A machine learning-based prediction method for multilayer nanofilm properties

Abstract

This invention discloses a machine learning-based method for predicting the properties of multilayer nanofilms. This method employs a machine learning neural network prediction model. Using an ultrafast laser ultrasonic experimental platform based on pump-probe technology and theoretical calculation methods, the method acquires instantaneous reflectivity change data. The dataset is then preprocessed and feature extracted. Finally, the obtained feature dataset is mapped to the film material properties. Using a network training algorithm, the method enables automated prediction of the material properties of multilayer nanofilm structures. This method is suitable for detecting the material parameters of multilayer nanofilms and can rapidly determine the internal structure, optical, and acoustic parameters of multilayer nanofilms with high detection accuracy.

Description

Technical Field

[0001] The present invention relates to a method for predicting properties of a multilayer nanofilm, and in particular to a method for predicting properties of a multilayer nanofilm based on machine learning. Background Art

[0002] Nano-thin film materials are the foundation of the semiconductor industry. With the improvement of semiconductor integration, accurately measuring device performance parameters (such as electrical properties, optical properties, restricted substances, etc.) has become an indispensable link in the process of optimizing design solutions and improving the success rate of product production.

[0003] Currently, mainstream methods for testing nanothin films include X-ray diffraction, electron microscopy, and ellipsometry. Ellipsometry is widely used in the semiconductor industry, but it still has several drawbacks: sample pretreatment is costly and time-consuming if the film substrate is transparent; it is difficult to measure organic thin films; it is not suitable for homogeneous materials; and analysis of test data is difficult, requiring high-level expertise from the measurement personnel. Ultrafast photoacoustic testing, a novel micro-nano nondestructive testing technique, utilizes pulsed lasers with femtosecond temporal resolution to generate and detect high-frequency ultrasound waves in materials, enabling high-resolution photoacoustic interaction detection and imaging. While achieving nanoscale resolution for microstructural imaging, it overcomes the radiation damage to samples caused by X-ray diffraction and electron microscopy, and also addresses many of the shortcomings of ellipsometry. However, due to the complex physical processes involved in the excitation, propagation, and detection of high-frequency ultrasound waves, the detected signals are also relatively complex. Manual interpretation of the signals generated by ultrafast photoacoustic testing is time-consuming and labor-intensive, and human error can significantly reduce the accuracy of test results, making it difficult to meet current production needs. Therefore, the need to automatically identify ultrafast photoacoustic detection signals through machines has become urgent. Summary of the Invention

[0004] Purpose of the invention: The present invention aims to provide a method for predicting the properties of multilayer nanofilms based on machine learning with high detection accuracy.

[0005] Technical solution: The method for predicting properties of multilayer nanofilms based on machine learning described in the present invention comprises the following steps:

[0006] S1: Select a multilayer nanofilm material sample and use an ultrafast femtosecond laser detection device to obtain data on the instantaneous reflectivity change of the sample surface; obtain simulated data of the instantaneous reflectivity of the sample through theoretical calculation methods to expand the experimental data and increase the diversity of the data set; the nanofilm can be transparent or non-transparent film or organic film;

[0007] S2: Construct a feature parameter dataset for the machine learning model and map the calculated feature values ​​to the properties of the multilayer nanofilm.

[0008] S3: Build the initial prediction network structure, train, test and evaluate the network model, create a graphical user interface based on the network model, and realize the automated prediction of the properties of multilayer nanofilms.

[0009] Preferably, the specific process of step S1 is:

[0010] S11: Prepare samples using multilayer nanofilms with different optical and acoustic properties, and use ultrafast femtosecond lasers to build a pump-probe system to detect the reflectivity or transmittance of the samples;

[0011] S12: Record the experimental parameters of the specimen as a sample parameter set, fine-tune the experimental parameters, input the adjusted sample parameter set into a theoretical calculation program, and obtain a theoretical data set of instantaneous reflectivity changes through a theoretical calculation method;

[0012] S13: Expand experimental data with theoretical calculation data to increase the diversity of the data set;

[0013] S14: Select a new sample for ultrafast photoacoustic detection, and obtain its instantaneous reflectivity or refractive index data through S12 and S13.

[0014] Preferably, the pump light pulse width used by the pump detection system in step S11 is 7 to 200 femtoseconds. In step S11, an ultrafast femtosecond laser is used to build a pump detection system. The pump light is used to excite a high-frequency acoustic pulse in the sample. After a certain time delay, the probe light is directed to the sample. Since the high-frequency acoustic pulse transmitted inside the sample will locally change the refractive index of the sample, the probe light will be reflected at the high-frequency acoustic pulse. Factors such as the interference formed by the reflection of the probe light at the high-frequency acoustic pulse and the reflection on the sample surface, displacement of the sample surface, and temperature changes on the sample surface will cause the total reflectivity or transmittance of the sample to change. A photodetector is used to detect the reflectivity or transmittance of the sample under different time delays.

[0015] Preferably, the experimental parameters in step S12 include light wavelength, incident angle, and the refractive index, thickness, and photoelastic coefficient of the specimen.

[0016] Preferably, the theoretical calculation formula in step S13 is as follows:

[0017]

[0018] Among them, k n is the wave vector; a n 、b n is the electric field constant; is the photoelastic coefficient of the material; ∈ n is the dielectric constant; u is the displacement; and n is the layer number in the thin film material.

[0019] Preferably, the specific process of step S2 is:

[0020] S21: calculating the mean of the instantaneous reflectivity change data of each sample, and quantizing the data set using a zero-mean method to obtain a zero-mean data set;

[0021] S22: Calculate the time domain envelope of the zero-mean data set to obtain an envelope data set;

[0022] S23: Select the time domain feature TD of the envelope data of each sample i (i=1, 2, 3, ..., 15), frequency domain feature FD j (j=1, 2, 3, ..., 6) time-frequency characteristics TFD k (k=1, 2, 3) construct feature data set;

[0023] S24: Corresponding the extracted feature dataset to the multilayer nanofilm properties.

[0024] Preferably, the specific process of step S3 is:

[0025] S31: Build a fully connected network with m hidden layers, each of which consists of n neurons, to construct an initial network structure; the value range of m is greater than 2, and the value of n follows the following rules:

[0026]

[0027] Among them, n s is the number of samples in the training dataset; n i is the number of input neurons; n o is the number of output neurons; α is an arbitrary value variable ranging from 2 to 10.

[0028] S32: randomly dividing the processed envelope data set into a training data set, a validation data set, and a test data set, and randomly inputting them into the network;

[0029] S33: Use the training data set to train the network parameters of the initial network structure, select the Levenberg-Marquardt function as the training method, and use the validation data set to compare and judge the performance of the network model under various parameters to obtain the optimal network structure; the test data set is used to objectively evaluate the performance of the neural network;

[0030] S34: Save the best network model obtained through training and create a graphical user interface to achieve automated prediction of multilayer nanofilm properties.

[0031] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) High precision. The present invention can measure the thickness of nanofilms in the range of 10 to 2500 nm, and the minimum measured thickness of the nanofilm can reach 10 nm, with high detection precision; (2) Unrestricted detection objects. Since the present invention adopts a pump detection system built with ultrafast femtosecond laser, its detection object can be a transparent or non-transparent film or an organic film; in addition, there are no strict requirements for the film substrate; (3) The present invention applies a machine learning algorithm to ultrafast photoacoustic detection of multilayer nanofilms, reveals the correspondence between the properties of multilayer nanofilms and the detection signals, and can quickly obtain the internal structure and optical and acoustic properties of multilayer nanofilms, solving the problems of difficult analysis and slow processing speed of ultrafast photoacoustic detection data, which is of great significance to the semiconductor industry and online non-destructive testing of integrated circuits. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart of a method for predicting properties of a multilayer nanofilm based on machine learning in an embodiment;

[0033] Figure 2 Graphs showing the experimental signal (top) and theoretical simulation signal (bottom) of a 500 nm thick SiO2 nanofilm sample obtained in the embodiment;

[0034] Figure 3 Graph showing the predicted properties of the multilayer nanofilm in the embodiment;

[0035] Figure 4 Schematic diagram of a graphical user interface for automated prediction of multilayer nanofilm properties in an embodiment. DETAILED DESCRIPTION

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0037] The following describes a method for predicting properties of multilayer nanofilms based on machine learning according to an embodiment of the present invention with reference to the accompanying drawings. Figure 1 The specific implementation steps are as follows:

[0038] S1: Select a SiO2 nanofilm sample and use an ultrafast femtosecond laser detection device to obtain the instantaneous reflectivity change data of the sample surface; obtain simulated data of the instantaneous reflectivity of the sample through theoretical calculation methods to expand the experimental data and increase the diversity of the data set. The specific steps are as follows:

[0039] S11: Prepare samples using multilayer nanofilms with different optical and acoustic properties, and use ultrafast femtosecond lasers to detect the instantaneous reflectivity or transmittance of the samples;

[0040] S12: Record the experimental parameters of the specimen as a sample parameter set, such as the light wavelength, incident angle, and refractive index, thickness (1:1:2000 nm), and photoelastic coefficient of the specimen, and fine-tune the thickness value (or light wavelength or incident angle, etc.). The thickness after fine-tuning is 1.5:1:1999.5 nm. Input the fine-tuned sample parameter set into a theoretical calculation program to obtain a theoretical data set of instantaneous reflectivity changes.

[0041] S13: Expand experimental data with theoretical calculation data to increase the diversity of the data set.

[0042] S14: Select a new sample for ultrafast photoacoustic detection, and obtain its instantaneous reflectivity or refractive index data through S12 and S13.

[0043] Examples of sample signals obtained through experimental and theoretical calculation methods are as follows: Figure 2 shown.

[0044] S2: Construct a feature parameter dataset for the machine learning model, and match the calculated feature parameter dataset with the properties of the multilayer nanofilm.

[0045] S21: calculating the mean of the instantaneous reflectivity change data of each sample, and quantizing the data set using a zero-mean method to obtain a zero-mean data set;

[0046] S22: Calculate the time domain envelope of the zero-mean data set to obtain an envelope data set;

[0047] S23: Select the time domain feature TD of the envelope data of each sample i (i=1,2,3,…,15), frequency domain feature FD j (j=1,2,3,…,6) Time-frequency characteristics TFD k (k=1, 2, 3) construct feature dataset;

[0048] S24: Corresponding the extracted feature dataset to the multilayer nanofilm properties.

[0049] S3: Construct the initial prediction network structure, divide the envelope data set into training data set, validation data set, and test data set in the ratio of 80%:10%:10%, and train, test, and evaluate the network model. According to the model prediction results, the predicted results of the multilayer nanofilm properties are plotted and compared with the measured values. Figure 3 It can be seen that the range of nanofilm thickness that can be predicted with high precision is 10 to 2500 nm, and the minimum measured nanofilm thickness can reach 10 nm. A graphical user interface is made based on the optimal network model to realize the automatic prediction of multilayer nanofilm properties (such as Figure 4 shown).

Claims

1. A method for predicting properties of multilayer nanofilms based on machine learning, characterized in that: The steps include: S1: Select a multilayer nanofilm material sample and use an ultrafast femtosecond laser detection device to obtain the change data of the instantaneous reflectivity of the sample surface; obtain simulated data of the instantaneous reflectivity of the sample through theoretical calculation methods to expand the experimental data; this process includes: S11: Prepare samples using multilayer nanofilms with different optical and acoustic properties, and use ultrafast femtosecond lasers to build a pump-probe system to detect the reflectivity or transmittance of the samples; S12: Record the experimental parameters of the specimen as a sample parameter set, fine-tune the experimental parameters, input the adjusted sample parameter set into a theoretical calculation program, and obtain a theoretical data set of instantaneous reflectivity changes through a theoretical calculation method; The theoretical calculation formula is as follows: Among them, k n is the wave vector; a n 、b n is the electric field constant; is the photoelastic coefficient of the material; ∈ n is the dielectric constant; u is the displacement; n is the layer number in the film material; S13: Expand experimental data with theoretical calculation data to increase the diversity of the data set; S14: Select a new sample for ultrafast photoacoustic detection, and obtain its instantaneous reflectivity or refractive index data through S12 and S13; S2: Constructing a feature parameter dataset for the machine learning model and mapping the calculated feature values ​​to the properties of the multilayer nanofilm. This process includes: S21: calculating the mean of the instantaneous reflectivity change data of each sample, and quantizing the data set using a zero-mean method to obtain a zero-mean data set; S22: Calculate the time domain envelope of the zero-mean data set to obtain an envelope data set; S23: Select the time domain feature TD of the envelope data of each sample i (i=1,2,3,…,15), frequency domain feature FD j (j=1,2,3,…,6) Time-frequency characteristics TFD k (k=1, 2, 3) construct feature dataset; S24: matching the extracted feature dataset with the multilayer nanofilm properties; S3: Build the initial prediction network structure, train, test, and evaluate the network model, create a graphical user interface based on the network model, and realize the automated prediction of multilayer nanofilm properties. This process includes: S31: Build a fully connected network with m hidden layers, each of which consists of n neurons, to construct an initial network structure; the value range of m is greater than 2, and the value of n follows the following rules: Among them, n s is the number of samples in the training dataset; n i is the number of input neurons; n o is the number of output neurons; α is an arbitrary value variable ranging from 2 to 10; S32: randomly dividing the processed envelope data set into a training data set, a validation data set, and a test data set, and randomly inputting them into the network; S33: Use the training data set to train the network parameters of the initial network structure, select the Levenberg-Marquardt function as the training method, and use the validation data set to compare and judge the performance of the network model under various parameters to obtain the desired network structure; S34: Save the obtained network model and create a graphical user interface to achieve automatic prediction of multilayer nanofilm properties.

2. The method for predicting properties of multilayer nanofilms based on machine learning according to claim 1, characterized in that: The pump light pulse width used by the pump detection system in step S11 is 7 to 200 femtoseconds.

3. The method for predicting properties of multilayer nanofilms based on machine learning according to claim 1, characterized in that: The experimental parameters in step S12 include the wavelength of light, the incident angle, and the refractive index, thickness, and photoelastic coefficient of the specimen.

Images