硅酸盐学报2025，Vol.53Issue(7)：1779-1785,7.DOI:10.14062/j.issn.0454-5648.20240746
基于机器学习势函数的沸石框架钠离子导体筛选
Screening of Zeolite Framework Sodium Ion Conductor Based on Machine Learning Potential Function
摘要
Abstract
Introduction All-solid-state batteries(ASSBs),which replace flammable organic-based liquid electrolytes into ion-conducting solids,are expected to address the safety challenges for conventional lithium-ion batteries.Solid-state electrolyte materials as the most important constituent in ASSBs can be classified as oxide,sulfide,and halide solid-state electrolytes according to the anion type in the structural framework.To improve the efficiency of solid-state electrolytes in ion transport,some research efforts focus on the structural features suitable for cation migration events.Zeolite is a type of inorganic crystalline material made up of vertex-connected TO4 tetrahedra(i.e.,T=Si,Al,P,etc.)with ordered microporous structures.They have one-,two-,or three-dimensional channel systems and cation exchange capabilities,which make them ideal for use as fast ion conductors.Zeolite membranes with a close-packed structure have a low electronic conductivity(i.e.,1.5 × 10-10 S/cm),making them an ideal option for solid-state batteries due to the high stability to air components and ability to reduce dendrite growth.In this paper,the molecular dynamics(MD)simulations of zeolite structures of relatively large sizes were accelerated via training a machine learning interatomic potential function(MLIP)with a good generalization ability,in turn screening zeolite structural frameworks suitable for Na+transport as promising solid-state electrolytes for sodium ions in ASSBs. Methods The zeolite structure was characterized using the Smooth Overlap of Atomic Positions(SOAP)descriptor method provided by a software package named DScribe.Ab initio arithmetic molecular dynamics(AIMD)data used to construct the training dataset were computed by the Vienna Ab initio Simulation Package(VASP).The projector-augmented-wave(PAW)pseudopotential method and the Perdew-Burke-Ernzerhof(PBE)generalized gradient approximation(GGA)function were used for the first-principles calculations,and the centre of the Γ-point in the Brillouin zone was chosen as the k-point,the NVT systematic and the Nosé-Hoover thermostat were used for the simulations,and a Nosé-mass corresponding to a period of 40 time steps was chosen(SMASS=0,a simulation time-step of 1 fs,and a simulation temperature of 1000 K).The MLIP model was based on the Nosé-Hoover model. The MLIP was fine-tuned based on a pretrained model of the Crystal Hamiltonian Graph Neural Network(CHGNet)with the energy and force calculated by AIMD as target values,and the model was trained with a learning rate of 10-3.The trained CHGNet potential function was embedded into the Atomic Simulation Environment(ASE)library to carry out the MD simulations under the NVT regime with a simulation step of 1 fs and a total simulation duration of 100 ps. Results and discussion The CHGNet pretrained model is based on the first-principles computational data of more than 1.5 million inorganic structures in the Materials program database,and it needs to be fine-tuned via adding the relevant data to achieve the better results when used for the computation of specific systems.To ensure that the data in the training set are applicable to all zeolite structures in the Na-Si-Al-O quaternary,the SOAP descriptors are used to select the ten structures with the differences in structural frameworks for the construction of the dataset,and the model obtained by training is capable of achieving a superior generalization in terms of simulated temperatures and systems.The model is used to select 18 representative structures from 124 zeolite structures for MD simulations at 1200 K.In most of the structures,Na+is moved slightly and adsorbed to the edge of the pore channel,and only Na+in ACO framework(ICSD 027717)is able to achieve a high enough mean-square displacement(MSD)of more than 800 Å2,which is extrapolated to obtain a potential barrier of 0.25 eV and an extrapolated room temperature ionic conductivity of 2.66 mS/cm after supplementing the multi-temperature simulation data.This is promising to be investigated as an important zeolite-structured solid-state electrolyte. Conclusions In this work,different structures selected by SOAP descriptors were used to fine-tune the CHGNet potential function model to obtain MLIP with a better generalization ability in the Na-Si-Al-O quaternary system.Using this MLIP for MD simulations indicated that in most of the zeolite structures Na ions could be adsorbed far away from the cavity centre and could not form a continuous transport channel,but the ACO framework(ICSD 027717)with a migration barrier of 0.25 eV and an extrapolated room-temperature ionic conductivity of 2.66 mS/cm occurred.Na+ions could be transported rapidly with the help of a pore structure,and it could be considered as a solid-state sodium battery electrolyte material for further studies.关键词
钠离子导体/沸石结构/机器学习势函数/固态电解质筛选Key words
sodium ion conductors/zeolite structure/machine learning interatomic potential/solid electrolyte screening分类
化学化工引用本文复制引用
符晓,肖睿娟,李泓..基于机器学习势函数的沸石框架钠离子导体筛选[J].硅酸盐学报,2025,53(7):1779-1785,7.基金项目
国家自然科学基金面上项目(52172258). （52172258）
