A method of obtaining superhard doped DLC coatings based on deep neural networks

Abstract

The application discloses a method for obtaining superhard doped DLC coating based on a deep neural network, which comprises the following steps: collecting historical experimental data information related to doped DLC and hardness and first principle calculation information, calculating input features, and constructing an original data set; and performing standardization processing on the optimized data set to obtain a training set and a test set of doped DLC hardness data; building a DNN model, training the model with the training set data, optimizing the model hyperparameters with a Bayesian optimization algorithm, and obtaining an optimal parameter model; and predicting new doped DLC with the trained model and designing the optimal element content under the same elements. The application solves the design problem of doped DLC. At present, the design of doped DLC mostly adopts an experimental trial-and-error method and first principle simulation, which has the defects of long design cycle, large resource consumption, high cost and the like, and it is difficult to obtain doped DLC coating with expected performance. Therefore, the method can meet the rapid, accurate and efficient design of new doped DLC coating.

Description

Technical Field

[0001] This invention relates to a doped DLC coating technology, and more particularly to a method for obtaining coating formulation through machine learning, specifically a method for obtaining ultra-hard doped DLC coatings based on deep neural networks. Technical Background

[0002] Currently, diamond-like carbon (DLC) films are widely used due to their high wear resistance and chemical stability. As the service environments of critical components become increasingly demanding, the protection and reliability of these components face new challenges. In recent years, the corrosion resistance, wear resistance, and mechanical properties of DLC have received increasing attention. Current design methods for doped DLC coatings primarily employ experimental trial-and-error and first-principles simulations, both of which suffer from drawbacks such as long design cycles, high resource consumption, and high costs, making it difficult to obtain doped DLC coating materials with the desired performance. Therefore, traditional material design methods are insufficient for the rapid, accurate, and efficient design of novel high-performance doped DLC coatings. The large compositional space of doped DLC provides extensive tunability in physical, chemical, and mechanical properties, making this type of material promising for various applications.

[0003] The large compositional space also poses a significant challenge to designing novel doped DLCs with customized properties or functions. Due to the numerous combinations, traditional methods struggle to achieve the required design efficiency. Currently, machine learning plays an increasingly important role in materials science, gradually being applied to various coating designs. For example, a Chinese patent discloses an ultrahard pentagonal transition metal carbide single-phase high-entropy ceramic material and its preparation method (Publication No.: CN110194667A), which obtains an ultrahard pentagonal transition metal carbide single-phase high-entropy ceramic material, improving the density and mechanical properties of the carbide; another Chinese patent discloses transition metal carbide high-entropy ceramic particles and their preparation method (Publication No.: CN110484796A), which prepares single-phase solid solution transition metal carbide high-entropy ceramic particles through a one-step heating method. However, in the design of doped DLC coatings, there is no similar intelligent design method, resulting in long development cycles, low efficiency, and high costs for novel doped DLC coating materials. Summary of the Invention

[0004] The purpose of this invention is to propose a design method for doped DLC coatings based on deep neural networks, in order to solve the problems of difficult design, long R&D cycle, low efficiency and high cost of novel high-performance doped DLC coating materials.

[0005] The present invention adopts the following technical solution:

[0006] A method for obtaining an ultrahard doped DLC coating based on a deep neural network is characterized by comprising: collecting historical experimental data and first-principles simulation calculation data of doped metal elements; calculating input feature descriptors based on the DLC doping element composition ratio data and element chemical information; and constructing an original dataset.

[0007] The initial dataset is cleaned and normalized, and feature engineering is performed on the acquired data to construct an input feature dataset. The input features are then divided to construct a training set and a test set.

[0008] Build a DNN model. To prevent overfitting, a regularization penalty term and a Dropout function are added to the model. The DNN network model is evaluated and trained using the test set.

[0009] The hyperparameters of the DNN are processed using the Bayesian optimization (BO) method to obtain the DNN model;

[0010] A hardness prediction dataset for doped DLC with different element contents was established based on the DNN model, and combined with feature analysis, an ultra-hard doped DLC coating was obtained.

[0011] The specific steps are as follows:

[0012] Step 1: Collect historical experimental data and first-principles simulation calculations on the hardness of previous metal carbide-doped DLC coatings. The metal elements used in the DLC coatings include Si, Mo, Ti, Ag, Al, Cr, W, Nb, Cu, Ta, and La.

[0013] Step 2: Clean the collected elemental chemical properties data, removing missing and outlier values.

[0014] Step 3: Build a feature engineering system that can be used for prediction and design of doped DLC coatings. Calculate the statistical values ​​of the coating composition elements using the chemical properties of each element. The statistical values ​​include the average value and the deviation.

[0015] Step 4: Build the basic DNN model. For now, set the mode parameters to two hidden layers with 64 neurons in each hidden layer. Use L2 regularization and add a Dropout layer to prevent overfitting.

[0016] Step 5: Use the BO (Bayesian) optimization algorithm to optimize the hyperparameters of the DNN model. The optimization parameter ranges include the number of hidden layers [1, 3], the number of hidden layer neurons [16, 512], the learning rate [1e-5, 1e-1], the batch size [2, 30], the optimizer selection [Adam, SGD], and the weight decay coefficient range [0, 0.2].

[0017] Step Six: Rebuild the DNN model using the optimized parameters, design various doped DLC coatings, and use the trained model to predict the coatings. Predict the hardness of doped DLC coatings with different ratios under the same composition to obtain the optimal ratio parameters. Taking Cr-DLC as an example, set different Cr element contents, predict its hardness, and return the ratio parameters.

[0018] The doped metal elements include Si, Mo, Ti, Ag, Al, Cr, W, Nb, Cu, Ta, and La; the doped DLC is a single-element doping, and the consideration parameter is the hardness value of a certain atomic percentage of different elements.

[0019] The chemical properties of the collected elements or element carbides include relative atomic mass, atomic radius, electron affinity, melting point, density, mass, and electronegativity.

[0020] The characteristic values ​​of the coating's constituent elements are calculated using the chemical properties of each element.

[0021] The basic DNN model is built using L2 regularization and a Dropout layer is added to prevent overfitting.

[0022] The hyperparameters of the DNN model were optimized using the BO optimization algorithm. The optimized parameter ranges included the number of hidden layers [1, 3], the number of hidden layer neurons [16, 512], the learning rate [1e-5, 1e-1], the batch size [2, 30], the optimizer [Adam, SGD], and the weight decay coefficient [0, 0.2]. The optimized results were: number of hidden layers [2], number of hidden layer neurons [16, 96], learning rate [1e-5], batch size

[10] , optimizer [Adam], and weight decay coefficient [0.2].

[0023] The DNN model was rebuilt using the optimized parameters, and a doped DLC coating was designed. The trained model was used to predict the coating, and the hardness of the doped DLC coating with different atomic contents under the same element was predicted to obtain the optimal ratio parameters.

[0024] The present invention has the following effects:

[0025] This invention solves the problems of difficult design, long R&D cycle, low efficiency, and high cost of novel high-performance doped DLC coating materials by using machine learning methods to predict and design doped DLC coatings. Attached Figure Description

[0026] Figure 1 This is the overall flowchart of the present invention.

[0027] Figure 2 This is a flowchart of the BO-optimized DNN of the present invention.

[0028] Figure 3 This is a schematic diagram of the Dropout principle of the present invention.

[0029] Figure 4 This is a graph showing the change in loss value during the training process of this invention. Detailed Implementation

[0030] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0031] Please see Figure 1-2 .

[0032] A method for obtaining an ultra-hard doped DLC coating based on a deep neural network includes the following steps:

[0033] First, historical experimental data and first-principles simulation calculations on the hardness of doped DLC were collected. Elements included Si, Mo, Ti, Ag, Al, Cr, W, Nb, Cu, Ta, and La. The concentrations of all collected dopant elements were less than 50 at.%. Data included hardness, Young's modulus, and residual stress. Input feature descriptors were calculated based on the dopant element concentrations and their inherent properties, including relative atomic mass, atomic radius, melting point, electron affinity, atomic density, and electronegativity, to construct the original dataset.

[0034] Secondly, the initial dataset is cleaned by removing missing and outlier values. Feature engineering is then implemented for doped DLC, redundant features are removed, and the optimized dataset is standardized to obtain training and testing sets for doped DLC hardness prediction.

[0035] Third, build the basic DNN model. Initially, set the mode parameters to two hidden layers, with 64 neurons in each hidden layer. Use L2 regularization, which penalizes the weights to help suppress overfitting. Add a Dropout layer, the principle of which is as follows... Figure 3 Dropout refers to the process of turning off random neurons based on predetermined probabilities during training. It can also lead to unbiased fitting of the network by using irregular combinations of selected neurons in each iteration, improving the model's generalization ability. Several hyperparameters are considered in doping hardness prediction models: learning rate, optimizer, number of neurons, number of network layers, etc.

[0036] Fourth, the BO optimization algorithm was used to optimize the hyperparameters of the DNN model. The optimized parameter ranges included the number of hidden layers [1, 3], the number of hidden layer neurons [16, 512], the learning rate [1e-5, 1e-1], the batch size [2, 30], the optimizer selection [Adam, SGD], and the weight decay coefficient range [0, 0.2]. The optimized parameters were: 2 hidden layers with 16 and 96 neurons respectively, a learning rate of 1e-5, the Adam optimizer selection, and a weight decay coefficient of 0.2. The final network structure is as follows. Figure 4 .

[0037] Finally, the DNN model was rebuilt using the optimized parameters, and various doped DLC coatings were designed. The trained model was then used to predict the hardness of the coatings, specifically the DLC coatings with different concentrations of the same element, and the concentration corresponding to the maximum hardness was returned. Taking Ti-DLC doping as an example, the concentration of each Ti element was set from 0% to 50 at.%. Its hardness was predicted. For example, unexplored doping elements were selected, their inherent elemental properties were collected, and different concentrations were designed as input features to predict their corresponding hardness values.

[0038] Experiments have shown that the error between the actual tested hardness of the coating and the predicted value is between 1% and 5%, which meets the usage requirements.

[0039] The parts not covered in this invention are the same as or can be implemented using existing technologies.

Claims

1. A method for obtaining an ultra-hard doped DLC coating based on a deep neural network, characterized in that, It includes: Collect historical experimental data and first-principles simulation data of doped metal elements, calculate input feature descriptors based on DLC doping element composition ratio data and element chemical information, and construct the original dataset. The initial dataset is cleaned and normalized, and feature engineering is performed on the acquired data to construct the input feature dataset. The input features are then divided to construct the training set and the test set. Build a DNN model. To prevent overfitting, a regularization penalty term and a Dropout function are added to the model. Train and evaluate the DNN model using the training set and the test set. The hyperparameters of the DNN are processed using the Bayesian optimization (BO) method to obtain the DNN model; A hardness prediction dataset for doped DLC with different element contents was established based on the DNN model, and combined with feature analysis, an ultra-hard doped DLC coating was obtained.

2. The method according to claim 1, characterized in that: The doped metal elements include Si, Mo, Ti, Ag, Al, Cr, W, Nb, Cu, Ta, and La; the doped DLC is a single-element doping, and the consideration parameter is the hardness value of a certain atomic percentage of different elements.

3. The method according to claim 1, characterized in that: The chemical properties of the collected elements or element carbides include relative atomic mass, atomic radius, electron affinity, melting point, density, mass, and electronegativity.

4. The method according to claim 1, characterized in that: The characteristic values ​​of the coating's constituent elements are calculated using the chemical properties of each element.

5. The method according to claim 1, characterized in that: The DNN model was built using L2 regularization and a Dropout layer was added to prevent overfitting.

6. The method according to claim 1, characterized in that: The DNN model was rebuilt using the optimized parameters, and a doped DLC coating was designed. The trained model was used to predict the coating, and the hardness of the doped DLC coating with different element contents under the same element was predicted to obtain the optimal ratio parameters.

Images