A method for a machine learning to identify the crystal phase distribution of polycrystalline thin films in nanodevices

Abstract

The present invention discloses a method for machine learning to identify the crystalline phase distribution of polycrystalline thin films in nanodevices. By automatically adjusting the simulation parameters of transmission electron microscopy (TEM) simulation software, a series of simulated TEM images of nano-polycrystalline functional thin films are obtained as a database file. Deep learning convolutional neural networks are constructed for each of the crystal structure, tilt direction, sample thickness, and other crystallographic parameters of the polycrystalline functional thin film. This neural network is trained to simulate the constructed TEM image database. After the neural network training is completed, TEM images of grains at different locations in the actual polycrystalline functional thin film are used to identify the crystalline phase distribution. The present invention can replace manual methods to achieve automatic, rapid, and reliable identification of the crystalline phase of nanocrystals within the polycrystalline functional thin film.

Description

Technical Field

[0001] The present invention belongs to the field of microscopic image structure recognition of polycrystalline functional materials in nanodevices, and relates to a method for machine learning-assisted recognition of the crystal phase distribution of polycrystalline functional materials. Background Art

[0002] Since the 21st century, information industries such as computers and the internet have experienced rapid growth. The semiconductor industry has gradually become a crucial foundation for national economic development, symbolizing the advancement of modern science and technology. As the fundamental building blocks of semiconductor integrated circuits, electronic components such as resistors, capacitors, inductors, and transistors have also been shrinking in size in line with Moore's Law, reaching the nanometer scale. For example, the gate size of mainstream transistors is currently 12 nanometers, while the core memory cells in new phase-change memories and ferroelectric memories are typically tens of nanometers. Device miniaturization can improve device density and performance. Further optimization of these performances requires understanding their operating principles and failure mechanisms at a more microscopic scale, placing higher demands on device structural characterization. In Ge2Sb2Te5-based phase-change memories (thickness of approximately 50 nanometers), the thin film material undergoes a reversible phase transition between amorphous and face-centered cubic nanocrystals (grain size approximately 10 nanometers). Excessive electric field strength results in the formation of thermodynamically more stable hexagonal phase grains. While these two phases have similar structures, the hexagonal phase consumes more power, hindering information storage, and therefore, it is desirable to minimize this transition. In hafnium zirconium oxide (HZO)-based ferroelectric memory (film thickness approximately 10 nanometers), ferroelectric orthorhombic (O) nanocrystals (grain size approximately 10 nanometers) within the film undergo atomic orientation shifts under the action of an electric field, enabling information storage. Currently, due to the immaturity of device processing, nanocrystals with other crystalline structures, such as the paraelectric monoclinic (M) and antiferroelectric tetragonal (T), also exist within the HZO film. The presence of these similar structures hinders the full application of ferroelectricity.

[0003] Accurately measuring the distribution of the crystal structure within these functional films can help understand their working principles and is the basis for further process optimization. However, due to the small size of the devices and the similarity of the crystal phase structures, conventional macroscopic experimental methods are often difficult to effectively distinguish. Spherical aberration-corrected transmission electron microscopy (Cs-TEM) can obtain the microscopic morphology and structure of the material in three dimensions. For example, high-resolution transmission electron microscopy (HREM), high-angle annular dark field (HAADF) and annular bright field (ABF) can show the changes in atomic arrangement at different positions in similar structures, such as the slip between two atomic layers when the face-centered cubic phase transforms to the hexagonal phase in phase change memory, and the atomic dislocation within the unit cell when the ferroelectric O phase transforms to the antiferroelectric T phase in ferroelectric memory. By identifying the arrangement characteristics of atoms in local areas, the crystal structure can be accurately distinguished at the nanometer or even sub-nanometer scale without the need for Fourier transform of the image. The four-dimensional scanning transmission electron microscopy (4D-STEM) technology that has emerged in Cs-TEM in recent years can record the convergent beam diffraction, nanobeam diffraction, and Ronchigram of the corresponding sites while capturing real-space images. Since these images are very sensitive to the crystal structure, thickness, and tilt direction, the crystal structure can be directly determined by analyzing the overall characteristics of these images. The resolution depends on the step size of the electron probe and can reach the atomic scale.

[0004] Because the crystal structure of nanocrystals can be arbitrarily oriented in three-dimensional space, when using images such as HREM and HAADF for crystal phase identification, transmission electron microscopy simulation software is usually used to first obtain theoretical images for crystal phase comparison. However, due to the random orientation and thickness of crystals, it is inevitable to manually modify the crystal phase structure, orientation, thickness and other parameters during simulation, which is repetitive and time-consuming. Based on the mainstream computer system programming software development platform, the simulation software can be used to traverse the simulation parameters with a set step size to obtain a series of transmission electron microscopy simulation images as a database file, which will greatly facilitate subsequent image comparison tasks. The size of the step size determines the precision of the database file.

[0005] Generally, the size of nanocrystal grains is relatively small compared to the entire device unit. Usually, a device will contain several nanocrystal grains. Therefore, even calibrating the crystal phase structure inside a device will take a very long time. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for machine learning to identify the crystal phase distribution of polycrystalline thin films in nanodevices. By automatically adjusting the transmission electron microscope simulation software parameters through software calls, a series of transmission electron microscope simulation images are obtained as a database. A deep learning convolutional neural network is constructed for the crystal structure, tilt direction, sample thickness and other crystallographic parameters of the polycrystalline functional thin film to be identified. The constructed neural network is trained using the transmission electron microscope simulation image database as a data set. After the neural network training is completed, feature judgment is performed on transmission electron microscope photos of grains at different positions in the actual polycrystalline functional thin film, thereby completing automatic, rapid and reliable identification of the crystal phase distribution of the actual polycrystalline functional thin film.

[0007] Specific technical solutions for achieving the purpose of the present invention:

[0008] A method for machine learning-assisted identification of the crystal phase distribution of a polycrystalline functional thin film in a nanodevice comprises the following steps:

[0009] Step 1: Build a corresponding atomic model based on the atomic arrangement of the constituent atoms in different crystal structures of the polycrystalline functional film in three-dimensional space;

[0010] Step 2: Import the atomic model established in step 1 into commercial or open-source transmission electron microscope image simulation software, and adjust the parameters according to the actual transmission electron microscope shooting conditions.

[0011] Step 3: Based on a mainstream computer system programming software development platform, the image simulation software used in Step 2 is used as the development object. For different crystal structures, the image simulation software is used to traverse the tilt direction (1, φ, θ) and sample thickness parameters represented by the three-dimensional spherical coordinate system with a set step size, and execute the transmission electron microscope image simulation operation process. A series of transmission electron microscope simulation images under different transmission electron microscope shooting modes are obtained and automatically saved;

[0012] Step 4: Construct N corresponding deep learning convolutional neural networks in parallel for the N crystal phase parameters of the polycrystalline functional film. Use the series of transmission electron microscope simulated images obtained in step 3 as the label data input set of a certain crystal phase parameter neural network, and extract the image features of the M sub-category crystal phase parameters in the local area through machine learning. The Softmax layer of the neural network outputs the probability of different sub-category crystal phase parameters in the total category M, and the error is obtained based on the output value and the input label value. Use the back propagation of the error to update the weight of the neural network and train the neural network again until the error value is less than the set accuracy to complete the recognition training of the neural network. Finally, use the same simulated image label data set to complete the recognition training of other crystal phase parameter neural networks.

[0013] Step 5: Use standard semiconductor processes to prepare nanodevices, such as phase change memory and ferroelectric memory, where the thickness of the polycrystalline functional film is less than 30 nanometers, and use an electrical test system to verify its performance;

[0014] Step 6: Use standard focused ion beam processing technology to prepare the device with good performance verified in step 5 into nanosheets;

[0015] Step 7: Observe the nanosheet prepared in step 6 under a transmission electron microscope. After adjusting the state of the electron microscope, select a suitable imaging mode to continuously capture microstructure images of different positions of the polycrystalline functional film.

[0016] Step 8: When the overall microstructure image captured in step 7 is used as data input for the deep learning convolutional neural network, feature recognition is performed directly, and real-time processing is performed to obtain the crystallographic parameters of different locations of the polycrystalline functional thin film. When the local microstructure image captured in step 7 is used as data input for the deep learning convolutional neural network, the entire image is first saved and the image grid is segmented. The grid images at different locations in the image are then sequentially used as data input for feature recognition to obtain the crystallographic parameters of different locations of the polycrystalline functional thin film.

[0017] In the method for machine learning-assisted identification of the crystal phase distribution of a polycrystalline functional film in a nanodevice, in step 1, the constructed crystal structure model is a phase structure of the polycrystalline functional film with the same components but different states.

[0018] In the method for machine learning-assisted identification of the crystal phase distribution of polycrystalline functional films in nanodevices, in step 2, the simulation parameters that need to be adjusted include acceleration voltage, spherical aberration, chromatic aberration, astigmatism, defocus, exposure time, image size, and signal-to-noise ratio.

[0019] In the method for machine learning-assisted identification of the crystal phase distribution of a polycrystalline functional film in a nanodevice, in step 3, the simulated structure is a phase structure of the polycrystalline functional film with the same components but different states.

[0020] In the method for machine learning-assisted identification of the crystal phase distribution of polycrystalline functional thin films in nanodevices, in step 3, the simulated sample thickness is 1-30 nanometers.

[0021] In the method for machine learning-assisted identification of the crystal phase distribution of polycrystalline functional thin films in nanodevices, in step 3, the simulated image can be a high-resolution image, a high-angle annular dark-field image, annular bright-field image, a convergent beam diffraction image, a nanobeam diffraction image, and a Ronchigram.

[0022] In the method for machine learning-assisted identification of the crystal phase distribution of polycrystalline functional thin films in nanodevices, in step 4, the crystal phase parameters that require image recognition training include crystal structure, thickness, and spatial orientation.

[0023] In the method for machine learning-assisted identification of the crystal phase distribution of polycrystalline functional thin films in nanodevices, in step 4, the deep learning convolutional neural network constructed can be an AlexNet, ResNet or VGG model.

[0024] The method for machine learning-assisted identification of the crystal phase distribution of polycrystalline functional films in nanodevices, in step 8, the imaging mode images of the overall microstructure image as data input include convergent beam diffraction images, nanobeam diffraction images and Ronchigrams at different scanning positions in the transmission electron microscope scanning transmission mode.

[0025] In the method for machine learning-assisted identification of the crystal phase distribution of polycrystalline functional thin films in nanodevices, in step 8, the imaging mode images of the local microstructure image as data input include high-resolution image, high-angle annular dark field image and annular bright field image.

[0026] Compared with the prior art, the beneficial effects obtained by the present invention are as follows: a series of simulated images are automatically obtained as database files by traversing the simulation parameters in the transmission electron microscope simulation software, and this process does not require frequent manual modification of the simulation parameters. Based on the convergent beam diffraction, nanobeam diffraction and Ranchi diagram as the input data for image recognition, the crystallographic parameters of different positions can be processed in real time, and the accuracy depends on the step size of the electron probe, which can reach the atomic scale; based on the grid image of HREM, HAADF or ABF as the input data for image recognition, there is no need to perform data conversion operations such as Fourier transform, and the phase structure is distinguished by directly comparing the arrangement rules of the atomic columns in the local area, so it has nanometer or even atomic level resolution. Machine learning is used to construct a deep learning convolutional neural network for various crystal phase parameters such as crystal structure, crystal orientation, thickness, etc., and the transmission electron microscope images such as HREM, HAADF or ABF of the simulated polycrystalline functional material film are used as the data set to complete the training of the neural network, replacing manual means to achieve automatic, rapid and reliable recognition of the crystal phase of nanocrystals inside the polycrystalline functional film.

[0027] Deep learning convolutional neural networks are constructed for the crystallographic parameters of polycrystalline functional films, such as the crystal structure, tilt direction, and sample thickness. The obtained microscopic simulated HREM, HAADF, or ABF images are used as database files to train the neural network for feature recognition of atomic arrangement rules. After the training is completed, the phase structure of the actual film microstructure images taken by TEM can be quickly and reliably automatically calibrated to obtain the crystal phase distribution inside the film. This invention will be helpful in guiding the subsequent optimization work of nanodevices more quickly. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Atomic structure models of different crystal structures of polycrystalline functional films;

[0029] Figure 2 It is the parameter setting interface of qstem image simulation software;

[0030] Figure 3 Typical HAADF images of different orientations simulated by qstem image simulation software;

[0031] Figure 4 This is the workflow diagram of the deep learning convolutional neural network AlexNet;

[0032] Figure 5 Schematic diagram of the device of hafnium zirconium oxide ferroelectric capacitor;

[0033] Figure 6 Schematic diagram of continuous photography of hafnium zirconium oxide ferroelectric thin films under a transmission electron microscope;

[0034] Figure 7 The distribution of the grains to be tested in the hafnium zirconium oxide ferroelectric film; DETAILED DESCRIPTION

[0035] The present invention is described in detail below with reference to specific embodiments.

[0036] See Figure 1-Figure 7 The present invention proposes a method for machine learning-assisted identification of the crystal phase distribution of polycrystalline functional thin films in nanodevices. By automatically adjusting the simulation parameters of the transmission electron microscope simulation software, a series of simulated transmission electron microscope images of nano-polycrystalline functional thin films are obtained as a database file. Deep learning convolutional neural networks are constructed based on the crystal structure, tilt direction, sample thickness and other crystallographic parameters of the polycrystalline functional thin films to simulate the neural network trained on the transmission electron microscope image database. After the neural network training is completed, the crystal phase distribution is identified by transmission electron microscope images of grains at different locations in the actual polycrystalline functional thin film. The method comprises the following steps:

[0037] Step 1: According to the atomic arrangement of anions and cations in three-dimensional space in the three crystal structures of hafnium zirconium oxide (HZO) ferroelectric material, such as orthorhombic phase (O), tetragonal phase (T) and monoclinic phase (M), build the corresponding atomic model. Figure 1 shown.

[0038] Step 2: Import the atomic model created in step 1 into the open source Qstem software for transmission electron microscope image simulation, and adjust the simulation parameters (accelerating voltage, spherical aberration, chromatic aberration, astigmatism, defocus, exposure time, image size, signal-to-noise ratio) according to the actual transmission electron microscope shooting conditions. The parameter setting interface is as follows: Figure 2 shown.

[0039] Step 3: Based on the Python programming environment, the qstem simulation software is used as the development object. The parameters such as the tilt direction (1, φ, θ) and sample thickness (1-30 nm) expressed in the spherical polar coordinate system in the three-dimensional space are traversed in the qstem software with a set step size for different structures (O phase, T phase and M phase), and the subsequent transmission electron microscope image simulation operation process is performed. A series of high-angle annular dark field (HAADF) simulation images are obtained and automatically saved. The typical HAADF image of hafnium zirconium oxide material is shown as follows: Figure 3 shown.

[0040] Step 4: Construct three deep learning convolutional neural networks AlexNet in parallel based on the crystal structure, tilt direction, and sample thickness of the hafnium zirconium oxide material. The corresponding network workflow diagram is as follows: Figure 4 As shown. Using the series of HAADF simulation images obtained in step 3 as the label data input set, the image features of the three structures (O phase, T phase and M phase) in the local area are extracted through machine learning. The probabilities of the three structures of O phase, T phase and M phase are output in the Softmax layer of the neural network, and the error is obtained according to the output value and the input label value. The error feedback algorithm is used to update the weights of the neural network and train the neural network again until the error value is less than the set accuracy to complete the recognition training of the neural network. Finally, the HAADF image label data set is used to complete the recognition training of the tilt direction and sample thickness neural network.

[0041] Step 5: Use mature atomic layer deposition process or pulsed laser deposition method to deposit bottom electrode, hafnium zirconium oxide thin film and top electrode on silicon wafer substrate in sequence, and then process them through standard semiconductor process to obtain hafnium zirconium oxide ferroelectric capacitor. The thickness of hafnium zirconium oxide thin film is 1-20 nanometers. The schematic diagram of capacitor structure is shown in the figure below. Figure 5 shown.

[0042] Step 6: Use a commercial ferroelectric tester to test and verify the ferroelectric properties of the HfZrO ferroelectric capacitor.

[0043] Step 7: Use standard focused ion beam processing to prepare the hafnium zirconium oxide ferroelectric capacitor into a nano-thin sheet with a thickness of less than 20 nanometers.

[0044] Step 8: The hafnium zirconium oxide nanosheets prepared in step 7 are observed under a transmission electron microscope. After adjusting the electron microscope state (acceleration voltage, magnification, exposure time, and image size), microstructure images of the hafnium zirconium oxide thin film at different positions are continuously taken in the transmission electron microscope HAADF imaging mode under the same shooting conditions, such as Figure 6 shown.

[0045] Step 9: Save the HAADF image obtained in step 8 and further perform grid segmentation on the image, such as Figure 7 As shown, the grid images at different positions in the picture are input as the images to be tested into the deep learning convolutional neural network trained in step 4 to perform feature recognition of crystallographic parameters, and the film thickness, crystal structure and crystal orientation of different positions in hafnium zirconium oxide are obtained.

[0046] In summary, this embodiment proposes a method for machine learning-assisted identification of the crystalline phase distribution of polycrystalline functional thin films in nanodevices. By automatically adjusting the simulation parameters of the transmission electron microscope simulation software, a series of transmission electron microscope images of polycrystalline functional materials are generated as a database. A deep learning convolutional neural network is constructed based on the crystal structure, tilt direction, sample thickness, and other crystallographic parameters of the actual polycrystalline functional materials. The neural network constructed using the simulated image database is then used to perform feature analysis on transmission electron microscope images of grains at different locations in the actual polycrystalline functional material thin film, completing the task of automatically identifying the crystalline phase distribution.

[0047] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A method for machine learning to identify the crystal phase distribution of polycrystalline thin films in nanodevices, characterized in that: The following steps are involved: Step 1: Build a corresponding atomic model based on the atomic arrangement of the constituent atoms in different crystal structures of the polycrystalline functional film in three-dimensional space; Step 2: Import the atomic model created in step 1 into commercial or open-source transmission electron microscopy image simulation software and adjust the parameters according to the actual transmission electron microscopy imaging conditions; Step 3: Based on a mainstream computer system programming software development platform, the image simulation software used in Step 2 is used as the development object. For different crystal structures, the image simulation software is used to traverse the tilt direction (1, φ, θ) and sample thickness parameters represented by the three-dimensional spherical coordinate system with a set step size, and execute the transmission electron microscope image simulation operation process. A series of transmission electron microscope simulation images under different transmission electron microscope shooting modes are obtained and automatically saved; Step 4: Construct N corresponding deep learning convolutional neural networks in parallel for the N crystal phase parameters of the polycrystalline functional film; use the series of transmission electron microscope simulated images obtained in step 3 as the label data input set of a certain crystal phase parameter neural network, and extract the image features of the M sub-category crystal phase parameters in the local area through machine learning; output the probability of different sub-category crystal phase parameters in the total category M in the softmax layer of the neural network, and obtain the error based on the output value and the input label value; use the back propagation of the error to update the weight of the neural network and train the neural network again until the error value is less than the set accuracy to complete the recognition training of the neural network; finally, use the same simulated image label data set to complete the recognition training of other crystal phase parameter neural networks; Step 5: Use standard semiconductor processes to prepare nanodevices, such as phase change memory and ferroelectric memory, where the thickness of the polycrystalline functional film is less than 30 nanometers, and use an electrical test system to verify its performance; Step 6: Use standard focused ion beam processing technology to prepare the device with good performance verified in step 5 into nanosheets; Step 7: Observe the nanosheet prepared in step 6 under a transmission electron microscope. After adjusting the state of the electron microscope, select a suitable imaging mode to continuously capture microstructure images of different positions of the polycrystalline functional film. Step 8: When the overall microstructure image taken in step 7 is used as the data input of the deep learning convolutional neural network, feature recognition is performed directly, and real-time processing is performed to give the crystallographic parameters of different positions of the polycrystalline functional film; when the local microstructure image taken in step 7 is used as the data input of the deep learning convolutional neural network, the data of the entire image is first saved and the image grid is segmented, and then the grid images of different positions in the image are used as data input in turn, and feature recognition is performed to obtain the crystallographic parameters of different positions of the polycrystalline functional film.

2. The method according to claim 1, characterized in that In the step 1, the crystal structure model constructed is a phase structure of the same component but different states of the polycrystalline functional thin film.

3. The method according to claim 1, characterized in that In step 2, the simulation parameters that need to be adjusted include acceleration voltage, spherical aberration, chromatic aberration, astigmatism, defocus, exposure time, image size, and signal-to-noise ratio.

4. The method according to claim 1, wherein In step 3, the simulated structure is a phase structure of the polycrystalline functional thin film with the same components but different states.

5. The method according to claim 1, wherein In step 3, the simulated sample thickness is 1-30 nanometers.

6. The method according to claim 1, characterized in that In step 3, the images obtained by simulation are high-resolution images, high-angle annular dark-field images, annular bright-field images, convergent beam diffraction images, nanobeam diffraction images and Ronchigrams.

7. The method according to claim 1, characterized in that In step 4, the crystal phase parameters that need to be trained for image recognition include crystal structure, thickness and spatial orientation.

8. The method according to claim 1, characterized in that In step 4, the deep learning convolutional neural network constructed is an AlexNet, ResNet or VGG model.

9. The method according to claim 1, characterized in that In step 8, the imaging mode images of the overall microstructure image as data input include convergent beam diffraction images, nanobeam diffraction images and Ronchigrams at different scanning positions in the transmission electron microscope scanning transmission mode.

10. The method according to claim 1, characterized in that In step 8, the imaging mode images of the local microstructure image as data input include high-resolution image, high-angle annular dark field image and annular bright field image.

Images