应用化学 ›› 2025, Vol. 42 ›› Issue (6): 757-775.DOI: 10.19894/j.issn.1000-0518.250064

• 综合评述 • 上一篇    

机器学习驱动的钙钛矿发光材料研究进展:智能设计、性能优化与产业化应用

王一铭1, 杨博翔2, 张华久3, 孙立恒1(), 董彪1()   

  1. 1.吉林大学电子科学与工程学院,长春 130012
    2.吉林大学化学学院,长春 130015
    3.长春吉大附中实验学校,长春 130021
  • 收稿日期:2025-02-21 接受日期:2025-05-11 出版日期:2025-06-01 发布日期:2025-07-01
  • 通讯作者: 孙立恒,董彪
  • 作者简介:第一联系人:共同第一作者
  • 基金资助:
    国家自然科学基金(52250077);吉林省自然科学基金(20220402005GH);吉林省科技厅项目(20210204095YY)

Research Progress of Machine Learning-Driven Perovskite Luminescent Materials: Intelligent Design, Performance Optimization and Industrial Application

Yi-Ming WANG1, Bo-Xiang YANG2, Hua-Jiu ZHANG3, Li-Heng SUN1(), Biao DONG1()   

  1. 1.College of Electronic Science and Engineering,Jilin University,Changchun 130012,China
    2.College of Chemistry,Jilin University,Changchun 130015,China
    3.Affiliated Middle School to Jilin University,Changchun 130021,China
  • Received:2025-02-21 Accepted:2025-05-11 Published:2025-06-01 Online:2025-07-01
  • Contact: Li-Heng SUN,Biao DONG
  • About author:dongb@jlu.edu.cn
    sunlh24@jlu.edu.cn
  • Supported by:
    the National Natural Science Foundation of China(52250077);the Natural Science Foundation of Jilin Province(20220402005GH);Jilin Provincial Department of Science and Technology(20210204095YY)

摘要:

机器学习技术为钙钛矿发光材料的高效开发提供了突破性解决方案。 本文系统综述了该技术在材料设计、性能优化及实验指导中的创新应用,重点解决传统试错法效率低下(如CsPbBr3单组分优化需320次实验,耗时14个月)与理论计算精度不足(如传统密度泛函理论(DFT)对Eu3+掺杂体系发光效率预测误差达40%)的瓶颈问题。 在材料结构与性能预测方面,基于神经网络的跨尺度建模显著提升了关键参数的预测精度。 开发的深度卷积神经网络(CNN)模型,通过128×128×128电子密度网格解析晶体畸变,实现带隙预测平均绝对误差(MAE)0.08 eV,对CsPbBr3的预测误差仅1.3%。 图神经网络(GNN)进一步量化层状钙钛矿堆垛角度与带隙的关联,预测误差<0.05 eV; 材料筛选与优化中,多目标算法实现了性能指标的协同提升。 采用NSGA-II算法筛选出Cs2SnGeI6,其外量子效率(EQE)达24%,器件寿命(T??)延长至1200 h。 贝叶斯优化 (Bayesian optimization, BO) 框架结合机器人平台,将CsPbBr3-x I x 量子点的光致发光量子产率(PLQY)从45%提升至89%,实验迭代次数减少85%; 实验设计方面,微流控机器人平台通过动态参数调控(流速10~100 μL/min,混合时间110 s),24 h内筛选出CsPbBr1.5I1.5最佳配比,发射波长误差±3 nm,PLQY达92%。 柔性器件工艺优化中,贝叶斯算法将光电转换效率(PCE)从18.2%提升至21.5%,工艺时间缩短62.5%。

关键词: 机器学习, 钙钛矿材料, 智能材料筛选, 应用研究

Abstract:

Machine learning techniques provide a breakthrough solution for the efficient development of perovskite light-emitting materials. This paper systematically reviews the innovative applications of this technology in material design, performance optimization, and experimental guidance, focusing on solving the bottlenecks of the inefficiency of the traditional trial-and-error method (320 experiments are required for the optimization of a single component of CsPbBr3, which takes 14 months) and the insufficient accuracy of theoretical calculations (the error in the prediction of the luminous efficiency of the Eu3+-doped system by DFT is as high as 40%). In terms of material structure and property prediction, cross-scale modeling based on neural networks has significantly improved the prediction accuracy of key parameters. The developed deep convolutional neural network (CNN) model analyzes crystal distortion through a 128×128×128 electron density grid, achieving a mean absolute error (MAE) of 0.08 eV for band gap prediction, with a prediction error of only 1.3% for CsPbBr3. The graph neural network (GNN) further quantifies the correlation between the stacking angle of layered perovskites and the band gap, with a prediction error of <0.05 eV. In material screening and optimization, the multi-objective algorithm achieves a synergistic improvement in performance indicators. The NSGA-II algorithm was used to screen Cs2SnGeI6, which achieved an external quantum efficiency (EQE) of 24% and extended the device lifetime (T??) to 1200 h. The Bayesian optimization framework combined with a robotic platform increased the photoluminescence quantum yield (PLQY) of CsPbBr3-xIx quantum dots from 45% to 89%, and the number of experimental iterations was reduced by 85%. In terms of experimental design, the microfluidic robotic platform screened the optimal ratio of CsPbBr1.5I1.5 within 24 h by dynamically adjusting the parameters (flow rate 10~100 μL/min, mixing time 110 s), with an emission wavelength error of ±3 nm and a PLQY of 92%. In the process optimization of flexible devices, the Bayesian algorithm increased the photoelectric conversion efficiency (PCE) from 18.2% to 21.5%, and the process time was shortened by 62.5%.

Key words: Machine learning, Perovskite materials, Intelligent material screening, Application research

中图分类号: