Search

CN-121999946-A - Hafnium-zirconium oxide full-component ferroelectric property and fatigue optimization method based on machine learning potential

CN121999946ACN 121999946 ACN121999946 ACN 121999946ACN-121999946-A

Abstract

The invention relates to a full-component ferroelectric characteristic and fatigue optimization method of hafnium-zirconium oxide based on machine learning potential, which comprises the steps of 1, obtaining atomic configuration, energy, stress and electric dipole moment data of Hf x Zr 1‑x O 2 under different components x (x is more than or equal to 0 and less than or equal to 1) as an original dataset, 2, training a multi-task machine learning potential model by using a deep learning algorithm, wherein the model comprises an energy-force prediction branch and an electric dipole moment prediction branch, 3, carrying out molecular dynamics evolution under different temperatures and external electric fields to obtain a variation curve of polarization intensity along with a cycle period, and 4, screening an optimal material component according to a nonlinear corresponding relation between fatigue rate obtained through simulation and Hf content. Compared with the prior art, the invention obtains the machine learning potential model capable of synchronously describing energy, force and electric dipole moment by constructing a training data set covering the whole component space and utilizing deep neural network training, reveals the regulation rule of Hf doping proportion to crystalline phase stability through large-scale molecular dynamics simulation, quantifies the relation between domain wall movement rate and fatigue rate, and provides theoretical guidance of atomic scale for component design of FeRAM devices.

Inventors

  • YUAN ZHEN

Assignees

  • 北京大学
  • 袁真

Dates

Publication Date
20260508
Application Date
20260211

Claims (10)

  1. 1. A method for optimizing the full-component ferroelectric property and fatigue of hafnium-zirconium oxide based on machine learning potential is characterized by comprising the following steps: Step S1, constructing an original training data set covering a full-component space and a multi-dimensional phase configuration, wherein the data set comprises an atomic configuration and corresponding energy, stress, wiry tensor and local electric dipole moment labels of atoms; S2, constructing a multi-task deep neural network comprising potential energy surface prediction branches and electric dipole moment prediction branches, and performing iterative training of a potential function model by utilizing an active learning strategy to obtain a high-precision potential function model considering mechanical properties and electric response; Step 3, deploying a trained potential function model in molecular dynamics simulation, and performing polarization inversion dynamic simulation under electric field coupling by responding to an external electric field changing along with time in real time through an electric dipole moment prediction branch; and S4, performing multi-period electric field cyclic loading simulation on different component models, quantitatively calculating the attenuation rate of the residual polarization intensity, establishing a mapping relation between the components and the fatigue life, and outputting optimal component interval parameters considering both high residual polarization and low fatigue rate according to the mapping relation.
  2. 2. The method for optimizing the full-component ferroelectric characteristics and fatigue of hafnium-zirconium oxide based on machine learning potential according to claim 1, wherein in step S1, the specific process of constructing the original training data set is as follows: s1.1, constructing an HZO solid solution supercell model by using a monoclinic phase, a tetragonal phase and an orthoferroelectric phase of a block HfO 2 and ZrO 2 as parent phase structures through an atomic substitution method, wherein the value range of x is more than or equal to 0 and less than or equal to 1; S1.2, introducing oxygen vacancy defects into a part of the supercell model or constructing domain wall models with different orientations; S1.3, static self-consistent calculation or first sex principle molecular dynamics simulation is carried out on the model by utilizing first sex principle calculation software, and energy, atomic stress, wiry tensor and atomic Boen effective charge or Wannier center data of a system are collected and used as training labels.
  3. 3. The method for optimizing the full-component ferroelectric characteristics and fatigue of the hafnium-zirconium oxide based on the machine learning potential according to claim 2, wherein the first principle of principle calculation adopts a projection prefix plus plane wave (PAW) method, the exchange correlation functional is PBEsol, and the atomic born effective charge or Wannier central data is obtained through calculation by a modern polarization theory (Berry Phase) method.
  4. 4. The method for optimizing the full-component ferroelectric characteristics and fatigue of hafnium-zirconium oxide based on machine learning potential according to claim 1, wherein in step S2, the multi-tasking deep neural network comprises a descriptor network and a fitting network; the fitting network comprises two independent output branches, wherein the first branch is used for outputting an energy scalar, an atomic stress vector and a wiry tensor of a system; The training process adopts a dynamic weighting loss function that L=p e ΔE 2 + p f ΔF 2 + p v ΔV 2 + p d ΔD 2 Wherein Δe, Δf, Δv, Δd are prediction errors of energy, stress, wiry, and dipole moment, respectively, and p e 、p f 、p v 、p d is a corresponding dynamic weight coefficient.
  5. 5. The method for optimizing the full-component ferroelectric characteristics and fatigue of hafnium-zirconium oxide based on machine learning potential according to claim 1, wherein in step S2, the active learning strategy adopts a closed loop iterative mechanism (DP-GEN) of "explore-tag-train", and specifically comprises: s2.1, performing molecular dynamics sampling under the conditions of wide temperature range and different pressures by using a current potential function model; S2.2, screening candidate configurations according to Force Deviation (Force device) of configuration prediction by a model, and if the maximum Force Deviation of the configuration falls into a preset threshold value interval sigma low ≤ε max ≤σ high , calling a first sexual principle to calculate and carry out high-precision marking; s2.3, training, namely merging the marked new data into a training set, and retraining the potential function model until the proportion of the candidate configuration converges.
  6. 6. The method for optimizing the full-component ferroelectric characteristics and fatigue of hafnium-zirconium oxide based on machine learning potential according to claim 1, wherein the specific process of step S3 is as follows: s3.1, under an isothermal isobaric (NPT) ensemble, applying an alternating electric field E (t) in the form of triangular waves or trapezoidal waves along the ferroelectric polarization axis direction; S3.2, calling an electric dipole moment prediction branch of a potential function model in real time in a simulation process, calculating the instantaneous total polarization intensity P (t) of the system, and drawing a P (t) -E (t) curve to obtain an atomic-scale electric hysteresis loop; S3.3, visually analyzing atomic displacement or central symmetry parameters, and observing nucleation, growth and pinning processes of domain walls under the drive of an electric field.
  7. 7. The method for optimizing the full-component ferroelectric characteristics and fatigue of hafnium-zirconium oxide based on machine learning potential according to claim 1, wherein in step S4, the extraction of the fatigue failure rule comprises: S4.1, carrying out multi-period electric field cyclic loading on models of different components x, and recording an attenuation curve of the residual polarization intensity P r along with the cyclic times N; S4.2, calculating a fatigue rate factor lambda, and determining that a main mechanism causing fatigue is irreversible transformation of an orthogonal phase to a monoclinic phase by combining phase transformation energy barrier analysis; And S4.3, screening out the component with the highest remaining polarization intensity retention rate and the initial polarization value meeting the preset threshold value as the optimal material formula.
  8. 8. The method for optimizing the full-component ferroelectric characteristics and fatigue of hafnium-zirconium oxide based on machine learning potential according to claim 7, wherein the optimal component interval is Hf 0.3 Zr 0.7 O 2 with Hf content of about 30%.
  9. 9. A computing system implementing the method of any of claims 1 to 8, comprising: The data construction module is used for generating an initial configuration covering the full component space and executing first sexual principle calculation to acquire training data; The model training module is used for training a multi-task deep learning potential function model with an electric dipole moment prediction function based on an active learning strategy; The dynamic simulation module is used for loading a potential function model and applying an external electric field, and executing molecular dynamics simulation to generate a hysteresis loop and a domain wall evolution track; And the analysis optimization module is used for counting fatigue characteristic data and outputting an optimal material component suggestion.
  10. 10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method steps of any of claims 1 to 8.

Description

Hafnium-zirconium oxide full-component ferroelectric property and fatigue optimization method based on machine learning potential Technical Field The invention belongs to the technical field of molecular simulation calculation of material science, and particularly relates to a full-component ferroelectric property and fatigue optimization method of hafnium-zirconium oxide based on machine learning potential. Background As integrated circuit processes continue to shrink to nanometer scale, conventional ferroelectric memories (e.g., PZT, baTiO 3, etc.) have been developed into bottlenecks due to poor compatibility with CMOS processes, limited thickness scaling, etc. Hafnium-based ferroelectric materials (such as Hf xZr1-xO2 (HZO)) found in 2011 are considered as ideal candidates for next generation low power consumption, high density nonvolatile memories because they still maintain excellent ferroelectricity in ultra-thin state and are perfectly compatible with existing CMOS processes. Performance studies on hafnium-based ferroelectric materials have gradually been directed from physical discovery to reliability optimization and mechanism interpretation. In the field of theoretical simulation, the first principle calculation can provide high-precision ground state energy information, but is limited by a calculation scale (hundred atomic magnitude) and a time scale (picosecond magnitude), and cannot simulate real domain wall dynamic evolution, long-term fatigue cycle and imprinting effect. In recent years, the development of machine learning potential functions provides a new approach for solving the above problems, which can realize large-scale and long-period molecular dynamics simulation on the premise of keeping the accuracy of the first-order principle calculation stage. Studies have shown that polarization inversion paths and domain wall structures in HfO2 systems can be explored using deep learning potential energy models (e.g., deepMD), proving that microscopic domain wall behavior is a key factor in determining the macroscopic reliability of HZO devices. In the prior art, the simulation of the HZO system still has the following limitations that firstly, the coverage of components is incomplete, and the electrical property of the HZO is strongly regulated and controlled by the doping proportion of Zr. Experiments show that the zirconium-rich component tends to exhibit antiferroelectric properties toward the tetragonal phase (t-phase), while the hafnium-rich component is stabilized in the orthogonal phase (o-phase) to exhibit ferroelectric properties. The existing simulation research focuses on specific components, lacks a unified potential function model covering the space of all components from x=0 to x=1, is difficult to systematically reveal the influence of continuous changes of components on competing laws and electrical characteristics, lacks quantitative prediction of dynamic failure characteristics, namely the existing machine learning potential simulation focuses on static structures or characterization of single polarization turning paths, is difficult to realize quantitative prediction of polarization attenuation (fatigue) and loop deviation (imprinting) processes under multiple alternating electric field cycles, and is difficult to couple defects and evolution mechanisms, namely experimental means are difficult to capture transient atomic motion processes under electric field driving, especially near complex defects such as domain walls, phase interfaces and the like, so that physical association among material components, microstructure evolution and macroscopic failure laws is still unclear. Therefore, a simulation method which can be compatible with full-component description, accurately capture dynamic polarization flip response and quantitatively predict fatigue and imprinting effect is developed, and has important scientific significance and application value for optimizing the formula of the HZO material and improving the reliability of the ferroelectric memory. Disclosure of Invention The invention aims at solving the problems that the simulation component of Hafnium Zirconium Oxide (HZO) in the prior art is single, and the capability of quantitatively predicting dynamic failure process (fatigue and imprinting) under a cyclic electric field is lacking. The method realizes high-precision simulation of polarization turning paths, domain wall dynamics and fatigue characteristics in the full-component range of HZO on the atomic scale by constructing a multi-task deep learning potential energy model. The aim of the invention can be achieved by the following technical scheme: a method for optimizing the full-component ferroelectric property and fatigue of hafnium-zirconium oxide based on machine learning potential is characterized by comprising the following steps: step S1, constructing an original training data set covering the full component space and the multi-dimensional phase configu