, which proves the feasibility of predicting the antibacterial rate of materials through surface structure pictures.

[0137] Through multiple repeated verifications, the best antibacterial structure on the surface of polyetheretherketone is a depth of 100 - 200 nm and a surface average roughness of 30 - 45 nm.

[0138] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predicting and preparing an antibacterial surface of polyetheretherketone based on machine learning and femtosecond laser, characterized in that: The method includes the following steps: S1. Establish a machine learning tool: Obtain the surface structure information and corresponding antibacterial performance data of known antibacterial materials, and use the surface structure information and corresponding antibacterial performance data of the antibacterial materials to select and train a machine learning model. The surface structure information of the antibacterial materials includes statistics based on quasi-three-dimensional and true three-dimensional topography information, statistics of surface texture information expressed by a gray-level co-occurrence matrix, and these statistics form a high-dimensional antibacterial data set; S2. Prepare the polyetheretherketone surface: Obtain polyetheretherketone surfaces with different topographies and roughnesses by linearly adjusting the energy density of laser pulses; S3. Obtain the structure information of the polyetheretherketone surfaces with different topographies, including: 1) Use an emission scanning electron microscope to obtain the quasi-three-dimensional topography information of the surface, where the third-dimensional information, that is, the protrusion height or depression depth of the material surface, is reflected by the gray value of the generated pixel points; 2) Use an atomic force microscope to obtain the true three-dimensional topography information of the surface, where the third-dimensional information, that is, the protrusion height or depression depth information of the material surface, is reflected by the real height value or depth value; S4. Prediction and analysis: According to the three-dimensional structure information of the material surface obtained in step 3, use the machine learning tool for classification prediction, predict the antibacterial ability of the polyetheretherketone surfaces with different topographies, and select the optimal antibacterial structure; S5. Through actual antibacterial experiments, verify the antibacterial performance of the polyetheretherketone surfaces with different topographies, and test the accuracy of the prediction by the machine learning method.

2. The prediction and preparation method of a polyetheretherketone antibacterial surface based on machine learning and femtosecond laser according to claim 1, characterized in that: The process of obtaining the quasi-three-dimensional topography information of the material surface includes reading each SEM image to generate a two-dimensional gray matrix, and the matrix elements take discrete values; the process of obtaining the true three-dimensional topography information of the material surface includes reading each AFM image to generate a two-dimensional matrix describing the surface undulation information, and the matrix elements take continuous values. Based on these two types of two-dimensional matrices, the specific statistics are: 1) Variance: The variance of all matrix element values in the matrix, which reflects the contrast information of the material surface and characterizes the unevenness of the material surface undulation. where m and n are the dimensions of the matrix; Q(x, y) is the matrix element value at the x-th row and y-th column of the matrix, corresponding to the gray value of the pixel at the corresponding position in the SEM image, or the actual height value of the corresponding point in the AFM image; is the average value of all matrix elements, reflecting the average brightness of each point in the SEM image, or the average height of each point in the AFM image; 2) One-dimensional information entropy: Reflects the richness and complexity of the information contained in an image; The one-dimensional information entropy calculated based on the SEM gray image information of the quasi-three-dimensional topography information is: In the formula, f(i) represents the proportion of pixels with gray value i: #(i) represents the number of pixels with gray value i. m and n are the dimensions of the gray matrix. In an 8-bit gray image, the value range of i is an integer between 0 and 255.

3. The prediction and preparation method of a polyetheretherketone antibacterial surface based on machine learning and femtosecond laser according to claim 2, characterized in that: The statistics describing the surface texture information of the material also include four statistics of energy, contrast, correlation, and uniformity obtained through the gray-level co-occurrence matrix. The specific expressions are: 1) Gray-level co-occurrence matrix: The gray-level co-occurrence matrix generated based on the SEM gray image of the quasi-three-dimensional topography information is defined as follows Among them In an 8-bit gray image, the maximum dimension of the obtained gray-level co-occurrence matrix is 256×256; To generate a gray-level co-occurrence matrix based on the AFM image of the true three-dimensional information, it is necessary to first discretize the values of the matrix elements; The characteristic scale of the material surface related to antibacterial is usually between 0.1μm and 2μm. The gray-level co-occurrence matrices at different spatial scales are calculated by taking all possible pixel point spacings between 0.1μm and 2μm for both the SEM image and the AFM image respectively; At the same spatial scale, four different gray-level co-occurrence matrices are calculated in four different spatial directions: horizontal, vertical, 45-degree left, and 45-degree right. The statistics obtained are averaged to obtain the most general results. 2) Statistics describing surface texture information: ① Energy: ② Contrast ③ Correlation where μ i , μ j , are the mean and variance of the elements in the i-th row or j-th column of the gray-level co-occurrence matrix, respectively. ④ Homogeneity 4. The prediction and preparation method of a polyetheretherketone antibacterial surface based on machine learning and femtosecond laser according to claim 1, characterized in that: The machine learning models used include support vector machine classifier, naive Bayes classifier, gradient boosting algorithm, adaptive boosting algorithm, random forest classifier, and multi-layer perceptron. Test steps: For each machine learning model, in 50 independent processes respectively, using the train_test_split tool of the Scikit-learn library, the dataset is randomly divided into a training set and a test set. The training set is used to train the model, and the test set is used to evaluate the performance and generalization ability of the trained model; according to the scale of the collected dataset, 25% of the test set data is selected from the entire dataset. In each independent process, first through 10-fold cross-validation, the GridSearchCV tool is used to perform grid search on the hyperparameter space to determine the most suitable hyperparameter combination for the current task for each model; then the selected optimal hyperparameter combination is used to train the final model and predict the test set data. Finally, based on the existing classification labels, the performance of the model on the test set is evaluated, including: accuracy, precision, and recall; each performance metric is averaged over 50 independent processes.

5. The prediction and preparation method of a polyetheretherketone antibacterial surface based on machine learning and femtosecond laser according to claim 4, characterized in that: During the test, the classification threshold is incremented by 0.1, and the classification threshold range is 0.1 - 0.

9.

6. The prediction and preparation method of a polyetheretherketone antibacterial surface based on machine learning and femtosecond laser according to claim 1, characterized in that: In step 1, the change range of the laser energy density is 0 to 4 J / cm 2 .

7. A prediction and preparation method of a polyetheretherketone antibacterial surface based on machine learning and femtosecond laser according to any one of claims 1-6, characterized in that: When preparing the surface, a linear scanning method is used to establish a multi-level micro-nano composite structure, with a scanning line spacing of 13 - 18 μm, a scanning speed of 12 - 16 mm / s, a spot size of 26 - 35 μm, and 2 - 5 pulses in each spot irradiation area.

8. A prediction and preparation method for a polyetheretherketone antibacterial surface based on machine learning and femtosecond laser according to any one of claims 1 - 6, characterized in that: The surface area for obtaining three-dimensional information is 20×20 μm 2 .

Images