Construction method of chalcogenide phase change storage material potential model, material characteristic calculation method and electronic equipment

Abstract

The invention discloses a chalcogenide phase change storage material potential model construction method, a material characteristic calculation method and electronic equipment, and belongs to the field of semiconductor materials. On the basis of a potential model trained by a first training set collected through molecular dynamics simulation based on a first principle calculation method, a current potential model is fully utilized, molecular dynamics simulation is carried out based on the current potential model, training samples are further collected, representative training samples are screened out, and a second training set is formed; according to the method, the training efficiency is high, the generalization ability is high, and a large amount of training data can be obtained with low cost. Meanwhile, verification of physical properties is introduced into convergence conditions of model training, so that the physical consistency and interpretability of the potential model are improved; on the basis, the training set which is good in diversity and sufficient in data volume can be obtained with low cost, and meanwhile the accuracy of the potential model in an actual application scene is further improved.

Description

Technical Field

[0001] The present invention belongs to the field of semiconductor materials, and more specifically, relates to a method for constructing a potential model of a chalcogenide phase change storage material, a method for calculating material properties, and electronic equipment. Background Art

[0002] In recent years, the ever-increasing demand for storage and computing has driven the exploration and optimization of novel phase-change memory materials. Phase-change memory (PCRAM), an emerging non-volatile memory technology, is widely considered a strong contender for next-generation memory due to its rapid phase transitions, ultra-low power consumption, excellent stability, and long cycle life. To achieve better performance, constructing potential models of phase-change memory materials is of significant significance in device exploration and mechanism research.

[0003] The evolution of artificial intelligence technology has further transformed the field of research and development. In particular, the use of machine learning to fit large-scale ab initio molecular dynamics (AIMD) databases to train potential models has become a very promising method that can significantly expand the spatial and temporal scales of molecular dynamics simulations while maintaining quantum mechanical accuracy.

[0004] Constructing a potential model through learning and training requires a large number of diverse training samples and labels. Existing technologies often rely on first-principles calculations to obtain a large number of training samples and labels. However, using first-principles calculations to obtain a large number of training samples and labels is computationally intensive, time-consuming, and labor-intensive, with high both time and economic costs. Furthermore, the data is often redundant, with weak specificity between samples, and often fails to cover rare or extreme states in the structural space (such as defects, crystallization transition states, etc.), resulting in low training efficiency and weak generalization of the potential model. Furthermore, existing technologies often only consider whether the output error of the potential model converges when training the potential model, while ignoring physical rationality. This results in the potential model having a high accuracy in output evaluation in scientific research and application fields, but when solving actual physical problems, the potential model is prone to collapse or derives a non-physical structure. These problems not only increase the cost of investment in the training set, but also easily lead to some erroneous conclusions in the process of material research and development. Summary of the Invention

[0005] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a method for constructing a potential model of sulfur-based phase change storage materials, a material property calculation method, and an electronic device. The purpose is to obtain a training set with good diversity and sufficient data volume at a relatively low cost, while further improving the accuracy of the potential model in actual application scenarios.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for constructing a potential model of a chalcogenide phase change memory material, comprising:

[0007] S1. Obtaining an initial potential model of the chalcogenide phase change memory material as the current potential model; wherein the initial potential model is trained on a first training set; the first training set is obtained by performing picosecond molecular dynamics simulations of the chalcogenide phase change memory material using a first principles calculation method;

[0008] S2. Based on the current potential model, nanosecond-level molecular dynamics simulations are performed on the chalcogenide phase change memory material under multiple sets of different influencing parameters to obtain the physical configurations of the chalcogenide phase change memory material under each set of influencing parameters, forming a first physical configuration set; each set of influencing parameters includes: phase state and external environment parameters;

[0009] S3. Filtering out a plurality of candidate physical configurations from the first physical configuration set, which are non-outliers and have a distance in the potential model feature space greater than or equal to a first preset distance;

[0010] S4. Perform single-point energy calculations on each candidate physical configuration to obtain the corresponding potential energy surface and force field, thereby constructing the second training set;

[0011] S5. Using the second training set to train the current potential model, or using the total of the second training set and the first training set to train the current potential model;

[0012] S6. Calculate the potential energy surface output error and force field output error of the current potential model based on the pre-collected validation set. If the output errors are both less than the corresponding preset output errors, proceed to S7; otherwise, adjust the screening parameters and proceed to S3; wherein the screening parameters include: a first preset distance; the validation set includes: a physical configuration and the corresponding real potential energy surface and force field;

[0013] S7, verifying the physical properties of the chalcogenide phase change memory material based on the current potential model; if the verification passes, go to S8; otherwise, adjust the influencing parameters in S2 and go to S2;

[0014] S8. Output the current potential model as the final potential model.

[0015] Further preferably, the verification of the physical properties of the chalcogenide phase change memory material based on the current potential model includes:

[0016] A1. Based on the current potential model, multiple molecular dynamics simulations are performed on each type of physical property of the chalcogenide phase change memory material to obtain the physical characteristics of the chalcogenide phase change memory material. The physical characteristics include: characteristic data reflecting each type of physical property of the chalcogenide phase change memory material;

[0017] A2. When the errors between all feature data in the physical features and the corresponding preset reference feature data are less than the corresponding preset errors, the verification is determined to be passed; otherwise, it is failed.

[0018] Further preferably, the above-mentioned adjustment of the influencing parameters in S2 includes: adjusting one or more groups of influencing parameters in S2 to an influencing parameter group that can reflect the physical property X of the sulfur-based phase change memory material during the molecular dynamics simulation process, so as to increase the proportion of physical configurations that can reflect the physical property X in the first physical configuration set obtained after re-executing S2; wherein the physical property X is a physical property category that has not passed verification.

[0019] Further preferably, the above-mentioned physical characteristics include: local structural fingerprint characteristics of the chalcogenide phase change memory material;

[0020] The local structural fingerprint features include one or more of the SOAP descriptor, local bond orientation order parameter and coordination number of the chalcogenide phase change memory material.

[0021] Further preferably, the local structural fingerprint feature is a vector consisting of the SOAP descriptor, the local bond orientation order parameter and the coordination number of the chalcogenide phase change memory material.

[0022] Further preferably, the above physical characteristics also include: energy-volume curve, ring distribution, melting point and crystallization rate of the sulfur-based phase change memory material.

[0023] Further preferably, the method for determining an outlier of a physical configuration includes: calculating an outlier value of the physical configuration, and when the outlier value is greater than a preset threshold, determining that the physical configuration is an outlier; otherwise, determining that the physical configuration is not an outlier;

[0024] The outlier calculation formula is ||E-E'||; where E is the potential energy surface obtained by inputting the physical configuration into the current potential model; E' is the potential energy surface obtained by performing single-point energy calculation on the physical configuration;

[0025] Alternatively, the outlier value is calculated as Where F is the force field obtained by inputting the physical configuration into the current potential model; F' is the force field obtained by performing single-point energy calculations on the physical configuration; and σ(F) is the standard deviation of the force field for each atom of the physical configuration contained in the force field F of the physical configuration.

[0026] Further preferably, the initial potential model of the above-mentioned sulfur-based phase change memory material is constructed by the following method: using the first training set to train the machine learning model to obtain the initial potential model.

[0027] Further preferably, the first training set is obtained by:

[0028] The physical configurations of the sulfur-based phase change memory material under different phase states and external environmental parameters are obtained to form a second physical configuration set; the second physical configuration set includes: a crystalline configuration subset and an amorphous configuration subset; wherein, the amorphous configurations and the corresponding potential energy surfaces and force fields in the amorphous configuration subset are obtained by performing picosecond molecular dynamics simulations of the sulfur-based phase change memory material in the amorphous state under different external environmental parameters based on the first-principles calculation method; the crystalline configuration subset includes: the crystalline configuration of the sulfur-based phase change memory material and the crystalline configuration after data enhancement of the crystalline configuration of the sulfur-based phase change memory material; the crystalline configuration is obtained by directly obtaining it from a material crystalline configuration library and constructing it using atomic simulation software for the sulfur-based phase change memory material; wherein, the data enhancement includes: adding defects, adding vacancies and adding strain to the crystalline configuration.

[0029] screening the physical configurations in the second set of physical configurations to retain physical configurations whose physical features are separated by a distance greater than or equal to a second preset distance;

[0030] Perform single-point energy calculations on each crystal configuration in the screened second physical configuration set to obtain the corresponding potential energy surface and force field;

[0031] The physical configurations in the screened second physical configuration set are used as training samples, and the corresponding potential energy surfaces and force fields are used as labels to construct the first training set.

[0032] In a second aspect, the present invention provides a method for calculating the characteristics of a chalcogenide phase change memory material, comprising:

[0033] Inputting the physical configuration of the chalcogenide phase change memory material to be calculated into the potential model to obtain the potential energy surface and force field of the chalcogenide phase change memory material;

[0034] The potential model is constructed using the construction method provided in the first aspect of the present invention.

[0035] In a third aspect, the present invention provides an electronic device comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the method provided in the first aspect or the second aspect of the present invention when executing the computer program.

[0036] In a third aspect, the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method provided in the first aspect or the second aspect of the present invention.

[0037] In a fourth aspect, the invention further provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the method provided in the first or second aspect of the invention.

[0038] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0039] 1. The present invention provides a method for constructing a potential model for a sulfur-based phase-change memory material. Based on a potential model trained using a first training set collected through molecular dynamics simulations based on first-principles calculations, the method fully utilizes the current potential model, performs molecular dynamics simulations based on the current potential model, further collects training samples, and selects representative training samples to form a second training set with high diversity and low redundancy for further training the potential model. This method has high training efficiency and strong generalization capabilities. Furthermore, the acquisition of the second training set does not rely on computationally complex first-principles calculations, enabling the acquisition of a large amount of training data at a low cost. Furthermore, physical property verification is incorporated into the convergence conditions for model training, thereby enhancing the physical consistency and interpretability of the potential model, ensuring not only high accuracy in output results but also that its physical behavior reflects actual physical properties. Consequently, the present invention can obtain a training set with high diversity and sufficient data at a low cost, while further improving the accuracy of the potential model in practical application scenarios.

[0040] 2. Furthermore, the method for constructing a potential model of sulfur-based phase change storage materials provided by the present invention, when the verification of the physical properties fails, increases the proportion of physical configurations that can reflect the physical properties of the category that failed the verification in the first physical configuration set obtained after re-executing S2, thereby further refining the abundance of the data set and the expressive power of the potential model, and further improving the accuracy of the potential model in actual application scenarios.

[0041] 3. Furthermore, in the method for constructing a potential model of a chalcogenide phase-change memory material provided by the present invention, the physical characteristics include local structural fingerprint characteristics of the chalcogenide phase-change memory material; the local structural fingerprint characteristics include one or more of the SOAP descriptor, local bond orientation order parameter, and coordination number of the chalcogenide phase-change memory material. The SOAP descriptor captures continuous density variation characteristics, the local bond orientation order parameter distinguishes local symmetry, and the coordination number distinguishes inter-class topology. Any one of the three can more accurately reflect the physical property of the short-range structure of the chalcogenide phase-change memory material, while multiple of the three can further complement each other in terms of physical meaning and numerical stability, providing a more robust short-range structural representation.

[0042] 4. Furthermore, in the method for constructing the potential model of the sulfur-based phase change storage material provided by the present invention, when obtaining the first training set, the crystalline configuration subset also includes crystalline configurations that have undergone data enhancement. One or more operations of adding defects, adding vacancies, and adding strain can further improve the diversity of the data, thereby further improving the accuracy of the initial potential model. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flow chart of a method for constructing a potential model of a chalcogenide phase change memory material provided by an embodiment of the present invention;

[0044] Figure 2 Schematic diagram of the large-scale amorphous structure of GeTe provided in an embodiment of the present invention;

[0045] Figure 3 A schematic diagram of farthest point sampling of a data set using GeTe amorphous configuration as an example provided by an embodiment of the present invention;

[0046] Figure 4 Schematic diagram of energy field and force field RMSE verification of the potential model using amorphous GeTe as an example provided in an embodiment of the present invention;

[0047] Figure 5 A schematic diagram of verifying the physical properties of a potential model based on the coordination number distribution function using amorphous GeTe as an example provided in an embodiment of the present invention;

[0048] Figure 6 A schematic diagram of verifying the physical properties of a potential model based on an energy-volume curve using amorphous GeTe as an example provided in an embodiment of the present invention;

[0049] Figure 7 A schematic diagram of verifying the physical properties of a potential model based on ring distribution using amorphous GeTe as an example provided in an embodiment of the present invention;

[0050] Figure 8 A schematic diagram of verifying the physical properties of a potential model based on melting point using amorphous GeTe as an example provided in an embodiment of the present invention;

[0051] Figure 9 A schematic diagram of amorphous GeTe used as an example to verify the physical properties of the potential model based on the crystallization rate provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0053] To achieve the above objectives, in a first aspect, the present invention provides a method for constructing a potential model of a chalcogenide phase change memory material, comprising:

[0054] S1. Obtaining an initial potential model of the chalcogenide phase change memory material as a current potential model;

[0055] Specifically, the initial potential model can be an existing trained initial potential model, such as Deepmd-kit, ACE, GAP, Deep Potential Model (Deep Potential), Neuroevolution Potential Model (Neuroevolution Potential, NEP) and other potential function models. The present invention further trains and optimizes on the basis of it, and can also be trained on the first training set. Specifically, the first training set is used to train the machine learning model to obtain the initial potential model. It should be noted that the above-mentioned machine learning model can be a traditional machine learning model such as SVM, or a deep learning model such as a neural network model and Transformer, which is not limited here.

[0056] The first training set is obtained by performing picosecond molecular dynamics simulations of chalcogenide phase change memory materials using first-principles calculations. Specifically, in one optional embodiment, first-principles molecular dynamics simulations of chalcogenide phase change memory materials are performed using first-principles calculation software such as VASP software.

[0057] Preferably, in an optional implementation manner, the first training set is obtained by:

[0058] The physical configurations of the chalcogenide phase change memory material under different phase states and external environmental parameters are obtained to form a second physical configuration set; the second physical configuration set includes: a crystalline configuration subset and an amorphous configuration subset; wherein, the amorphous configurations and the corresponding potential energy surfaces and force fields in the amorphous configuration subset are obtained by performing picosecond molecular dynamics simulations of the chalcogenide phase change memory material in the amorphous state under different external environmental parameters based on the first-principles calculation method; taking the chalcogenide phase change memory material as GeTe as an example, Figure 2 The diagram shows a large-scale amorphous structure of GeTe (4096 atoms). The external environmental parameters include one or more of temperature (or temperature gradient) and pressure.

[0059] The crystalline configuration subset includes: the crystalline configuration of the chalcogenide phase change memory material and the crystalline configuration of the chalcogenide phase change memory material after data enhancement; the crystalline configuration is obtained by directly obtaining it from the material crystalline configuration library and constructing it using atomic simulation software for the chalcogenide phase change memory material; wherein the data enhancement includes: adding one or more operations of defects, adding vacancies and adding strain to the crystalline configuration;

[0060] The physical configurations in the second physical configuration set are screened to retain physical configurations whose physical features are greater than or equal to a second preset distance; wherein the preset distance is an empirical value, and in an optional embodiment, the second preset distance is 5% of the physical feature space.

[0061] Perform single-point energy calculations on each crystal configuration in the screened second physical configuration set to obtain the corresponding potential energy surface and force field;

[0062] The physical configurations in the screened second physical configuration set are used as training samples, and the corresponding potential energy surfaces and force fields are used as labels to construct the first training set.

[0063] S2. Based on the current potential model, performing nanosecond molecular dynamics simulations on the chalcogenide phase change memory material under multiple sets of different influencing parameters to obtain physical configurations of the chalcogenide phase change memory material under each set of influencing parameters, constituting a first physical configuration set; each set of influencing parameters includes: phase state and external environmental parameters; wherein the external environmental parameters include: one or more of: temperature (or temperature gradient) and pressure;

[0064] Specifically, in an optional embodiment, the current potential model is used in combination with molecular dynamics simulation software such as Lammps or GPUMD to perform molecular dynamics simulation on the sulfur-based phase change memory material based on the current potential model. The specific process includes: taking a certain physical configuration of the sulfur-based phase change memory material as its initial physical configuration, inputting the initial physical configuration into the current potential model to obtain the corresponding potential energy surface and force field, inputting the obtained potential energy surface and force field and each influencing parameter group into the molecular dynamics simulation software to obtain the physical configuration of the sulfur-based phase change memory material under the corresponding influencing parameter group. Among them, the certain physical configuration of the sulfur-based phase change memory material is the physical configuration of the sulfur-based phase change memory material that has been collected or calculated in advance.

[0065] In one alternative implementation, the current potential model file is placed into the simulation project and confirmed to contain the correct network parameters and descriptor configuration. In molecular dynamics simulation software such as GPUMD or LAMMPS, the potential model, influencing parameter set, time step, and ensemble type (e.g., NVT / NPT) are specified, along with an appropriate output frequency. During the simulation, trajectory files are periodically output in a common format (e.g., XYZ, LAMMPS dump, exyz), and the second physical configuration set is sampled from the trajectory file.

[0066] It should be noted that the present invention can construct a corresponding potential model for a specific target chalcogenide phase-change memory material. In this case, the chalcogenide phase-change memory material in the present invention is fixed to the target chalcogenide phase-change memory material. Furthermore, the present invention can also construct a corresponding potential model for a class of chalcogenide phase-change memory materials. In this case, to improve the generalizability of the model, the chalcogenide phase-change memory materials involved in the first physical configuration set typically include chalcogenide phase-change memory materials of different compositions.

[0067] It should be noted that the sulfide phase change memory materials include binary or ternary compounds whose anions are S, Se, and Te and whose cations are from the third, fourth, and fifth main groups.

[0068] S3. Filter out multiple candidate physical configurations from the first physical configuration set that are non-outliers and have a distance in the potential model feature space greater than or equal to a first preset distance; wherein the preset distance is an empirical value. In an optional implementation, the first preset distance is 5% of the potential model feature space.

[0069] In an optional embodiment, the method for determining an outlier of a physical configuration includes: calculating an outlier value of the physical configuration; when the outlier value is greater than a preset threshold, determining that the physical configuration is an outlier; otherwise, determining that the physical configuration is a non-outlier; wherein the preset threshold is an empirical value.

[0070] The outlier calculation formula is ||E-E'||; where E is the potential energy surface obtained by inputting the physical configuration into the current potential model; E' is the potential energy surface obtained by performing single-point energy calculation on the physical configuration; in this case, in an optional embodiment, the preset threshold is 100 meV atom -1 .

[0071] Alternatively, the outlier value is calculated as Where F is the force field obtained by inputting the physical configuration into the current potential model; F' is the force field obtained by calculating the single point energy of the physical configuration; σ(F) is the standard deviation of the force field of each atom of the physical configuration contained in the force field F of the physical configuration. In this case, in an optional embodiment, the preset threshold is

[0072] In an optional implementation, a farthest point sampling method is used to screen multiple candidate physical configurations whose distance in the potential model feature space is greater than or equal to a first preset distance.

[0073] S4. Perform single-point energy calculations on each candidate physical configuration to obtain the corresponding potential energy surface and force field; construct a second training set using the candidate physical configurations as training samples and the corresponding potential energy surface and force field as labels;

[0074] It should be noted that single-point energy calculations can be obtained using first-principles calculation software such as VASP software.

[0075] In an optional embodiment, data enhancement is also performed on some or all of the physical configurations in the second training set; data enhancement includes: adding defects, adding vacancies, and adding strain (such as stretching, perturbation, etc.) to the crystalline configuration to further improve the diversity of the data set.

[0076] S5. Using the second training set to train the current potential model, or using the total of the second training set and the first training set to train the current potential model;

[0077] It should be noted that there is no limitation on the training method. It can be retraining, incremental training or fine-tuning.

[0078] S6. Calculate the potential energy surface output error and force field output error of the current potential model based on the pre-collected validation set. If the output errors are both less than the corresponding preset output errors, proceed to S7; otherwise, adjust the screening parameters and proceed to S3; wherein the screening parameters include: a first preset distance; the validation set includes: a physical configuration and the corresponding real potential energy surface and force field;

[0079] In an optional implementation manner, the screening parameters may further include: the above-mentioned preset threshold, parameters in the farthest point sampling method, etc., which are not limited here.

[0080] It should be noted that the output error can be measured in a variety of ways, such as mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), etc., which are not limited here.

[0081] Preferably, the preset output error corresponding to the potential energy surface and the force field is an empirical value. In an optional embodiment, the preset output error corresponding to the potential energy surface is 15 meV atom -1 ; The preset output error corresponding to the force field is

[0082] S7, verifying the physical properties of the chalcogenide phase change memory material based on the current potential model; if the verification passes, go to S8; otherwise, adjust the influencing parameters in S2 and go to S2;

[0083] In an optional embodiment, the verification of the physical properties of the chalcogenide phase change memory material based on the current potential model includes:

[0084] A1. Based on the current potential model, multiple molecular dynamics simulations are performed on each type of physical property of the chalcogenide phase change memory material to obtain the physical characteristics of the chalcogenide phase change memory material. The physical characteristics include: characteristic data reflecting each type of physical property of the chalcogenide phase change memory material;

[0085] Similarly, if the present invention constructs a corresponding potential model for a specific target chalcogenide phase change memory material, the physical properties of the target chalcogenide phase change memory material can be verified. If the present invention constructs a corresponding potential model for a class of chalcogenide phase change memory materials, the physical properties of one or more chalcogenide phase change memory materials can be verified.

[0086] A2. If the error between all characteristic data in the physical characteristics and the corresponding preset baseline characteristic data is less than the corresponding preset error, the verification is considered passed; otherwise, it is considered failed. The preset baseline characteristic data are empirical values and can be obtained from literature, experimentally, or based on first-principles calculations, using characteristic data of corresponding physical properties collected by performing multiple molecular dynamics simulations similar to A1 for each type of physical property of the chalcogenide phase change memory material. This is not limited here.

[0087] Preferably, the above-mentioned adjustment of the influencing parameters in S2 includes: adjusting one or more groups of influencing parameters in S2 to an influencing parameter group that can reflect the physical property X of the sulfur-based phase change memory material during the molecular dynamics simulation process, so as to increase the proportion of physical configurations that can reflect the physical property X in the first physical configuration set obtained after re-executing S2; wherein the physical property X is a physical property category that has not passed verification.

[0088] In an optional embodiment, the above-mentioned physical characteristics include: local structural fingerprint characteristics of the chalcogenide phase change memory material;

[0089] The local structural fingerprint features include: one or more of the SOAP descriptor, local bond orientation order parameter and coordination number of the chalcogenide phase change memory material;

[0090] The SOAP descriptor of the chalcogenide phase change memory material is the average value of the SOAP descriptors of each physical configuration obtained by performing multiple molecular dynamics simulations on the chalcogenide phase change memory material in A1 above;

[0091] The local bond orientation order parameter of the chalcogenide phase change memory material is the average value of the local bond orientation order parameters of each physical configuration obtained by performing multiple molecular dynamics simulations on the chalcogenide phase change memory material in A1 above;

[0092] The coordination number of the sulfur-based phase change memory material is the average of the coordination numbers of various physical configurations obtained by performing multiple molecular dynamics simulations on the sulfur-based phase change memory material in A1.

[0093] Preferably, the above physical characteristics further include: energy-volume curve, ring distribution, melting point and crystallization rate of the sulfur-based phase change memory material.

[0094] S8. Output the current potential model as the final potential model.

[0095] In order to further illustrate the method for constructing the potential model of the chalcogenide phase change memory material provided by the present invention, a detailed description is given below in conjunction with a specific embodiment:

[0096] like Figure 1 As shown, the method for constructing the chalcogenide phase change memory material potential model in this embodiment includes the following steps:

[0097] S1. Obtaining an initial potential model of the chalcogenide phase change memory material as the current potential model; wherein the initial potential model is trained on a first training set; the first training set is obtained by performing picosecond molecular dynamics simulations of the chalcogenide phase change memory material using a first principles calculation method;

[0098] The initial potential model is obtained through the following process:

[0099] Based on the first-principles calculation method (the first-principles calculation method in this embodiment is the DFT algorithm, which is implemented using VASP software), a molecular dynamics simulation of the chalcogenide phase change memory material in the amorphous state is performed at the picosecond level under different external environmental parameters to obtain the amorphous configurations under different external environmental parameters, constituting an amorphous configuration subset;

[0100] In this embodiment, the method for constructing the amorphous configuration subset of the chalcogenide phase change memory material is as follows:

[0101] 1) Using the NVT ensemble (fixed number of atoms, simulation unit volume, and temperature), based on first-principles calculation methods, a 6ps molecular dynamics simulation was performed at 2000K to melt the sulfur-based phase change memory material into a liquid, completely disrupting the original atomic arrangement.

[0102] 2) Molecular dynamics simulation was performed at a cooling rate of 283K / ps to cool the sulfur-based phase change memory material from 2000K to 300K to simulate the quenching process.

[0103] 3) The geometric structure of the amorphous structure obtained during the quenching process is optimized, and the interatomic stress is used as the iterative convergence standard. The convergence criterion is that the difference between the magnitude of the force on all atoms between two iterations is less than a preset force threshold (in this embodiment, the value is When the stress between all atoms in the entire unit cell is less than the preset stress (3 kbar in this embodiment), the structural optimization is considered complete, and a stable amorphous configuration is obtained. At the same time, the trajectory of the amorphous configuration is sampled during the construction process to obtain stable amorphous configurations at different temperatures, forming an amorphous configuration subset. In the above process, the potential energy surface and force field of each amorphous configuration in the amorphous configuration subset calculated by the first principles calculation method are recorded.

[0104] Furthermore, the amorphous configurations in the obtained amorphous configuration subset were subjected to molecular dynamics simulation at 300K, 500K, and 700K, respectively, to obtain the amorphous configurations at 300K, 500K, and 700K, and added to the amorphous configuration subset; wherein, the molecular dynamics simulation generates the trajectory of each step, that is, the coordinate information of all atoms at each step.

[0105] In this embodiment, the crystalline configuration subset of the chalcogenide phase change memory material includes: the crystalline configuration of the chalcogenide phase change memory material and the crystalline configuration of the chalcogenide phase change memory material after data enhancement; wherein the data enhancement includes: adding one or more operations of defects, adding vacancies, and adding strain to the crystalline configuration;

[0106] In this embodiment, the sulphur-based phase change memory material is constructed using atomic simulation software to obtain the corresponding crystalline configuration.

[0107] In this embodiment, molecular dynamics simulations are performed using molecular dynamics simulation software such as LAMMPS or GPUMD. This embodiment employs the atomistic simulation software ASE or GPUMD-Wizard to construct the crystalline structures of different components of the chalcogenide phase change memory material, such as mono- or binary FCC and BCC crystal structures. The resulting physical configuration data can be collected and organized using the dpdata tool.

[0108] The farthest point sampling method is used to screen the physical configurations in the second physical configuration set to retain physical configurations whose physical features are greater than or equal to a preset distance; the physical features in this embodiment are SOAP descriptors of the physical configurations; the preset distance in this embodiment is 5% of the SOAP descriptor feature space, specifically 0.05.

[0109] Perform single-point energy calculations on each crystal configuration in the screened second physical configuration set to obtain the corresponding potential energy surface and force field;

[0110] The physical configurations in the screened second physical configuration set are used as training samples, and the corresponding potential energy surfaces and force fields are used as labels to construct the first training set.

[0111] The first training set is used to train the neural network model to obtain an initial potential model.

[0112] S2. Based on the current potential model, nanosecond-level molecular dynamics simulations are performed on the chalcogenide phase change memory material under multiple sets of different influencing parameters to obtain the physical configurations of the chalcogenide phase change memory material under each set of influencing parameters, forming a first physical configuration set; each set of influencing parameters includes: phase state and external environment parameters;

[0113] In this embodiment, the current potential model is combined with molecular dynamics simulation software (such as Lammps or GPUMD) to perform long-time nanosecond Monte Carlo molecular dynamics simulations (MCMD) and temperature exploration simulations to obtain the physical configurations of the sulfur-based phase change memory material under various sets of influencing parameters, forming a first set of physical configurations. MCMD combines the global sampling capability of the MC method with the local fine evolution capability of the MD method, enabling more efficient exploration of the potential energy surface of complex systems. Temperature exploration refers to the use of the current potential model for low-cost, high-throughput calculations to obtain the long-term relaxation structures of various crystalline or amorphous configurations at different temperatures. Compared with first-principles calculations, this greatly reduces the time and resources consumed by computation.

[0114] Specifically, the current potential model is used for temperature exploration, constructing different temperature gradients from room temperature (300K) to the melting point (e.g., 1000K). Multiple independent, long-term simulations at the nanosecond level are then performed to sample a temperature-dependent dataset. MCMD utilizes Monte Carlo methods such as atom exchange and morphological transformation to prevent the system from falling into local energy minima, providing a more randomized structure for the dataset. The sampled dataset is then subjected to strain, perturbation, and defect addition operations using the open-source software packages ASE and dpdata, and then used for iterative model training.

[0115] S3. Filtering out a plurality of candidate physical configurations from the first set of physical configurations that are non-outliers and have a distance greater than or equal to a preset distance in the potential model feature space;

[0116] In this embodiment, the outlier value calculation formula is ||E-E'||, and the outlier value of each physical configuration in the first physical configuration set is calculated respectively. When the outlier value is greater than a preset threshold, the physical configuration is determined to be an outlier; otherwise, the physical configuration is determined to be a non-outlier; where E is the potential energy surface obtained by inputting the physical configuration into the current potential model; E' is the potential energy surface obtained by performing single-point energy calculation on the physical configuration;

[0117] In this embodiment, the farthest point sampling method is used to obtain multiple candidate physical configurations whose distance in the potential model feature space is greater than or equal to a preset distance. The preset distance in this embodiment is 0.05.

[0118] Through the above process, redundant physical configurations in the first physical configuration set can be deleted, while improving the expression capability.

[0119] like Figure 3 Shown is a schematic diagram of the farthest point sampling for a data set taking the amorphous configuration of GeTe as an example.

[0120] S4. Perform single-point energy calculations on each candidate physical configuration to obtain the corresponding potential energy surface and force field, thereby constructing a second training set with the candidate physical configuration as training samples and the corresponding potential energy surface and force field as labels;

[0121] S5. Using the second training set to train the current potential model, or using the total of the second training set and the first training set to train the current potential model;

[0122] S6. Calculate the potential energy surface output error and force field output error of the current potential model based on the pre-collected validation set. If the output errors are both less than the corresponding preset output errors, proceed to S7. Otherwise, adjust the screening parameters and proceed to S3. The screening parameters include: a preset distance; the validation set includes: a physical configuration and the corresponding real potential energy surface and force field.

[0123] In this embodiment, the output error is measured by RMSE.

[0124] S7, verifying the physical properties of the chalcogenide phase change memory material based on the current potential model; if the verification passes, go to S8; otherwise, adjust the influencing parameters in S2 and go to S2;

[0125] The above-mentioned verification of the physical properties of chalcogenide phase change memory materials based on the current potential model includes:

[0126] A1. Based on the current potential model, multiple molecular dynamics simulations are performed on each type of physical property of the chalcogenide phase change memory material to obtain the physical characteristics of the chalcogenide phase change memory material. The physical characteristics include: characteristic data reflecting each type of physical property of the chalcogenide phase change memory material;

[0127] A2. When the errors between all feature data in the physical features and the corresponding preset reference feature data are less than the corresponding preset errors, the verification is determined to be passed; otherwise, it is failed.

[0128] Based on the potential model, various physical properties of chalcogenide phase change memory materials (such as GeTe) such as structural characteristics, thermodynamic properties, and kinetic properties are extensively evaluated until the model can be tested and applied to long-term, large-scale simulation tasks such as crystallization. The process is applicable to a range of materials.

[0129] In this embodiment, the physical characteristics include: local structural fingerprint characteristics, energy-volume curve, ring distribution, melting point and crystallization rate of the chalcogenide phase change memory material.

[0130] 1) Local structural fingerprint features

[0131] The local structural fingerprint feature is used to accurately express and identify the physical property of short-range order difference in chalcogenide phase change memory materials, and can be one or more of the SOAP (Smooth Overlap of Atomic Positions) descriptor, the local bond orientation order parameter (Bond Orientation Order, BOO) and the coordination number (Coordination Number, CN), preferably all three. In this embodiment, the local structural fingerprint feature is a vector composed of the SOAP descriptor, the local bond orientation order parameter and the coordination number of the chalcogenide phase change memory material.

[0132] The SOAP descriptor of the chalcogenide phase change memory material is the average value of the SOAP descriptors of each physical configuration obtained by performing multiple molecular dynamics simulations on the chalcogenide phase change memory material in A1 above;

[0133] The local bond orientation order parameter of the chalcogenide phase change memory material is the average value of the local bond orientation order parameters of each physical configuration obtained by performing multiple molecular dynamics simulations on the chalcogenide phase change memory material in A1 above;

[0134] The coordination number of the sulfur-based phase change memory material is the average of the coordination numbers of various physical configurations obtained by performing multiple molecular dynamics simulations on the sulfur-based phase change memory material in A1.

[0135] The SOAP descriptor of the physical configuration is centered on the atom, and the distribution of neighboring atoms is represented as a superposition of Gaussian clouds. The invariant power spectrum is constructed by expanding the spherical harmonic function. For any atom in the physical configuration, the calculation formula of its SOAP descriptor is:

[0136]

[0137] Among them, p nn'l is the SOAP descriptor of an atom in a physical configuration with angular momentum quantum number l and the nth and n'th radial basis functions. n, n' are the indices of the radial basis functions. l is the angular momentum quantum number, which determines the order of the spherical harmonic expansion. m is the magnetic quantum number, -l≤m≤l. c nlm are the coefficients of the local density projection onto the radial basis and spherical harmonics after Gaussian broadening. Indicates c nlm The conjugate complex number of means that if c nlm =a+bj, then This is to ensure that the final result is a real number.

[0138] The SOAP descriptor of the physical configuration is composed of the SOAP descriptors of all atoms in the physical configuration. The SOAP descriptor is invariant to translation, rotation, and atomic arrangement, and is suitable for continuously expressing the difference in short-range order (SRO) between crystals and amorphous materials.

[0139] The local bond orientation order parameter is a scalar quantity of local order extracted by statistically analyzing the bond directions using spherical harmonic functions. For any atom in a physical configuration, the calculation formula for the local bond orientation order parameter is:

[0140]

[0141] Among them, Q l is the local bond orientation order parameter of an atom in the physical configuration under the angular momentum quantum number l and magnetic quantum number m; l represents the angular momentum quantum number, which determines the order of the spherical harmonic expansion and is usually 4 or 6; m is the magnetic quantum number, -l≤m≤l. b is the number of neighbor atoms of the atom. lm (θ, φ) represents the spherical harmonic function, whose inputs are two angles: θ is the polar angle, representing the angle between the atom's neighbor and the z-axis; φ is the azimuthal angle, representing the angle between the neighbor in the xy plane and the x-axis. q4 and q6 are commonly used to distinguish between tetrahedral, octahedral, and amorphous structures.

[0142] The local bond orientation order parameters of all atoms in the physical configuration constitute the local bond orientation order parameters of the physical configuration.

[0143] The coordination number can be used to distinguish local coordination configurations such as tri-, tetra-, and penta-coordinates, and is particularly sensitive to local polymorphism in amorphous states. The number of neighbors of the central atom is counted using Voronoi analysis or the cutoff radius method: For any atom in the physical configuration, the coordination number is calculated as:

[0144]

[0145] Among them, CN i is the coordination number of the i-th atom in the physical configuration; r ij is the distance between the i-th atom and the j-th atom in the physical configuration. c is the cutoff radius;

[0146] The coordination number of all atoms in a physical configuration constitutes the coordination number of the physical configuration.

[0147] In this embodiment, the SOAP descriptor is calculated for each atom in the physical configuration, and the vector dimension D can reach tens to hundreds of dimensions. The local bond orientation order parameters of the atom, such as Q4 and Q6, are added as an extended dimension (2D). The coordination number is added as a single-dimensional scalar. Finally, the structural fingerprint vector of each atom in the physical configuration is formed:

[0148]

[0149] The resulting fingerprint vector is normalized for all atoms (e.g., Z-score or min-max). This fingerprint design has good structural resolution and physical interpretability, making it suitable for describing local evolution between amorphous and crystalline states and for verifying molecular dynamics physical properties.

[0150] 2) Energy-volume curve

[0151] For each physical configuration obtained during the molecular dynamics simulation in A1, which may be amorphous, crystalline (hexagonal, cubic phase), etc., scale its volume, calculate the energy under each volume, and fit the Birch–Murnaghan formula The energy-volume curve of each physical configuration is obtained. Where E(V) is the energy under volume V, E0 is the energy under equilibrium state (minimum energy), V0 is the volume under equilibrium state (minimum energy), B0 is the bulk modulus (BulkModulus) in equilibrium state, B 0′ is the first derivative of the bulk modulus with respect to pressure (pressure derivative of the bulk modulus).

[0152] The energy-volume curve of each physical configuration obtained by performing multiple molecular dynamics simulations on the sulfur-based phase change memory material in A1 is averaged to obtain the energy-volume curve of the sulfur-based phase change memory material.

[0153] 3) Ring distribution (ring statistical analysis)

[0154] During the molecular dynamics simulations in A1, a relaxed structure of a stable amorphous material was constructed. Correspondingly, using first-principles calculations, a similar relaxed structure of a stable amorphous material was simulated and constructed, serving as the corresponding preset baseline characteristic data. The statistical results of the two structures were compared for the four-membered ring structure, as well as the ABAB ring. A four-membered ring is a closed ring structure consisting of four atoms connected end-to-end by chemical bonds. The presence of the four-membered ring (particularly the ABAB four-membered ring) facilitates the rearrangement of the material during crystallization, contributing to rapid phase transitions.

[0155] For each physical configuration obtained by multiple molecular dynamics simulations of the sulfur-based phase change memory material in A1 above, a stable amorphous relaxed structure was constructed, the corresponding ring distribution was obtained, and the ring distribution of the sulfur-based phase change memory material was obtained by averaging.

[0156] 4) Melting point

[0157] The melting point is used to reflect the thermodynamic properties of sulfur-based phase change memory materials. During the molecular dynamics simulation in A1, the melting point of the sulfur-based phase change memory material (in this embodiment, the GeTe system material) was calculated using the single-phase method and the phase coexistence method based on the potential model. When the material is in a single-phase state (no phase change occurs), the energy usually shows a smooth curve as the temperature changes. However, near the melting point, the energy will suddenly change, and the point of the sudden change is the melting point of the system. The phase coexistence method refers to a method of constructing a two-phase coexistence structure such as crystalline-liquid-crystalline, and then performing multiple independent simulations at different temperatures. When the applied temperature is higher than the melting point, the crystalline part is expected to melt. When the temperature is lower than the melting point, the crystalline part will grow, causing the liquid part to crystallize, thereby estimating the melting point of the system.

[0158] The corresponding melting points of each physical configuration obtained from multiple molecular dynamics simulations of the sulfur-based phase change memory material in A1 were calculated using the single-phase method and the phase coexistence method, and the averaged values were calculated as the melting points of the sulfur-based phase change memory material.

[0159] 5) Crystallization rate

[0160] The crystallization rate is used to reflect the kinetic properties of sulfur-based phase change memory materials. During the molecular dynamics simulation in A1, based on the potential model, the changing trend of the potential energy surface of the sulfur-based phase change memory material along with the crystallization process, that is, the crystallization rate, was obtained.

[0161] In this embodiment, multiple molecular dynamics simulations are performed to obtain multiple crystallization rates of the chalcogenide phase change memory material, and the average is calculated as the final crystallization rate of the chalcogenide phase change memory material.

[0162] In this embodiment, the above-mentioned adjustment of the influencing parameters in S2 includes: adjusting one or more groups of influencing parameters in S2 to an influencing parameter group that can reflect the physical property X of the sulfur-based phase change memory material during the molecular dynamics simulation, so as to increase the proportion of physical configurations that can reflect the physical property X in the first physical configuration set obtained after re-execution of S2; wherein the physical property X is a physical property category that has not passed verification.

[0163] For example, when the physical property of the melting point fails to be verified, a set of influencing parameters that can reflect the physical property of the melting point is given. Taking GeTe material as an example, the influencing parameter group {(crystalline, 600K), (crystalline, 700K), (liquid, 1000K), (liquid, 1100K)} can be given.

[0164] If the requirements for a particular physical property are not met, steps S2 through S7 are repeated using active learning to enhance the potential model's ability to describe that specific property. Once the accuracy requirements and physical property constraints are met, the potential model is able to capture the interactions of chalcogenide phase change materials under multiphase and multitemperature conditions, enabling its application to more complex, long-term, and large-scale crystallization and aging tasks.

[0165] This example uses a restart file to restart training of the current potential model, reduces the initial learning rate for fine-tuning, and compares the similarity of the DFT-MD (density functional simulation) and NN-MD (potential model simulation) correlation functions and coordination numbers in the liquid and amorphous states after training. Energy-volume curves (Birch–Murnaghan formula) are fitted to compare the errors in the characteristic data. Ring statistical analysis is performed to view the distribution of four-membered rings. The melting point is measured using the single-phase method and the phase coexistence method. By simulating crystallization for a long time, the energy change trend with crystallization at different temperatures is observed.

[0166] S8. Output the current potential model as the final potential model.

[0167] The trained potential model is capable of capturing the interactions between atoms and can be applied to simulation problems that are difficult to solve using first principles.

[0168] After the obtained potential model parameters are instantiated, they can be embedded as independent executable files in classic molecular dynamics software such as Lammps and GPUMD to conduct long-term and large-scale crystallization AI For Science research at the nanosecond level.

[0169] In order to further illustrate the properties of the potential model construction method provided by the present invention, the following comparative experiments were performed after the potential model training was completed:

[0170] This experiment takes GeTe material as an example.

[0171] Verification of output error:

[0172] like Figure 4 The figure shows the energy field and force field RMSE verification of the potential model using amorphous GeTe as an example. As can be seen from the figure, the energy root mean square error (RMSE) between the first-principles calculation and the potential model prediction is 9.96 meV atom for the training set and the test set, respectively. -1 and 7.66meV atom -1 The root mean square errors of the force field are and More than 70% of the configurations have energy and force field equilibrium potentials within 10.5 meV atom -1 and This shows that the current potential model has met the simulation requirements in terms of calculation error.

[0173] Verification of physical properties:

[0174] After the training, the similarities of DFT-MD (density functional simulation) and NN-MD (potential model simulation provided by the present invention) in the coordination number, energy-volume curve and ring distribution in the liquid and amorphous states were compared.

[0175] like Figure 5 The figure shows a schematic diagram of verifying the physical properties of the potential model based on the coordination number distribution function using amorphous GeTe as an example, where NN represents the coordination number distribution function obtained based on the potential model, and DFT represents the coordination number distribution function obtained based on DFT.

[0176] like Figure 6 The figure shows a schematic diagram of verifying the physical properties of the potential model based on the energy-volume curve using amorphous GeTe as an example, where NN represents the energy-volume curve obtained based on the potential model, and DFT represents the energy-volume curve obtained based on DFT.

[0177] like Figure 7 The figure shows a schematic diagram of the physical property verification of the potential model based on the ring distribution using amorphous GeTe as an example, where NN-MD represents the ring distribution obtained based on the potential model, and DFT-MD represents the ring distribution obtained based on DFT.

[0178] from Figure 5-Figure 7 It can be seen that the coordination number, energy-volume curve and ring distribution of GeTe materials presented by the simulation based on the potential model provided by the present invention are relatively similar to those based on the classical DFT simulation, reflecting the accuracy of the potential model provided by the present invention in actual application scenarios.

[0179] Further verification of melting point and crystallization rate:

[0180] like Figure 8A schematic diagram shows the physical property validation of the potential model based on melting point, using amorphous GeTe as an example. The figure includes melting point simulations using the single-phase method (top) and the phase coexistence method (bottom). The square points in the top figure represent all data points, and the triangle points represent sampling points from the total data. As can be seen in the top figure, the material's state begins to change significantly around 990K, which is close to the true melting point of 998K. In the bottom figure, IV represent different states of physical configuration. I represents the initial state, corresponding to the crystalline-liquid-crystalline configuration; II represents the fully melted state; and III represents the fully crystallized state. Comparing state IV with initial state I shows overall melting (more liquid in the center); comparing state V with initial state I shows overall crystallization (less liquid and more crystalline structure). The bottom figure shows that the melting point of the material should be higher than state V and lower than state IV, that is, between 990K and 1000K, which is close to the true melting point of 998K.

[0181] like Figure 9 A schematic diagram shows the physical validation of a potential model based on crystallization rate using amorphous GeTe as an example. The figure includes simulations of the energy changes for crystallization at multiple independent temperatures.

[0182] from Figure 8-Figure 9 It can be seen that the potential energy of amorphous GeTe gradually decreases with annealing time, which means that the system transforms into a more stable crystalline structure. As the annealing temperature increases, the rate of potential energy reduction gradually increases and the crystallization time point is delayed, which means that heating will lead to an increase in the crystallization rate of the system and also increase the difficulty of nucleation events. When the temperature reaches 650K, no crystallization phenomenon is observed at the simulated 10 nanosecond scale, which is consistent with the real physical properties of GeTe and reflects the accuracy of the potential model provided by the present invention in practical application scenarios.

[0183] In summary, this embodiment proposes a method for constructing a potential model of sulfur-based phase change storage materials based on machine learning. The GPUMD-Wizard, ASE and other software are used to greatly enrich the configuration of the data set, and outlier detection and farthest point sampling are applied to filter the data set to reduce redundant data and improve the expression efficiency of each configuration; a configuration sampling process with a longer time scale is constructed based on the potential model, and a wider range of material space is constructed through MCMD and temperature exploration, avoiding the high computing power loss of the first-principles iterative data set; an active learning mode of efficient interaction between the training set and the potential model, which gradually fills the gaps in specific phases in the data set through the configuration generated by the potential model, making the training set's expression of the system more specific and detailed, and at the same time starting model fine-tuning to refine the potential model's ability to describe the material system; based on the application and physical mechanism of sulfur-based phase change storage materials, a comprehensive analysis and verification scheme is proposed in combination with structure, thermodynamics and crystallization kinetics, which effectively characterizes the physical properties of the potential model under various environmental conditions and specific phases. The method proposed in this embodiment can efficiently generate a complete data set with as few first-principles calculations as possible. At the same time, the active learning method can continuously iteratively refine the abundance of the training set and the expressiveness of the potential model. Finally, based on the physical mechanism verification scheme of sulfur-based phase change storage materials, the characterization effect of the model and the application capability of large-scale molecular dynamics simulation can be detected in multiple dimensions.

[0184] The present invention aims to reduce the large amount of redundant data sets required when constructing potential models in the existing technology, reduce investment costs, and at the same time build an efficient data loop iterative working mode, apply active learning to gradually improve the model's ability to describe physical phenomena, and summarize the physical structure, thermodynamics, and kinetic verification schemes of each phase of sulfur-based phase change materials to maximize the evaluation of the feasibility and rationality of the potential model in large-scale molecular dynamics simulations.

[0185] In a second aspect, the present invention provides a method for calculating the characteristics of a chalcogenide phase change memory material, comprising:

[0186] Inputting the physical configuration of the chalcogenide phase change memory material to be calculated into the potential model to obtain the potential energy surface and force field of the chalcogenide phase change memory material;

[0187] The potential model is constructed using the construction method provided in the first aspect of the present invention.

[0188] The related technical method is the same as the construction method provided in the first aspect of the present invention and will not be described in detail here.

[0189] In a third aspect, the present invention provides an electronic device comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the method provided in the first aspect or the second aspect of the present invention when executing the computer program.

[0190] The related technical methods are the same as those provided in the first and second aspects of the present invention and will not be described in detail here.

[0191] In a third aspect, the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method provided in the first aspect or the second aspect of the present invention.

[0192] The related technical methods are the same as those provided in the first and second aspects of the present invention and will not be described in detail here.

[0193] In a fourth aspect, the invention further provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the method provided in the first or second aspect of the invention.

[0194] The related technical methods are the same as those provided in the first and second aspects of the present invention and will not be described in detail here.

[0195] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a potential model of a chalcogenide phase change memory material, characterized in that: include: S1. Obtaining an initial potential model of the chalcogenide phase change memory material as a current potential model; The initial potential model is trained on the first training set; The first training set is obtained by performing picosecond molecular dynamics simulation of chalcogenide phase change memory materials using a first-principles calculation method; S2. Based on the current potential model, nanosecond-level molecular dynamics simulations are performed on the chalcogenide phase change memory material under multiple sets of different influencing parameters to obtain the physical configurations of the different chalcogenide phase change memory materials under each set of influencing parameters, forming a first physical configuration set; each set of influencing parameters includes: phase state and external environment parameters; S3. Filtering out a plurality of candidate physical configurations from the first physical configuration set, which are non-outliers and have a distance in the potential model feature space greater than or equal to a first preset distance; S4. Perform single-point energy calculations on each candidate physical configuration to obtain the corresponding potential energy surface and force field, thereby constructing the second training set; S5. Using the second training set to train the current potential model, or using the total of the second training set and the first training set to train the current potential model; S6. Calculate the potential energy surface output error and force field output error of the current potential model based on the pre-collected validation set. If both output errors are less than the corresponding preset output errors, go to S8. Otherwise, adjust the screening parameters and go to S3. The screening parameters include: a first preset distance; S7, verifying the physical properties of the chalcogenide phase change memory material based on the current potential model; if the verification passes, go to S8; otherwise, adjust the influencing parameters in S2 and go to S2; S8. Output the current potential model as the final potential model.

2. The construction method according to claim 1, characterized in that The verification of the physical properties of the chalcogenide phase change memory material based on the current potential model includes: A1. Based on the current potential model, multiple molecular dynamics simulations are performed on each type of physical property of the chalcogenide phase change memory material to obtain the physical characteristics of the chalcogenide phase change memory material. The physical characteristics include: characteristic data reflecting each type of physical property of the chalcogenide phase change memory material; A2. When the errors between all feature data in the physical features and the corresponding preset reference feature data are less than the corresponding preset errors, the verification is determined to be passed; otherwise, it is failed.

3. The construction method according to claim 2, characterized in that The adjusting of the influencing parameters in S2 includes: adjusting one or more groups of influencing parameters in S2 to an influencing parameter group that can reflect the physical property X of the sulfur-based phase change memory material during the molecular dynamics simulation, so as to increase the proportion of physical configurations that can reflect the physical property X in the first physical configuration set obtained after re-execution of S2; wherein the physical property X is a physical property category that has not passed verification.

4. The construction method according to claim 2, characterized in that The physical characteristics include: local structural fingerprint characteristics of the chalcogenide phase change memory material; The local structural fingerprint features include: one or more of the SOAP descriptor, local bond orientation order parameter and coordination number of the chalcogenide phase change memory material.

5. The construction method according to claim 4, characterized in that The physical characteristics also include: energy-volume curve, ring distribution, melting point and crystallization rate of the chalcogenide phase change memory material.

6. The construction method according to any one of claims 1 to 5, characterized in that The method for determining an outlier of a physical configuration includes: calculating an outlier value of the physical configuration; when the outlier value is greater than a preset threshold, determining that the physical configuration is an outlier; otherwise, determining that the physical configuration is not an outlier; The calculation formula of the outlier is ||E-E'||; where E is the potential energy surface obtained by inputting the physical configuration into the current potential model; E' is the potential energy surface obtained by performing single-point energy calculation on the physical configuration; Alternatively, the calculation formula for the outlier is Where F is the force field obtained by inputting the physical configuration into the current potential model; F' is the force field obtained by performing single-point energy calculations on the physical configuration; and σ(F) is the standard deviation of the force field for each atom of the physical configuration contained in the force field F of the physical configuration.

7. The construction method according to any one of claims 1 to 5, characterized in that: The first training set is obtained by: The physical configurations of the sulfur-based phase change memory material under different phase states and external environmental parameters are obtained to form a second physical configuration set; the second physical configuration set includes: a crystalline configuration subset and an amorphous configuration subset; the amorphous configurations and corresponding potential energy surfaces and force fields in the amorphous configuration subset are obtained by performing picosecond molecular dynamics simulations of the sulfur-based phase change memory material in the amorphous state under different external environmental parameters based on a first-principles calculation method; the crystalline configuration subset includes: the crystalline configuration of the sulfur-based phase change memory material and the crystalline configuration after data enhancement of the crystalline configuration of the sulfur-based phase change memory material; the crystalline configuration is obtained by directly obtaining it from a material crystalline configuration library and constructing it using atomic simulation software for the sulfur-based phase change memory material; the data enhancement includes: adding defects, adding vacancies and adding strain to the crystalline configuration. screening the physical configurations in the second set of physical configurations to retain physical configurations whose physical features are separated by a distance greater than or equal to a second preset distance; Perform single-point energy calculations on each crystal configuration in the screened second physical configuration set to obtain the corresponding potential energy surface and force field; The physical configurations in the screened second physical configuration set are used as training samples, and the corresponding potential energy surfaces and force fields are used as labels to construct the first training set.

8. A method for calculating the characteristics of a chalcogenide phase change memory material, characterized in that: include: Inputting the physical configuration of the chalcogenide phase change memory material to be calculated into the potential model to obtain the potential energy surface and force field of the chalcogenide phase change memory material; Wherein, the potential model is constructed using the construction method described in any one of claims 1-7.

9. An electronic device, characterized in that: include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the method according to any one of claims 1 to 8 when executing the computer program.

10. A computer program product, characterized in that The method comprises a computer program / instruction, which implements the method according to any one of claims 1 to 8 when executed by a processor.

Images