Method for identification of mechanical parameters of thin films and coatings based on instrumented indentation technique

Abstract

A method for identifying and analyzing the mechanical parameters of thin films and coatings based on instrumented indentation technology is proposed. This method establishes a dimensionless equation for the indentation of thin films and coatings using a conical indenter, which includes material constitutive and indentation information. A finite element dataset is established to fit the dimensional analysis equation, comprising a mesh dataset and a real material dataset. A neural network is used to train the mesh dataset, forming a dimensionless fitting formula for the analysis method, from the load-depth curve input from the dimensional analysis equation to the material stress-strain curve output. The real material dataset is used as a test set to evaluate the accuracy of the model fitting formula. The load-depth curve is obtained using laboratory instruments. The load-depth curve is substituted into the dimensionless fitting formula obtained in the preceding steps to obtain the output dimensionless stress-strain curve. The required real material stress-strain curve is then calculated based on the dimensional relationships.

Description

Technical Field

[0001] This invention relates to a parameter analysis method, and more particularly to a method for identifying and analyzing the mechanical parameters of thin films and coatings based on instrumented indentation technology. It belongs to the field of micro / nano mechanical testing. Background Technology

[0002] Thin film and coating structures refer to structures on which functional materials are coated onto a substrate material, such as thin films used in chips and coatings on engine blades. Developing methods for testing and analyzing the mechanical properties of thin film and coating structures is beneficial for the development of their fabrication processes and the exploration of their applications. Instrumented indentation testing technology, with its surface, micro-area, and micro-destructive characteristics, is the most commonly used method for testing the mechanical properties of thin film and coating materials, especially chip materials. This testing method is applicable to nanometer indentation testing of thin films (thickness below 10 μm) and micrometer indentation testing of coatings (thickness above 10 μm).

[0003] Below, we will describe the methods used in the existing technology:

[0004] Empirical estimation method: For testing the hardness and elastic modulus of films and coatings, the national standard GB / T 25898—2010 recommends the use of continuous stiffness measurement method. According to different hardness / modulus curve forms, three types of empirical estimation methods for hardness and elastic modulus are selected: (1) Plateau method: Determine whether a plateau appears on the indentation hardness and elastic modulus curve. If it appears, the value can be taken within the indentation depth range where the plateau is located. See Figure 2 (2) Peak / Valuation Method: If no plateau appears, only the peak value (maximum value) is available. This value can be used as the minimum estimate of the indentation hardness or elastic modulus of the film and coating. See [link to relevant documentation]. Figure 3 If neither a plateau nor a peak occurs, and only the minimum trough value is found, the trough value of the elastic modulus can be used as the maximum estimate of the elastic modulus of the film and coating. See [link to relevant documentation]. Figure 4 (a) , and the valley value should not be used as a hardness estimate. If there is a transition point where the hardness increases rapidly with depth, then the value at that point can be used as the maximum estimate of the hardness of the film and coating. See [reference needed]. Figure 4 (b). (3) Extrapolation approximation method: If no plateau, peak, valley, or transition point appears, the indentation hardness and elastic modulus within a certain indentation depth range can be linearly extrapolated to zero depth to obtain estimated values ​​of the indentation hardness and elastic modulus of the film and coating. See [reference needed]. Figure 2 and Figure 3 The indentation depth range of the data used for extrapolation should not be too small (greater than 50 nm).

[0005] The series-parallel analogy of springs (Hay-Crawford method):

[0006] Jennifer Hay and Crawford of KLA obtained the elastic modulus of materials from indentation unloading curves using the Oliver-Pharr and Song-Pharr methods. They considered the film and coating as supporting springs, the substrate as parallel springs, and the entire structure as a series-parallel spring system. (See [link to relevant documentation]). Figure 6 The unloading section, from 25% to 50%, can be fitted as follows:

[0007]

[0008] In the formula, These are weighting factors derived by Gao et al.; see the notes for their expansion. =0.0626 is a constant obtained through finite element fitting. It is the folded shear modulus of the thin film and coating-substrate structure. It is the shear modulus of the substrate. This refers to the shear modulus of films and coatings; the conversion relationship between shear modulus and elastic modulus is as follows:

[0009]

[0010]

[0011]

[0012] In the formula, and These are the reduced elastic modulus and Poisson's ratio of the thin film and coating-substrate structure. and It refers to the elastic modulus and Poisson's ratio of the substrate. and These are the elastic modulus and Poisson's ratio of thin films and coatings. The fitting relationship is shown in the notes.

[0013] The series-parallel analogy method for identifying the elastic modulus of films and coatings is relatively stable and reliable, but the modeling method is based on a purely elastic model and cannot be extended to the identification of plastic parameters.

[0014] This section references:

[0015] Hay J, Crawford B. Measuring substrate-independent modulus of thinfilms[J]. Journal of Materials Research, 2011, 26(6): 727-738.

[0016] The KLA G200X product manual clearly states that it uses the Hay-Crawford model.

[0017] The empirical 10% principle method: This method controls the maximum indentation depth within 10% of the film and coating thickness. In this case, it is empirically assumed that the indentation deformation field is entirely within the thickness of the film and coating, and no deformation occurs inside the substrate. The semi-infinite assumption can be applied to the film and coating area, and its elastoplastic mechanical properties can be obtained by analyzing the bulk material using methods such as the Oliver-Pharr method recommended by ISO 14577, the Cao-Lu method, and the Jiang-Zhang spherical indentation method.

[0018] This section references:

[0019] ISO 14577-1:2015. Metallic materials -- Instrumented indentation test for hardness and materials parameters -- Part 1: Test method

[0020] Example of KLA's 10% principle: combined with ISO 14577.

[0021] Cao YP, Lu J. A new method to extract the plastic properties of metal materials from an instrumented spherical indentation loading curve. Acta Materialia, 2004, 52: 4023~4032

[0022] Jiang P, Zhang TH, Feng YH, Yang R, Liang N G. Determination ofplastic properties by instrumented spherical indentation: Expanding cavitymodel and similarity solution approach. Journal of Materials Research, 2009,24(3): 1030~ 1038

[0023] Dimensional Analysis + Least Squares Fitting Method: Zhao and Chen, based on dimensional analysis, used a conical indenter and introduced dimensionless total indentation work and unloading work to replace residual indentation as analytical parameters. For two different indentation depths, they fitted the relationships between the dimensionless maximum indentation load, total indentation work, and unloading work and the constitutive parameters. A total of six equations were formed, and three of them were selected to solve for the constitutive parameters. Due to the complexity of the dimensionless fitting form and the large number of fitting terms (more than 20), the fitting coefficients could not be directly solved from the equations, requiring multiple finite element simulation iterations, resulting in low computational efficiency and high solution difficulty.

[0024] This section references:

[0025] Zhao M, Chen X, Xiang Y, et al. Measuring elastoplastic properties ofthin films on an elastic substrate using sharp indentation[J]. ActaMaterialia, 2007, 55(18): 6260-6274.

[0026] However, the aforementioned existing technologies have the following drawbacks:

[0027] (1) The 10% rule cannot meet the cutting-edge mechanical testing requirements of thinner films: Due to factors such as indenter tip wear and strain gradient effects, the depth of instrumented indentation has a lower limit requirement, which cannot meet the requirements of nanometer indenters for testing thinner films and micrometer indenters for testing thinner coatings. Furthermore, the 10% rule method is primarily used for testing material hardness and elastic modulus. Most methods for testing the plasticity parameters of bulk materials have higher minimum indentation requirements to obtain more plastic indentation information. When this type of method is actually used to measure films and coatings, the lower limit requirement for thickness is even higher, resulting in a narrower testing range. The empirical 10% rule method is only suitable for thicker films and coatings and cannot meet the cutting-edge mechanical testing requirements of thinner films and coatings.

[0028] (2) There is a technological gap in the identification and analysis methods for the plastic mechanical parameters of films and coatings considering the substrate effect: The three estimation methods recommended by the national standard GB / T 25898—2010, and the analytical method recommended by the commercial instrument company KLA that treats the surface material and substrate as a series-parallel spring, can only test the hardness and elastic modulus of the film and coating structure, but cannot test the plastic parameters of the film and coating. Empirical estimation methods are based on empirical estimation and extrapolation of experimental data processing, lack theoretical basis, and can only analyze the elastic parameters of the sample.

[0029] (3) Solving the indentation elastoplastic model of thin films and coatings is difficult, and empirical formulas are scarce: Laboratory methods for elastoplastic analysis of thin film and coating structures are mostly based on the technical route of "dimensional analysis + least squares fitting". This method has been successful in the elastoplastic parameter analysis of bulk materials, but due to the difficulty in solving the indentation mechanical model of thin film and coating structures, there are few empirical formulas available to simplify the problem, and the results of fitting with simple polynomials are not good. Fitting relationships can only be established for specific types of materials with a small range of parameters. The mechanical properties of ultrathin films of various metal materials are very different from those of the same bulk material. They cannot be predicted solely based on the mechanical properties of the bulk material. They must be treated as unknown materials, and analytical methods that can predict a wide range of parameters must be used to predict their mechanical properties.

[0030] (4) The reliability of the data-driven fitting analysis method is yet to be determined: Neural network methods can fit large amounts of data and have strong convergence capabilities. This scheme is based on the technical route of "dimensional analysis + data-driven fitting", using neural network methods combined with training / test datasets with a wide range of parameters to fit the functional equations of dimensional analysis. Since the fitting form of neural networks is not generated by physical modeling and there are no empirical formulas to limit or reference, the reliability of the model is questionable when using universal methods to fit dimensional analysis equations.

[0031] This section covers four existing technologies: 1. Thin film and coating testing has current national standards; 2. Commercial instrumentation integrating the method proposed by KLA (Hay-Crawford method) with the latest G200X instrument is widely adopted; 3. The empirical 10% rule is currently used in many laboratories and is also recommended by KLA; 4. Dimensional analysis + least squares fitting is a cutting-edge technical approach, not yet commercially adopted, but closely related to this invention. Technical approach 4 solves the problems of 1-3 in a research sense, but its accuracy is low and it has not been widely used. Summary of the Invention

[0032] To address the shortcomings of existing technologies, this invention discloses a method for identifying and analyzing the mechanical parameters of thin films and coatings based on instrumented indentation technology. The technical solution is as follows:

[0033] A method for identifying and analyzing the mechanical parameters of thin films and coatings based on instrumented indentation technology includes the following steps:

[0034] Step 1: Establish a dimensionless equation for the indentation of thin films and coatings using a conical indenter, which includes material constitutive and indentation information;

[0035] Step 2: Based on the dimensionless equation obtained in Step 1, establish the finite element dataset required for neural network fitting. This finite element dataset consists of two parts: a mesh dataset and a real material dataset.

[0036] Step 3: Using a neural network, train the fitted mesh dataset to form an analytical method fitting formula from the load-depth curve input pressed into the dimensional analysis equation to the material stress-strain curve output; use the real material dataset as a test set to evaluate the accuracy of the model fitting formula.

[0037] Step 4: Obtain the load-depth curve using laboratory instruments;

[0038] Step 5: Substitute the load-depth curve obtained in Step 4 into the dimensionless fitting equation obtained from the neural network fitting in Step 3 to obtain the dimensionless output. - The curve is calculated based on dimensional relationships to obtain the required material reality. - curve. Beneficial effects

[0039] (1) Integrated parameter identification of elastic-plastic mechanical properties of film and coating: Only one parameter, the elastic modulus of the substrate, needs to be tested in advance, and only one cone indentation test is needed to obtain the elastic-plastic stress-strain relationship and parameters of film and coating.

[0040] (2) The model considers the substrate effect and allows for testing of thinner films and coatings: Finite element modeling is based on deformable substrates and allows for analysis of the elastic-plastic parameters of films and coatings by considering the substrate effect.

[0041] (3) Bypassing the physical modeling step, data-driven fitting of dimensional analysis formulas: Traditional bulk material analysis methods are based on physical modeling (such as the pore expansion model) or empirical relationships (such as representative strain). Thin film and coating indentation mechanics problems are difficult to simplify based on physical modeling, and there are no effective simplified empirical formulas. Least square fitting is difficult and the accuracy is hard to guarantee. Data-driven fitting methods are based on large-scale datasets and achieve good test results over a wide range of parameters.

[0042] (4) Establish a real-world materials dataset, consisting of 96 sets of data. Use a large-scale test dataset to verify the credibility of the data-driven approach. Attached Figure Description

[0043] Figure 1 Schematic diagram of indentation load-depth curve;

[0044] Figure 2 Schematic diagram of the structure of a conical indenter pressing in a thin film and coating;

[0045] Figure 3 Test results of SiO2 thin film samples thermally oxidized on Si(100) substrate

[0046] a) Indentation hardness-depth curve, b) Elastic modulus-depth curve;

[0047] Figure 4 Test results of approximately 320 nm thick DLC thin film samples sputtered by ion arc plating on GT35 steel substrate: a) indentation hardness-depth curve, b) elastic modulus-depth curve;

[0048] Figure 5 Test results of a 311 nm thick SiO2 thin film sample grown by thermal oxidation on a Si(100) substrate: a) Elastic modulus-depth curve, b) Indentation hardness-depth curve.

[0049] Figure 6 A schematic diagram of a series-parallel model of thin films and coatings with a substrate;

[0050] Figure 7 Schematic diagram of neural network fitting;

[0051] Figure 8 The fitting results of the neural network to the training set, validation set, and test set (training 15,000 steps);

[0052] Figure 9 Step-by-step flowchart. Detailed Implementation

[0053] Example 1

[0054] A method for identifying and analyzing the mechanical parameters of thin films and coatings based on instrumented indentation technology includes the following steps:

[0055] Step 1: Establish a dimensionless equation for the indentation of thin films and coatings using a conical indenter. This dimensionless equation includes material constitutive information and indentation information.

[0056] Test objects and requirements: Based on instrumented nanoindentation testing technology, test the elastoplastic mechanical parameters of chip-type thin films and coating structures.

[0057] Mechanical Model Establishment: A linear-power hardening constitutive monolayer film and coating cover a purely elastic substrate, and the elastic modulus of the substrate is known. First, the linear-power hardening assumption of the material is given.

[0058]

[0059] In the formula, The yield strength of the material. The elastic modulus of the material, For material strain, For the material's yield strain, Let be the material hardening index. Based on this constitutive assumption and the instrumented indentation parameters, the dimensional analysis formulas for the cone indentation of thin films and coating structures are written as follows:

[0060]

[0061] In the formula, For elastic modulus, For yield strength, The hardening index, Poisson's ratio, subscript The following table represents thin films and coatings. Indicates the base; For maximum indentation depth, For thin film and coating film thickness, Let be the equivalent half-cone angle of the conical indenter. Referring to Cheng and Cheng's dimensional analysis method for bulk materials, it can be dimensionless and written as:

[0062]

[0063] It is observed that introducing the assumption of a purely elastic substrate reduces two analytical parameters, greatly simplifying the analysis. Chip-type materials often use brittle materials like silicon as the substrate, whose breaking strength is lower than its yield strength. A softer metal is coated onto the substrate as the functional material. Since the breaking strength of the substrate is higher than the yield strength of the surface-coated metal, the assumption of a purely elastic substrate can be applied without introducing significant errors.

[0064]

[0065] Omitted constant term (equivalent half-cone angle of indenter) Given the ratio of 1 / 2 and Poisson's ratio (which has little effect), we get:

[0066]

[0067] Use analytical parameters on the left side of the equation. express The dimensionless form is obtained as follows:

[0068]

[0069] At this point, the variable on the left side of the equals sign is the load. and relative indentation depth The right side shows the load-depth curve corresponding to the indentation; the right side shows the elastic-plastic parameters of the film and coating, corresponding to the stress-strain curve of the material. Based on the dimensional analysis equation, a dimensional analysis neural network fitting relationship is selected, with the load-depth curve as input and the material stress-strain relationship as output. A framework for constructing the dataset is established based on this input-output relationship. This refers to the indentation depth during the indentation process.

[0070] Step 2: Establish the finite element dataset required to fit the dimensional analysis equations. This finite element dataset consists of two parts: a mesh dataset and a real material dataset.

[0071] The traditional "dimensional analysis + least squares fitting" approach generally has a strong physical background and simplified empirical formulas. Such problems are relatively easy to fit, and good results can be obtained with a small number of samples (dozens of sets of experimental results). However, the conical indentation problem of thin films and coatings, when the maximum indentation depth exceeds 10% of the film or coating thickness, becomes a large deformation problem, making it difficult to establish an effective physical model for analysis and solution. A common approach is to predefine a range of material parameters and fit the results within this prior judgment range to form an analytical method.

[0072] Establishing datasets with a wide parameter range: Thin films and coatings, due to their lower thickness, have different microstructures than bulk materials or thick films, often exhibiting different mechanical properties. In most cases, the elastic modulus, yield strength, and hardening index of films and coatings are much higher, behaving like a different material. Therefore, prior predictions of material parameters based on experience with bulk materials are likely to fail, necessitating the establishment of analytical methods with a wide parameter range. Within this range, using neural networks to fit large-scale datasets (thousands of sets) becomes an effective approach.

[0073] A finite element mesh dataset is established as both the training and validation set: The brittle material substrate is assumed to be purely elastic, i.e., the substrate plasticity parameter is not considered; simultaneously, the thin film or coated metallic material is assumed to have a linear elastic-power-hardening constitutive model. Considering the different film-substrate (or coating-substrate) material performance adaptation relationships, in...

[0074]

[0075]

[0076]

[0077] Within the parameter range, the finite element simulation material parameters used in the dataset are selected to ensure that the material parameter data in the training set approximates a normal distribution while maintaining parameter density. The values ​​are shown in the table. The interval between values ​​is approximately logarithmic. There are three reasons for this approach: First, equal intervals result in excessively large step sizes for smaller values ​​and excessively small step sizes for larger values, leading to a data distribution that is too biased towards the larger value range and can easily cause numerical mismatch. Second, the elastic modulus of many films or coatings is actually similar to that of the substrate, and the yield strength is also in the range of 0.5 to 2 times the elastic modulus (around 0.1%), effectively reducing the training cost for extreme values. Third, this method makes the numerical distribution of the training set tend towards the center, exhibiting an approximate normal distribution. This effectively increases the data density in the central region, meeting the needs of real material testing, and also makes the data distribution of the training set, validation set, and test set nearly consistent.

[0078] A finite element "real" material dataset was established as the test set: the aforementioned mesh dataset was obtained by taking integer points based on the upper and lower limits of the material's elastoplastic data. This dataset was then divided into training and validation sets for model training and result selection. The final model requires an independent test dataset for validation, which is not used in the training process. Six common brittle semiconductor materials were selected as substrates, and eight common chip metal materials were selected as coating films. Ninety-six sets of data were generated by randomly selecting n values ​​in two intervals: 0–0.25 and 0.25–0.5. Finite element simulation was used to generate the finite element "real" material dataset.

[0079] Step 3: Using a neural network, train the fitted mesh dataset to form an analytical method fitting formula that outputs the material stress-strain curve from the load-depth curve input by the dimensional analysis equation; use the real material dataset as a test set to evaluate the accuracy of the model fitting formula.

[0080] Dataset preprocessing includes: 1) Finite element isometric interpolation: The output of the finite element simulation yields load-depth curves with non-uniform indentation depths. A script is written to interpolate these curves to achieve uniform indentation depths. Since residual indentation depth varies for different materials, a uniform 10% unloading curve length is used to maximize the utilization of elastic information in the unloading segment. 2) Generating labeled stress-strain curves for training: A script is written to generate labeled stress-strain curve data files with a maximum strain of 1%. The 1% maximum strain output ensures a balance between the elastic and plastic segments. 3) Extrema are excluded: Data entries in the dataset with yield strain greater than 1% of the elastic-plastic parameter are removed. 4) The dataset is randomly divided into training and test sets in a 7:1 ratio.

[0081] A 5-layer DNN neural network was constructed: After multiple tests and parameter tuning, including switching between DNN (Deep Neural Network), LSTM (Long Short Memory Neural Network), and CNN (Convolutional Neural Network), the tests found that the 3-layer deep neural network performed best. ReLU activation function was used between layers, the AdamW optimizer was used for descent during training, and the Mean Absolute Percentage Error (MAPE) function was used as the loss function.

[0082]

[0083]

[0084] In the formula, Represents the ReLU activation function. As the output of the previous layer of the neural network, the ReLU activation function represents a 0-linear activation function. The sampling point number is the output side stress-strain curve. The number of sampling points represents the number of points on the curve. Calculate the average of the sampling points. The sampling point values ​​are for predicting the curve. Let be the conventional true value of the stress-strain curve. Using this neural network, a dimensionless fitting equation is derived, taking the dimensionless load-depth curve as input and the material stress-strain curve as output.

[0085] Noise and bias resistance verification: Write noise and bias modules, add Gaussian white noise with a signal-to-noise ratio of 30 to the training set, and add 3% to 5% relative bias to the validation and test sets to test the noise and bias resistance performance of the model.

[0086] Develop the analysis software: Save the fitted relationship, write it into an analysis software that can acquire input curves, obtain output curves, and include a data post-processing module. See step 5 for the data post-processing section.

[0087] Training Results: Using MAPE values ​​of 20%, 15%, 10%, and 7.5% as benchmarks, the percentage of data in the test set (real-world data set) that falls below each benchmark is presented. This illustrates the distribution of the training results. (See [link to relevant documentation]). Figure 7 , Figure 8 As shown in the figure, 93.8% of the test set data had a MAPE loss of less than 20%, 90.6% had a MAPE loss of less than 15%, 76% had a MAPE loss of less than 10%, and 51% had a MAPE loss of less than 7.5%. Furthermore, the model exhibits ideal resistance to bias and noise, capable of withstanding noise at a signal-to-noise ratio of 30 and bias of 3%-5%.

[0088] Step 4: Obtain the load-depth curve using laboratory instruments;

[0089] The load-depth curves of the instrumented indentation test were obtained in the laboratory. The data were processed using laboratory data processing methods (calibrating the indenter, determining the initial contact zero point, etc.) to obtain load-depth curves suitable for analysis using this method.

[0090] Step 5: Substitute the load-depth curve into the dimensionless fitting equation to obtain the output dimensionless stress. -strain The curve is calculated based on the dimensionless formula to obtain the required material reality. - curve.

[0091] Compared with existing technologies:

[0092] This invention effectively addresses the problem of identifying the elastoplastic parameters of thin films and coatings containing substrates by establishing a framework for neural network analysis methods. This framework includes, but is not limited to, dimensional analysis principles, dataset generation, neural network construction, and pre- and post-processing of data. It allows researchers and testers to obtain well-fitting analytical formulas without relying on difficult analytical methods.

[0093] The analytical method of this invention can effectively perform instrumented indentation tests on films and coatings containing substrates. By employing an analytical method developed using neural networks, the elastic-plastic parameters of films and coatings that effectively remove substrate effects can be obtained, enabling testers to obtain more accurate material parameters.

[0094] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for identifying and analyzing the mechanical parameters of thin films and coatings based on instrumented indentation technology, characterized by: Step 1: Establish a dimensionless equation for the indentation of thin films and coatings using a conical indenter. This dimensionless equation includes material constitutive and indentation information; the expression for the dimensionless equation is: ; In the formula, The pressing load during the pressing process, The indentation depth during the indentation process, expressed as a dimensionless subscript. Indicates the loading process. For elastic modulus, For yield strength, The hardening index is indicated by the subscript. Indicates a thin film or coating, subscript Indicates the base; The thickness of the film or coating; Step 2: Establish the finite element dataset required to fit the dimensionless equation. The finite element dataset includes a mesh dataset and a real material finite element dataset. Step 3: Using a neural network, train the mesh dataset to fit and form a dimensionless fitting formula for the analysis method, which is the input of the load-depth curve pressed into the dimensionless equation and the output of the material stress-strain curve; use the real material finite element dataset as a test set to evaluate the accuracy of the model fitting formula. Step 4: Obtain the sample load-depth curve using laboratory instrumented indentation equipment; Step 5: Substitute the sample load-depth curve obtained in Step 4 into the dimensionless fitting formula of the neural network obtained in Step 3 to obtain the output dimensionless stress-strain curve. Calculate the required true stress-strain curve of the material based on the dimensional relationship.

2. The method for identifying and analyzing the mechanical parameters of thin films and coatings based on instrumented indentation technology according to claim 1, characterized in that: Step 2 further includes the following: establishing a finite element mesh dataset as both a training and validation set: assuming the brittle material substrate is purely elastic, i.e., the substrate plasticity parameter is not considered; simultaneously assuming the film and coating metal materials are linearly elastic-power-hardening constitutive models; considering the different film-substrate or coating-substrate material performance compatibility relationships, ； ； ； Within the parameter range, the finite element simulation material parameters used in the dataset are selected to make the material parameters of the mesh dataset approximately normally distributed, while ensuring that the selection density of the material parameters in the mesh dataset is approximately logarithmically equidistant. Establish a real material finite element dataset as the test set: The aforementioned mesh dataset is obtained by taking integer points based on the upper and lower limits of the material's elastic-plastic data. It is then divided into a training set and a validation set for model training and result selection. The final model needs to be validated using an independent test dataset that is not used in training. Six brittle semiconductor materials were selected as substrates, and eight metallic materials used in chips were selected as coating films. The n values ​​were randomly selected in two intervals: 0~0.25 and 0.25~0.5, for a total of 96 sets of data. The real material dataset was generated by finite element simulation.

3. The method for identifying and analyzing the mechanical parameters of thin films and coatings based on instrumented indentation technology according to claim 1, characterized in that: Step 3 further includes the following: Dataset preprocessing includes: 1) Finite element isometric interpolation: The output of the finite element simulation yields load-depth curves with non-uniform indentation depths. A script is written to interpolate these curves to achieve uniform indentation depths. Since the residual indentation depth varies for different materials, a uniform 10% unloading curve length is used to maximize the utilization of elastic information in the unloading segment. 2) Generating labeled stress-strain curves for training: A script is written to generate labeled stress-strain curve data files with a maximum strain of 1%. The 1% maximum strain output ensures the balance between the elastic and plastic segments. 3) Extrema are excluded: Data entries in the dataset with yield strain greater than 1% are removed from the elastic-plastic parameter data. Building a 5-layer DNN neural network: A deep neural network with 3 hidden layers is used, with ReLU activation function between layers. The AdamW optimizer is used for training descent, and the Mean Absolute Percentage Error (MAPE) function is used as the loss function. ； ； In the formula, Represents the ReLU activation function. As the output of the previous layer of the neural network, the ReLU activation function represents a 0-linear activation function; The sampling point number is the output side stress-strain curve. The number of sampling points represents the number of points on the curve. Calculate the average of the sampling points. The sampling point values ​​are for predicting the curve. This represents the conventional true value of the stress-strain curve; The neural network is used to fit a dimensionless fitting formula with a dimensionless load-depth curve as input and a material stress-strain curve as output. Noise and bias resistance verification: Write noise and bias modules, add Gaussian white noise with a signal-to-noise ratio of 30 to the training set, and add 3% to 5% relative bias to the validation and test sets to test the noise and bias resistance performance of the model; Analysis software is developed by saving the fitted relationship, writing it to obtain the input curve, obtain the output curve, and including a data post-processing module.

4. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored program, wherein the program, when executed, controls the device on which the non-volatile storage medium resides to perform the method described in any one of claims 1 to 3.

5. An electronic device, characterized in that: It includes a processor and a memory; the memory stores computer-readable instructions, and the processor is configured to execute the computer-readable instructions, wherein the computer-readable instructions, when executed, perform the method according to any one of claims 1 to 3.

Images