A superhard high-entropy nitride coating designed based on machine learning and its preparation method

Abstract

The present invention discloses an ultra-hard high-entropy nitride coating designed based on machine learning and a preparation method thereof, belonging to the technical field of plating of metal materials. The present invention comprises the following steps: (1) establishing a data set for predicting the hardness of a high-entropy nitride coating; (2) evaluating various features in the database through feature engineering, optimizing the features to improve prediction accuracy; (3) performing hyperparameter optimization and ten-fold cross-validation on the machine learning model to select the model with the best prediction performance; (4) inputting a preset prediction space into the machine learning model with the best prediction performance to select the high-entropy nitride coating system with the highest predicted hardness; (5) preparing an ultra-hard high-entropy nitride coating based on the high-entropy nitride coating with the highest predicted hardness. The present invention utilizes the advantages of machine learning in terms of high efficiency, low cost, and the ability to process large-scale complex systems, providing important scientific insights and effective design methods for the development of advanced protective coatings.

Description

Technical Field

[0001] The present invention relates to the technical field of plating of metal materials, and in particular to an ultra-hard high-entropy nitride coating designed based on machine learning and a preparation method thereof. Background Art

[0002] Friction damage permeates modern industrial society, posing a persistent threat to the performance and service life of materials, further leading to the frequent replacement of mechanical components and the gradual aging of heavy equipment. This not only significantly reduces production efficiency but also causes huge energy losses, running counter to the environmental goals of sustainable development. It is reported that energy losses caused by frictional contact alone account for 23% of the world's total energy consumption each year, reaching a staggering 119 EJ. The development of various advanced wear-resistant materials and the application of surface engineering are expected to reduce energy losses by approximately 40% in the next decade. Transition metal nitride (TMN) coatings have excellent mechanical properties and are widely used in various extreme wear environments. However, traditional nitride coatings face increasing challenges in meeting the new demands brought about by rapidly developing industrial applications.

[0003] High-entropy alloys (HEAs), also known as multi-principalelement alloys (MPEAs), are a new type of alloy system proposed almost simultaneously by Taiwanese scholar Ye Junwei and British scholar Cantor in 2004. Over the past two decades, they have attracted widespread attention in the academic community for their exceptional performance. With the rapid advancement of coating preparation technology, the concept of HEAs has been applied to nitride coatings, resulting in the development of high-entropy nitride (HEN) coatings. Leveraging the nearly unlimited compositional design space of HEAs, HEN coatings have the potential to surpass the performance limitations of traditional coating materials. Trial-and-error is the primary approach for HEN coating design, but it is inefficient and resource-intensive. With the rise of computational materials science, methods such as molecular dynamics (MD) and density functional theory (DFT) have accelerated the discovery of new HEN coatings. However, the high computational cost and relatively low efficiency of these methods have limited the further development of high-performance HEN coatings.

[0004] With the launch of the Materials Genome Initiative (MGI) project ten years ago, emerging data-driven models led by machine learning have been applied to high-throughput screening of various advanced materials, greatly reducing the computational and experimental costs in the process of discovering new materials. In the field of high-entropy materials, Wen et al. optimized the hardness improvement problem of the AlCoCrCuFeNi system by establishing a machine learning model that includes composition, physical descriptors, and hardness. Huang et al. developed a machine learning model for the design of ultra-high hardness solid solution hardening (SSH) HEAs. Li et al. effectively reduced the hardness prediction error of the AlCoCrCuFeNi system by optimizing the genetic algorithm of machine learning.

[0005] Limited by the inefficiency of traditional trial-and-error methods and the unlimited compositional design space of HEAs, accelerating the discovery of high-performance HEN coatings remains a daunting challenge. Existing methods for coating preparation rely on other learning approaches to design parameter indicators. For example, Chinese invention patent publication number CN118070674A discloses a method and system for predicting the anti-ablation performance of ceramic coatings based on machine learning. The method includes: establishing a dataset; establishing a preliminary optimal prediction model; comparing the prediction accuracy of the preliminary optimal prediction model with preset requirements; when the preset requirements are met, the preliminary optimal prediction model becomes the model for predicting the anti-ablation performance of the ceramic coating; when the preset requirements are not met, analyzing the feature importance of the preliminary optimal prediction model, constructing new features and re-establishing the optimal prediction model, evaluating the prediction accuracy of the optimal prediction model, and comparing the preset requirements until the preset requirements are met to obtain a model for predicting the anti-ablation performance of the ceramic coating; and inputting features and obtaining predicted ablation results based on the predicted anti-ablation performance model of the ceramic coating. This invention predicts anti-ablation performance through machine learning methods. However, when using machine learning as a tool and designing within its framework, the prediction accuracy of the indicators is highly affected by the type of simulation parameters selected. In the design and preparation of ultra-hard high-entropy nitride coatings, the selection and adjustment of simulation parameters remain a huge challenge for machine learning.

[0006] Currently, the application of machine learning is primarily focused on predicting the hardness and phase structure of bulk HEAs, while research on the performance of HEN coatings, which are significantly influenced by different fabrication parameters, is limited. Therefore, to further improve computational efficiency and conserve resources, using machine learning to guide the design of HEN coatings and produce protective coatings with excellent mechanical properties has important scientific significance and application value. Summary of the Invention

[0007] In view of the above-mentioned defects of the prior art, in a first aspect of the present invention, a method for preparing an ultra-hard high-entropy nitride coating based on machine learning design is provided, which is efficient, accurate, and has good applicability, and comprises the following steps:

[0008] (1) Dataset establishment:

[0009] A data set for predicting the hardness of high-entropy nitride coatings was established. The characteristics of the data set included the composition characteristics, physical descriptor characteristics, process parameter characteristics, and hardness characteristics of the high-entropy nitride coatings. The composition characteristics included the atomic percentages of Al, Cr, Nb, Si, Ta, Ti, V, Zr, and N. The physical descriptor characteristics included the mixing entropy (ΔS mix ), mixing enthalpy (ΔH mix )、Mixed melting point(T m ), entropy-enthalpy ratio (Ω), atomic size difference (δ), valence electron concentration (VEC), electronegativity difference (Δχ); process parameter characteristics include substrate temperature, substrate bias, deposition gas pressure, argon-nitrogen ratio (Ar / N2);

[0010] The dataset is divided into subsets A to D; subset A includes composition features and hardness features; subset B includes composition features, physical descriptor features, and hardness features; subset C includes composition features, process parameter features, and hardness features; subset D includes composition features, physical descriptor features, process parameter features, and hardness features;

[0011] (2) Feature Engineering:

[0012] Evaluate the features of the dataset through feature engineering, analyze their relevance and importance ranking, and optimize the features to improve prediction accuracy;

[0013] (3) Machine Learning Modeling:

[0014] The subsets AD of the dataset were used as input values, and the predicted hardness was used as output values. A set proportion of the subset AD in the dataset was used as the test set, and the rest was used as the training set. A machine learning model was performed on the subset AD using the set algorithm and test set proportion. The hyperparameters of the constructed machine learning model were optimized and ten-fold cross-validated to evaluate its prediction performance. The machine learning model with the best prediction performance was obtained.

[0015] (4) Ingredient optimization:

[0016] Inputting the preset prediction space into the machine learning model with the best prediction performance, selecting the high-entropy nitride coating with the highest predicted hardness formed with nitrogen from each multi-element high-entropy alloy system; wherein the multi-element high-entropy alloy system is a five-element to eight-element high-entropy alloy system formed by at least five of Al, Cr, Nb, Si, Ta, Ti, V, and Zr;

[0017] (5) Coating preparation:

[0018] According to the predicted high-entropy nitride coating with the highest hardness, the corresponding multi-component high-entropy alloy target is selected, and an ultra-hard high-entropy nitride coating is prepared under the corresponding process parameter characteristics.

[0019] The hardness characteristic described in this method or the parameter index for predicting hardness is nanohardness, and its unit is GPa.

[0020] Based on the above technical scheme, the inventive concept of the present invention is to design and select three features containing 20 parameters according to the preparation characteristics of high-entropy nitride coatings, and combine the features including at least composition characteristics to form four specific subsets. By utilizing the high computing efficiency and low computing cost of machine learning and the ability to process large-scale complex systems, ultra-hard high-entropy nitride coatings whose performance is greatly affected by complex preparation process parameters are predicted and prepared, so as to save computing costs, improve computing efficiency, and accelerate the discovery of new high-performance high-entropy nitride coatings.

[0021] Preferably, in step (1), the data set includes 47 literature data on high entropy nitride coatings starting from 2006 and 17 experimental data conducted in the laboratory.

[0022] Dividing the dataset into four subsets is beneficial for fully studying the impact of different types of features on prediction performance. In the actual machine learning design process, the step of dividing the dataset into subsets AD can be performed in step (1) or step (3). This step is performed before machine learning modeling, and the final input values ​​of subsets AD remain consistent.

[0023] Preferably, in step (2), the feature engineering includes using the Pearson correlation coefficient (PCC) to analyze the correlation between each feature, using the random forest (RF) algorithm to analyze the importance ranking between features, and eliminating redundant features based on the importance ranking; redundant features are features that rank lower in feature pairs with a Pearson correlation coefficient value greater than 0.95.

[0024] Preferably, in step (3), the type of algorithm includes at least one of extra tree (ET), gradient boosting decision tree (GBDT), random forest (RF), support vector regression (SVR), and extreme gradient boosting (XGB).

[0025] Preferably, in step (3), the test set ratio includes 10%, 20%, 30%, 40%, and 50%.

[0026] Preferably, in step (3), the hyperparameter optimization method is Bayesian optimization.

[0027] Preferably, in step (3), the indicator of prediction performance is prediction accuracy, and the prediction accuracy is evaluated based on the root mean squared error (RMSE) of the test set. The machine learning model with the best prediction performance is the machine learning model with the smallest root mean squared error of the test set.

[0028] Preferably, in step (4), in the preset prediction space, the composition interval of the composition feature is established by the lower limit and upper limit of the atomic percentage of each element in the data set, and the sum of the atomic percentage of each element is 100 at.%; the parameter value of the process parameter feature is selected from the three most frequently occurring values ​​in the data set.

[0029] Further preferably, the composition intervals of the composition characteristics are as follows: Al 0 at.%-18 at.%, Cr 4 at.%-28 at.%, Nb 0 at.%-10 at.%, Si 0 at.%-12 at.%, Ta 0 at.%-26 at.%, Ti 0 at.%-20 at.%, V 0 at.%-12 at.%, Zr 0 at.%-12 at.%, N 40 at.%-60 at.%, and the change step size of each interval is 2 at.%; the parameter values ​​of the process parameter characteristics are as follows: substrate temperature is selected from 200 ℃, 300 ℃, 400 ℃; substrate bias is selected from -100 V, -150 V, -200 V; deposition gas pressure is selected from 0.5 Pa, 0.75 Pa, 1.0 Pa; argon-nitrogen ratio is selected from 1, 1.5, 2.

[0030] Preferably, in step (5), the characteristics of the multi-element high entropy alloy target and process parameters corresponding to the high entropy nitride coating with the highest hardness are predicted as follows: For the five-element high entropy alloy system, the target is Al10 Cr 48 Nb 10 Si 12 Ti 20 , substrate temperature is 200 ℃, substrate bias is -100 V, deposition pressure is 0.75 Pa, argon nitrogen ratio is 1.5; for the hexavalent high entropy alloy system, the target material is Al 15 Cr 35 Nb 10 Si 15 Ti 20 V5, substrate temperature is 200 ℃, substrate bias is -100 V, deposition pressure is 1.0 Pa, argon nitrogen ratio is 2; for the seven-element high entropy alloy system, the target material is Al 14 Cr 33 Nb 10 Si 14 Ta5Ti 19 V5, substrate temperature is 200 ℃, substrate bias is -100 V, deposition pressure is 0.75 Pa, argon nitrogen ratio is 1.5; for the octal high entropy alloy system, the target material is Al 10 Cr 32 Nb 10 Si 14 Ta5Ti 19 V5Zr5, substrate temperature is 200 ℃, substrate bias is -100 V, deposition gas pressure is 1.0 Pa, argon-nitrogen ratio is 2; the suffix value of the element in the target material represents the atomic percentage value of the element.

[0031] In a second aspect of the present invention, there is provided an ultra-hard high-entropy nitride coating with high hardness and excellent mechanical properties, which is prepared by the method of the first aspect of the present invention.

[0032] This method uses the scientific principle of design-guided experimentation to address the design difficulties posed by the unlimited compositional space of high-entropy materials, avoids the waste of resources caused by trial-and-error methods, and improves the efficiency of developing high-performance high-entropy nitride coatings. This method has good universality. Simply based on the performance requirements of the target high-entropy material, a suitable database can be established, feature engineering and model optimization can be performed, and ultimately, the optimal model can be used to guide the preparation of high-entropy materials for different applications.

[0033] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0034] The present invention provides a preparation method of an ultrahard high-entropy nitride coating based on machine learning design, which has the advantages of high efficiency, accuracy, and good applicability, and improves the efficiency of developing high-performance high-entropy nitride coatings.

[0035] The present invention provides a superhard high-entropy nitride coating with high hardness, excellent mechanical properties and good application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the RF magnetron sputtering equipment used in coating preparation; including: 1. Vacuum system; 2. Heater; 3. Ventilation duct; 4. Sample holder; 5. RF power supply; 6. High entropy alloy target; 7. Arc power supply; 8. Cr target; 9. Baffle;

[0037] Figure 2 A schematic diagram of the process for preparing ultra-hard high-entropy nitride coatings designed based on machine learning;

[0038] Figure 3 Pearson correlation heatmap between all features in machine learning modeling;

[0039] Figure 4 Build a ranking chart of the importance of all features in machine learning modeling;

[0040] Figure 5 Hyperparameter optimization results for machine learning modeling; (a)-(d) correspond to the optimization results under subsets AD respectively;

[0041] Figure 6 This is a graph of the ten-fold cross-validation results based on hyperparameter optimization results in machine learning modeling;

[0042] Figure 7 A comparison chart of the hardness of the ultra-hard high-entropy nitride coatings prepared in Examples 1-4 and the coatings in the database;

[0043] Figure 8 Surface and cross-sectional morphologies of the ultrahard high-entropy nitride coatings prepared in Examples 1-4; (a)-(d) correspond to the test results of Examples 1-4, respectively;

[0044] Figure 9 X-ray diffraction (XRD) patterns of the ultrahard high entropy nitride coatings prepared in Examples 1-4;

[0045] Figure 10 This is a high-resolution transmission electron microscope (HRTEM) image of the ultrahard high-entropy nitride coating prepared in Example 1. DETAILED DESCRIPTION

[0046] The present invention is further illustrated by way of examples below, but the present invention is not limited to the scope of the examples. Experimental methods in the following examples where specific conditions are not specified were performed according to conventional methods and conditions, or selected according to the product specifications.

[0047] In the following embodiments:

[0048] Use Figure 1 The RF magnetron sputtering equipment shown is used for coating preparation. The complete deposition process of the coating is as follows:

[0049] According to the prediction results, the corresponding high entropy alloy target is prepared: the powders of each element are mixed according to the corresponding atomic ratio and then sintered;

[0050] Pre-clean the substrate: ultrasonically clean the substrate in acetone, ethanol, and deionized water for 10 minutes each to remove dust, organic matter, and other stubborn contaminants on the substrate surface. Then, dry it in a nitrogen atmosphere and quickly place it in a vacuum chamber to prevent re-contamination.

[0051] Prepare the experimental environment: Place the RF power supply and sputtering target 1, arc power supply and target 2, sample holder 4 and substrate (i.e. sample) in the RF magnetron sputtering equipment in sequence. At this time, the substrate is parallel to the arc power supply target 2, and the sputtering target 1 is perpendicular to the arc power supply target 2. Turn on the cooling system, evacuate the chamber through the vacuum system 6, and turn on the heater 5 at the same time to make the vacuum degree in the chamber (the vacuum degree is generally 10 -3 Pa level) reaches the ideal value and the temperature reaches the preset value;

[0052] Substrate etching: High-purity Ar gas is introduced through the ventilation pipe 7, the substrate negative bias voltage and duty cycle are set, and the arc power supply 2 is turned on. Under the action of the voltage difference, an arc-enhanced glow discharge is formed to ionize Ar into Ar. + ,Ar + The ion beam bombards the substrate under the potential difference, removes the oxide on the substrate surface and activates the surface to increase the bonding strength between the substrate and the coating (in order to protect the substrate from the Cr sputtered from the arc power target 2). + During the etching process, the arc power target 2 is provided with a baffle 3 and the baffle 3 is in a closed state, thereby achieving the effect of Cr + Block);

[0053] Deposition of transition layer: After the substrate etching is completed, the baffle 3 is opened, and Cr is sputtered in the arc power target 2. +A uniform Cr layer is deposited on the substrate to enhance the bonding strength of the hard coating to be deposited; depositing an ultra-hard high-entropy nitride coating: turning off the arc power supply 2, rotating the sample holder 4 90° clockwise so that the substrate is parallel to the sputtering target 1, and setting specific experimental parameters according to the predicted results; turning on the RF power supply 1 and setting the parameters, introducing high-purity argon and high-purity nitrogen with a concentration of 99.99 wt.% into the chamber through the ventilation pipe 7, and under the action of the Ar / N2 mixed flow and the RF source, the material in the sputtering target 1 is sputtered out and deposited on the substrate to form a hard coating uniformly wrapped on the surface of the substrate; cooling system and sample: stopping the film deposition after the experiment reaches the set deposition time, turning off the RF power supply 1, ventilation pipe 7, heater 5, and vacuum system 6 in sequence, and after the temperature drops to room temperature, turning off the cooling system, taking out the sample, and completing the preparation of the hard coating.

[0054] Example 1

[0055] The preparation method of ultra-hard high entropy nitride coating based on machine learning design is as follows Figure 2 As shown, the steps are as follows:

[0056] (1) Dataset establishment:

[0057] By reviewing 47 papers on high-entropy nitride hard coatings from 2006 to date and conducting 17 high-entropy nitride coating preparations in the laboratory, a dataset of 172 high-entropy nitride coating composition characteristics, physical descriptor characteristics, process parameter characteristics, and hardness characteristics was established.

[0058] Among them, the composition characteristics include the atomic percentage of eight strong nitride-forming elements, namely Al, Cr, Nb, Si, Ta, Ti, V, and Zr, as well as the N element; the physical descriptor characteristics include the system's mixing entropy, mixing enthalpy, mixing melting point, entropy-enthalpy ratio, atomic size difference, valence electron concentration, and electronegativity difference; the process parameter characteristics include substrate temperature, substrate bias, deposition gas pressure, and argon-nitrogen ratio during coating deposition;

[0059] The dataset is divided into four subsets A and D, where subset A is composed of composition features and hardness features; subset B is composed of composition features, physical descriptor features and hardness features; subset C is composed of composition features, process parameter features and hardness features; subset D is composed of composition features, physical descriptor features, process parameter features and hardness features;

[0060] (2) Feature Engineering:

[0061] Feature engineering was performed to evaluate and screen various features in the dataset. The Pearson correlation coefficient was used to analyze the correlation between features, and the random forest algorithm was used to analyze the importance ranking of features. Redundant features were eliminated based on the importance ranking, that is, the features ranked lower among the feature pairs with a Pearson correlation coefficient greater than 0.95.

[0062] (3) Machine Learning Modeling:

[0063] Subsets AD were used as input and hardness as output. Five commonly used algorithms for predicting material properties, including extreme tree, gradient boosting decision tree, random forest, support vector regression, and extreme gradient boosting, were used to optimize hyperparameters on four subsets at five test set ratios of 10%, 20%, 30%, 40%, and 50%, respectively. The optimization method used was Bayesian optimization. Based on the hyperparameter optimization results, the models were further subjected to ten-fold cross-validation. The machine learning model with the best prediction accuracy, i.e., the smallest root mean square error value on the test set, was selected based on the root mean square error of the test set.

[0064] (4) Ingredient optimization:

[0065] The preset prediction space is input into the machine learning model with the best performance. In the preset space, the composition range of each element is the lower limit and upper limit in the database, that is, Al 0 at.%-18 at.%, Cr 4 at.%-28 at.%, Nb 0 at.%-10 at.%, Si 0 at.%-12 at.%, Ta 0 at.%-26 at.%, Ti 0 at.%-20 at.%, V 0 at.%-12 at.%, Zr 0 at.%-12 at.%, N 40 at.%-60 at.%, the change step size of each interval is 2 at.%, and the sum of the atomic percentages of each element must be 100 at.%; the process parameter range is the three most frequently appearing values ​​in the database, that is, substrate temperature is selected from 200 ℃, 300 ℃, and 400 ℃; substrate bias is selected from -100 V, -150 V, and -200 V; deposition gas pressure is selected from 0.5 Pa, 0.75 Pa, and 1.0 Pa; argon-nitrogen ratio is selected from 1, 1.5, 2;

[0066] (5) Coating preparation:

[0067] From the prediction results, the five-element (excluding N) high entropy nitride coating with the highest predicted hardness was selected, and the corresponding high entropy alloy (Al 10 Cr 48 Nb 10 Si 12 Ti 20) was used as the target material, and an ultra-hard high-entropy nitride coating (denoted as AlCrNbSiTi according to the target type, the same below) was prepared by sputtering under the corresponding process parameters (substrate temperature of 200 ℃, substrate bias of -100 V, deposition gas pressure of 0.75 Pa, and argon-nitrogen ratio of 1.5).

[0068] Figure 3 This is a heatmap of the Pearson correlations of all features used in the machine learning model, including compositional features, physical descriptor features, and process parameter features. Each box in the heatmap represents the correlation between two different input features, and there are no strongly correlated features within them (PCC > 0.95). The presence of highly correlated features can reduce model stability and potentially lead to overfitting, resulting in poor generalization performance for new data. Therefore, the input features used in this invention can effectively help machine learning models predict the hardness of high-entropy nitride coatings.

[0069] Figure 4 The importance ranking of all features in machine learning modeling is as follows: Since the present invention has selected physical descriptor features and process parameter features that have been proven to be highly correlated with the hardness of high entropy nitride coatings when establishing the data set, Figure 3 The features shown in do not satisfy the redundant feature pair. Therefore, Figure 4 Only the importance between the features is analyzed, such as bias voltage is the most important feature causing hardness change, which is consistent with the current research results of high entropy nitride coatings.

[0070] Figure 5 This is the hyperparameter optimization result for machine learning modeling, using Bayesian optimization. Machine learning modeling was performed simultaneously under three dimensions: 5 algorithms, 5 test set ratios, and 4 subsets. The RMSE of the test set was used to evaluate the accuracy of the model. Figure 5 (a) is the optimization result under subset A, Figure 5 (b) is the optimization result under subset B, Figure 5 (c) is the optimization result under subset C, Figure 5(d) shows the optimization results for subset D. A low RMSE on the test set indicates that the model has good generalization ability and is able to make accurate predictions on new data. Therefore, considering the research objectives of this invention, the inventors believe that the RMSE on the test set is a more effective assessment of the model's predictive accuracy. The RMSE increases with the test set ratio, and a 10% test set ratio achieves the highest accuracy across all algorithms and subsets. For general machine learning models, a large test set ratio provides more reliable and stable evaluation results, better reflecting the model's generalization ability. However, in small datasets, due to insufficient training data, the model may not fully learn the data's features and patterns. Therefore, the inventors analyzed the predictive performance of the test set ratio for four subsets and five algorithms in parallel to facilitate more prudent selection. The model based on the RF algorithm achieved the lowest RMSE for almost all subsets and test set ratios. XGB and GBDT generally have good predictive performance, but perform poorly on small datasets. Furthermore, GBDT is sensitive to outliers in the dataset. SVR can encounter challenges when processing high-dimensional data, requiring more appropriate parameter settings to capture the potential nonlinear relationship between features and hardness. ET is good at processing high-dimensional features, but it also requires a lot of data support and appropriate hyperparameter adjustment. As an ensemble algorithm, RF has the characteristics of reducing the risk of overfitting and enhancing the generalization ability of the model, which makes it perform well when processing complex data sets. In addition, RF can not only perform nonlinear segmentation, but also handle interactions and complex relationships, thereby capturing the nonlinear relationships in the model excellently. The relatively simple parameter adjustment requirements are the reason why RF performs well in predicting the hardness of high-entropy nitride coatings. For different subsets, the model based on subset C ( Figure 5 c) and the model based on subset D ( Figure 5 d), its RMSE is compared with the model based on subset A ( Figure 5 a) and the model based on subset D ( Figure 5 b) is significantly smaller. During the coating preparation process, complex process parameters play a crucial role in the change of hardness. Introducing relevant process parameters can effectively improve the accuracy of machine learning modeling.

[0071] Figure 6 Based on Figure 5 The ten-fold cross validation results after hyperparameter optimization are worth noting. Figure 5 This indicates that the prediction accuracy of the machine learning model decreases as the proportion of the test set increases, indicating that the model is highly sensitive to data size. Therefore, to better demonstrate the model's generalization ability, the inventors performed 10-fold cross-validation on the model with a 10% test set after hyperparameter optimization. Cross-validation involves dividing the dataset into k equal-sized subsets, each time using k-1 subsets as the training set and the remaining subset as the test set to test the model. This process is repeated k times to obtain the predicted value for each model. Figure 6 The results show that among all models, the RF algorithm performed the most stably, with the highest accuracy and the lowest RMSE. The model based on subset C had a significant advantage in prediction accuracy over the other subsets. Therefore, the model based on the RF algorithm, a 10% test set ratio, and subset C was selected for composition optimization of high-entropy nitride coatings. This also indicates that physical descriptors derived from high-entropy alloys have limited applicability in predicting the properties of high-entropy nitride coatings.

[0072] Example 2

[0073] This embodiment is basically the same as embodiment 1, except that the specific operation of step (5) in this embodiment is as follows:

[0074] The hexavalent (excluding N) high entropy nitride coating with the highest predicted hardness was selected from the prediction results, and the corresponding high entropy alloy (Al 15 Cr 35 Nb 10 Si 15 Ti 20 V5) as the target material, and an ultra-hard high-entropy nitride coating (AlCrNbSiTiV) was prepared by sputtering under the corresponding process parameters (substrate temperature of 200 ℃, substrate bias of -100 V, deposition gas pressure of 1.0 Pa, and argon-nitrogen ratio of 2).

[0075] Example 3

[0076] This embodiment is basically the same as embodiment 1, except that the specific operation of step (5) in this embodiment is as follows:

[0077] The seven-element (excluding N) high entropy nitride coating with the highest predicted hardness was selected from the prediction results, and the corresponding high entropy alloy (Al 14 Cr 33 Nb 10 Si 14 Ta5Ti 19 V5) as the target material, and an ultra-hard high-entropy nitride coating (AlCrNbSiTaTiV) was prepared by sputtering under the corresponding process parameters (substrate temperature of 200 ℃, substrate bias of -100 V, deposition gas pressure of 0.75 Pa, and argon-nitrogen ratio of 1.5).

[0078] Example 4

[0079] This embodiment is basically the same as embodiment 1, except that the specific operation of step (5) in this embodiment is as follows:

[0080] From the prediction results, the eight-element (excluding N) high entropy nitride coating with the highest predicted hardness was selected, and the corresponding high entropy alloy (Al 10 Cr 32 Nb10 Si 14 Ta5Ti 19 V5Zr5) was used as the target material, and an ultrahard high-entropy nitride coating (AlCrNbSiTaTiVZr) was prepared by sputtering under the corresponding process parameters (substrate temperature of 200 ℃, substrate bias of -100 V, deposition gas pressure of 1.0 Pa, and argon-nitrogen ratio of 2).

[0081] Example 5

[0082] This example analyzes and characterizes the relevant properties of the ultrahard high entropy nitride coatings prepared in Examples 1-4, and studies the morphology, structure, and hardness of the above coatings.

[0083] Figure 7 The hardness comparison results of the ultra-hard high-entropy nitride coatings prepared in Examples 1-4 and the coatings in the database show that the experimental hardness of all coatings is higher than 40 GPa, reaching the level of ultra-hard coatings, indicating that the model based on subset C can better predict the hardness of high-entropy nitride coatings. Figure 7 The maximum hardness of the newly prepared coating was compared with the five-element and six-element (excluding nitrogen) high-entropy nitride coatings in the dataset. Compared with the six-element coating, the hardness increased by 5.7% (2.4 GPa), and compared with the five-element coating, the hardness increased by 4.7% (2.1 GPa). In addition, the deposition of two new seven-element and eight-element system coatings not included in the database in the examples highlights the great potential of machine learning in predicting superhard high-entropy nitride coatings.

[0084] Figure 8 (a)-(d) are the surface and cross-sectional morphologies of the ultrahard high-entropy nitride coatings prepared in Examples 1-4, respectively. All coatings have smooth and dense surfaces and cross-sections, and have good adhesion between layers and substrates. No obvious defects such as cracks and holes are observed.

[0085] Figure 9 These are the X-ray diffraction results of the ultra-hard high-entropy nitride coatings prepared in Examples 1-4. Since the high-entropy effect promotes the mutual dissolution of various elements, all coatings show a single face-centered cubic result.

[0086] Figure 10 This is a high-resolution transmission electron microscopy image of the ultrahard high-entropy nitride coating prepared in Example 1, and a low-magnification image of the coating ( Figure 10 a) Shows a columnar growth pattern with a width of approximately 20 nm. During the magnetron sputtering coating process, the limited diffusion capacity of surface atoms leads to the vertical growth around the nucleation sites, which contributes to the formation of the columnar structure. Figure 10(b) is the selected area electron diffraction pattern of the coating, which shows several Debye rings corresponding to the (111), (200) and (220) planes in the face center cubic (FCC) structure, respectively. Figure 9 The X-ray diffraction results are consistent with those of . Figure 10 c clearly shows that the coating is composed of columnar crystals and boundary phases, with penetrating lattice fringes between them rather than obvious grain boundaries.

[0087] These results demonstrate that the present invention establishes a database for predicting the hardness of high-entropy nitride coatings based on literature and experimental data. Feature engineering is used to reduce the dimensionality of features to optimize prediction accuracy. The predictive performance of different models is evaluated, and based on the optimal model, new ultrahard high-entropy nitride coatings are explored within a pre-set prediction space. Four high-entropy nitride coatings with different compositions were successfully prepared, including two new systems not found in the database. Hardness tests demonstrate that the coatings designed using machine learning exhibit excellent mechanical properties, suitable for use in harsh industrial environments.

[0088] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for preparing an ultra-hard high-entropy nitride coating based on machine learning design, characterized in that: The steps include: (1) Dataset establishment: A data set for predicting the hardness of high-entropy nitride coatings is established. The features in the data set include compositional features, physical descriptor features, process parameter features, and hardness features of high-entropy nitride coatings. The compositional features include the atomic percentages of Al, Cr, Nb, Si, Ta, Ti, V, Zr, and N. The physical descriptor features include mixing entropy, mixing enthalpy, mixing melting point, entropy-enthalpy ratio, atomic size difference, valence electron concentration, and electronegativity difference. The process parameter features include substrate temperature, substrate bias, deposition gas pressure, and argon-nitrogen ratio. The data set is divided into subsets AD. Subset A includes compositional features and hardness features. Subset B includes compositional features, physical descriptor features, and hardness features. Subset C includes compositional features, process parameter features, and hardness features. Subset D includes compositional features, physical descriptor features, process parameter features, and hardness features. (2) Feature Engineering: Through feature engineering, the features in the dataset are evaluated, their relevance and importance ranking are analyzed, and the features are optimized to improve the prediction accuracy; (3) Machine learning modeling: subsets AD of the dataset are used as input values ​​and the predicted hardness is used as output values; subset AD with a set proportion in the dataset is used as the test set, and the rest is used as the training set. Machine learning modeling is performed on subset AD using the set algorithm and test set proportion. The constructed machine learning model is subjected to hyperparameter optimization and ten-fold cross validation to evaluate its prediction performance and obtain the machine learning model with the best prediction performance. (4) Composition optimization: Input the preset prediction space into the machine learning model with the best prediction performance, and select the high-entropy nitride coating with the highest predicted hardness formed with N from each multi-element high-entropy alloy system; wherein the multi-element high-entropy alloy system is a five-element to eight-element high-entropy alloy system formed by at least five of Al, Cr, Nb, Si, Ta, Ti, V, and Zr; (5) Coating preparation: According to the predicted hardness of the high-entropy nitride coating, the corresponding multi-component high-entropy alloy target is selected, and the ultra-hard high-entropy nitride coating is prepared under the corresponding process parameter characteristics.

2. The method for preparing a superhard high-entropy nitride coating based on machine learning design according to claim 1, characterized in that: In step (2), feature engineering includes using the Pearson correlation coefficient to analyze the correlation between each feature, using the random forest algorithm to analyze the importance ranking between features, and eliminating redundant features based on the importance ranking; Redundant features are the features ranked lower in a feature pair with a Pearson correlation coefficient value greater than 0.

95.

3. The method for preparing an ultra-hard high-entropy nitride coating based on machine learning design according to claim 1, characterized in that: In step (3), the type of algorithm includes at least one of extreme tree, gradient boosting decision tree, random forest, support vector regression, and extreme gradient boosting.

4. The method for preparing a superhard high-entropy nitride coating based on machine learning design according to claim 1, characterized in that: In step (3), the test set ratio includes 10%, 20%, 30%, 40%, and 50%.

5. The method for preparing an ultra-hard high-entropy nitride coating based on machine learning design according to claim 1, characterized in that: In step (3), the hyperparameter optimization method is Bayesian optimization; the indicator of prediction performance is prediction accuracy, which is evaluated based on the root mean square error of the test set, and the machine learning model with the best prediction performance is the machine learning model with the smallest root mean square error of the test set.

6. The method for preparing an ultrahard high-entropy nitride coating based on machine learning design according to claim 1, characterized in that: In the step (4), in the preset prediction space, the composition interval of the composition feature is established by the lower limit and upper limit of the atomic percentage of each element in the data set, and the sum of the atomic percentage of each element is 100 at.%; the parameter value of the process parameter feature is selected from the three most frequently occurring values ​​in the data set.

7. The method for preparing a superhard high-entropy nitride coating based on machine learning design according to claim 6, characterized in that: The composition ranges of the composition characteristics are as follows: Al 0 at.%-18 at.%, Cr 4 at.%-28 at.%, Nb0 at.%-10 at.%, Si 0 at.%-12 at.%, Ta 0 at.%-26 at.%, Ti 0 at.%-20 at.%, V 0 at.%-12 at.%, Zr 0 at.%-12 at.%, N 40 at.%-60 at.%, and the change step size of each interval is 2 at.%; the parameter values ​​of the process parameter characteristics are as follows: substrate temperature is selected from 200 °C, 300 °C, and 400 °C; substrate bias is selected from -100 V, -150 V, and -200 V; The deposition gas pressure is selected from 0.5 Pa, 0.75 Pa, and 1.0 Pa; and the argon-nitrogen ratio is selected from 1, 1.5, and 2.

8. The method for preparing an ultra-hard high-entropy nitride coating based on machine learning design according to claim 1, wherein: In step (5), the characteristics of the multi-element high entropy alloy target and process parameters corresponding to the high entropy nitride coating with the highest hardness are predicted as follows: For the five-element high entropy alloy system, the target is Al 10 Cr 48 Nb 10 Si 12 Ti 20 , substrate temperature is 200℃, substrate bias is -100 V, deposition pressure is 0.75 Pa, argon-nitrogen ratio is 1.5; for the hexavalent high entropy alloy system, the target material is Al 15 Cr 35 Nb 10 Si 15 Ti 20 V5, substrate temperature is 200 ℃, substrate bias is -100 V, deposition pressure is 1.0 Pa, argon nitrogen ratio is 2; for the seven-element high entropy alloy system, the target material is Al 14 Cr 33 Nb 10 Si 14 Ta5Ti 19 V5, substrate temperature is 200 ℃, substrate bias is -100 V, deposition pressure is 0.75 Pa, argon nitrogen ratio is 1.5; for the octal high entropy alloy system, the target material is Al 10 Cr 32 Nb 10 Si 14 Ta5Ti 19 V5Zr5, substrate temperature is 200 ℃, substrate bias is -100 V, deposition gas pressure is 1.0 Pa, argon-nitrogen ratio is 2; the suffix value of the element in the target material represents the atomic percentage value of the element.

9. An ultrahard high-entropy nitride coating, characterized by: The method is described in any one of claims 1 to 8.

Images