搜索

x

留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

基于迁移学习的钙钛矿材料带隙预测

孙涛 袁健美

引用本文:
Citation:

基于迁移学习的钙钛矿材料带隙预测

孙涛, 袁健美

Band gap prediction of perovskite materials based on transfer learning

Sun Tao, Yuan Jian-Mei
PDF
HTML
导出引用
  • 针对快速获取钙钛矿材料带隙值的问题, 建立特征融合神经网络模型(CGCrabNet), 利用迁移学习策略对钙钛矿材料的带隙进行预测. CGCrabNet从材料的化学方程式和晶体结构两方面提取特征, 并拟合特征和带隙之间的映射, 是一个端到端的神经网络模型. 在开放量子材料数据库中数据(OQMD数据集)预训练的基础上, 通过仅175条钙钛矿材料数据对CGCrabNet参数进行微调, 以提高模型的稳健性. 数值实验结果表明, CGCrabNet在OQMD数据集上对带隙的预测误差比基于注意力的成分限制网络(CrabNet)降低0.014 eV; 本文建立的模型对钙钛矿材料预测的平均绝对误差为0.374 eV, 分别比随机森林回归、支持向量机回归和梯度提升回归的预测误差降低了0.304 eV、0.441 eV和0.194 eV; 另外, 模型预测的SrHfO3和RbPaO3等钙钛矿材料的带隙与第一性原理计算的带隙相差小于0.05 eV, 这说明CGCrabNet可以快速准确地预测钙钛矿材料的性质, 加速新材料的研发过程.
    The band gap is a key physical quantity in material design. First-principles calculations based on density functional theory can approximately predict the band gap, which often requires significant computational resources and time. Deep learning models have the advantages of good fitting capability and automatic feature extraction from the data, and are gradually used to predict the band gap. In this paper, aiming at the problem of quickly obtaining the band gap value of perovskite material, a feature fusion neural network model, named CGCrabNet, is established, and the transfer learning strategy is used to predict the band gap of perovskite material. The CGCrabNet extracts features from both chemical equation and crystal structure of materials, and fits the mapping between feature and band gap. It is an end-to-end neural network model. Based on the pre-training data obtained from the Open Quantum Materials Database (OQMD dataset), the CGCrabNet parameters can be fine-tuned by using only 175 perovskite material data to improve the robustness of the model.The numerical and experimental results show that the prediction error of the CGCrabNet model for band gap prediciton based on the OQMD dataset is 0.014 eV, which is lower than that obtained from the prediction based on compositionally restricted attention-based network (CrabNet). The mean absolute error of the model developed in this paper for predicting perovskite materials is 0.374 eV, which is 0.304 eV, 0.441 eV and 0.194 eV lower than that obtained from random forest regression, support vector machine regression and gradient boosting regression, respectively. The mean absolute error of the test set of CGCrabNet trained only by using perovskite data is 0.536 eV, and the mean absolute error of the pre-trained CGCrabNet decreases by 0.162 eV, which indicates that the transfer learning strategy plays a significant role in improving the prediction accuracy of small data sets (perovskite material data sets). The difference between the predicted band gap of some perovskite materials such as SrHfO3 and RbPaO3 by the model and the band gap calculated by first-principles is less than 0.05 eV, which indicates that the CGCrabNet can quickly and accurately predict the properties of new materials and accelerate the development process of new materials.
      通信作者: 袁健美, yuanjm@xtu.edu.cn
    • 基金项目: 湖南省自然科学基金(批准号: 2023JJ30567, 2021JJ30650)资助的课题.
      Corresponding author: Yuan Jian-Mei, yuanjm@xtu.edu.cn
    • Funds: Project supported by the Natural Science Foundation of Hunan Province, China (Grant Nos. 2023JJ30567, 2021JJ30650).
    [1]

    范晓丽 2015 中国材料进展 34 689Google Scholar

    Fan X L 2015 Mater. China 34 689Google Scholar

    [2]

    万新阳, 章烨辉, 陆帅华, 吴艺蕾, 周跫桦, 王金兰 2022 物理学报 71 177101Google Scholar

    Wan X Y, Zhang Y H, Lu S H, Wu Y L, Zhou Q H, Wang J L 2022 Acta Phys. Sin. 71 177101Google Scholar

    [3]

    Xie T, Grossman J C 2018 Phys. Rev. Lett. 120 145301Google Scholar

    [4]

    Chen C, Ye W K, Zuo Y X, Zheng C, Ong S P 2019 Chem. Mater. 31 3564Google Scholar

    [5]

    Karamad M, Magar R, Shi Y T, Siahrostami S, Gates L D, Farimani A B 2020 Phys. Rev. Materials 4 093801Google Scholar

    [6]

    Jha D, Ward L, Paul A, Liao W K, Choudhary A, Wolverton C, Agrawal A 2018 Sci. Rep. 8 17593Google Scholar

    [7]

    Goodall R E A, Lee A A 2020 Nat. Commun. 11 6280Google Scholar

    [8]

    Wang A Y T, Kauwe S K, Murdock R J, Sparks T D 2021 NPJ Comput. Mater. 7 77Google Scholar

    [9]

    胡扬, 张胜利, 周文瀚, 刘高豫, 徐丽丽, 尹万健, 曾海波 2023 硅酸盐学报 51 452Google Scholar

    Hu Y, Zhang S L, Zhou W H, Liu G Y, Xu L L, Yin W J, Zeng H B 2023 J. Chin. Chem. Soc. 51 452Google Scholar

    [10]

    Guo Z, Lin B 2021 Sol. Energy 228 689Google Scholar

    [11]

    Gao Z Y, Zhang H W, Mao G Y, Ren J N, Chen Z H, Wu C C, Gates I D, Yang W J, Ding X L, Yao J X 2021 Appl. Surf. Sci. 568 150916Google Scholar

    [12]

    Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez A N, Kaiser L, Polosukhin I 2017 arXiv: 1706.03762v5 [cs. CL

    [13]

    Nix D A, Weigend A S 1994 Proceedings of 1994 Ieee International Conference on Neural Networks (ICNN’94) Orlando, FL, USA, 28 June–02 July, 1994 p55

    [14]

    You Y, Li J, Reddi S, et al. 2020 arXiv: 1904.00962v5 [cs. LG

    [15]

    Smith L N 2017 arXiv: 1506.01186v6 [cs. CV

    [16]

    Saal J E, Kirklin S, Aykol M, Meredig B, Wolverton C 2013 JOM 65 1501Google Scholar

    [17]

    Jain A, Ong S P, Hautier G, Chen W, Richards W D, Dacek S, Cholia S, Gunter D, Skinner D, Ceder G, Persson K A 2013 APL Mater. 1 011002Google Scholar

    [18]

    Yamamoto T 2019 Crystal Graph Neural Networks for Data Mining in Materials Science (Yokohama: Research Institute for Mathematical and Computational Sciences, LLC

    [19]

    Kirklin S, Saal J E, Meredig B, Thompson A, Doak J W, Aykol M, Rühl S, Wolverton C 2015 NPJ Comput. Mater. 1 15Google Scholar

    [20]

    Calfa B A, Kitchin J R 2016 AIChE J. 62 2605Google Scholar

    [21]

    Ward L, Agrawal A, Choudhary A, Wolverton C 2016 NPJ Comput. Mater. 2 16028Google Scholar

    [22]

    Tshitoyan V, Dagdelen J, Weston L, Dunn A, Rong Z Q, Kononova O, Persson K A, Ceder G, Jain A 2019 Nature 571 95Google Scholar

    [23]

    Breiman L 2001 Mach. Learn. 45 5Google Scholar

    [24]

    Wu Y R, Li H P, Gan X S 2013 Adv. Mater. Res. 848 122Google Scholar

    [25]

    孙涛, 袁健美 2023 物理学报 72 028901Google Scholar

    Sun T, Yuan J M 2023 Acta Phys. Sin. 72 028901Google Scholar

    [26]

    Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M, Duchesnay É 2011 J. Mach. Learn. Res. 12 2825Google Scholar

  • 图 1  CGCrabNet模型算法

    Fig. 1.  CGCrabNet model algorithm.

    图 2  CGCrabNet预训练损失值变化

    Fig. 2.  CGCrabNet pre-training loss value change.

    图 3  验证集预测带隙值

    Fig. 3.  Predicted band gap values on the validation set.

    图 4  测试集预测带隙值

    Fig. 4.  Predicted band gap values on the test set.

    图 5  钙钛矿材料带隙预测的平均绝对误差对比

    Fig. 5.  MAE of band gap prediction for perovskite materials.

    图 6  预测带隙与计算带隙散点图

    Fig. 6.  Predicting and calculating band gap scatter maps.

    表 1  超参数取值

    Table 1.  Hyperparameter value.

    超参数名称 含义
    $ {d}_{{\rm{m}}} $ 元素特征构造得到的向量维度 512
    $ {N}_{{{f}}} $ 化学式中最大元素种类 7
    N 注意力机制层数 3
    n 注意力机制头数 4
    I 参与训练的元素种类和 89
    T 图卷积层数 3
    $ {V}_{{\rm{c}}{\rm{g}}} $ 节点嵌入后元素向量维度 16
    $ {w}_{1} $∶$ {w}_{2} $ 权重比参数 7:3
    Epochs 最大迭代次数 300
    batch_size 批处理大小 256
    下载: 导出CSV

    表 2  元素嵌入法测试结果(单位: eV)

    Table 2.  Elemental embedding method test results (in eV).

    元素嵌入方法 Train MAE Val MAE Test MAE
    One-Hot 0.185 0.423 0.433
    Magpie 0.428 0.546 0.566
    Mat2vec 0.203 0.408 0.420
    下载: 导出CSV

    表 3  深度学习模型测试结果(单位: eV)

    Table 3.  Deep learning model test results (in eV).

    Train MAEVal MAETest MAE
    CGCNN0.5020.6050.601
    Roost0.1780.4470.455
    CrabNet0.2260.4220.427
    HotCrab0.1770.4220.440
    CGCrabNet0.1870.4080.413
    下载: 导出CSV

    表 4  回归模型参数

    Table 4.  Regression model parameters.

    机器学习方法超参数名称取值
    RF子学习器数量90
    SVR核函数多项式核
    多项式核次数3
    正则化强度2
    伽马参数2
    零系数1.5
    GBR子学习器数量500
    学习率0.2
    最大深度4
    损失函数绝对误差函数
    下载: 导出CSV

    表 5  钙钛矿材料预测值和计算值对比 (单位: eV)

    Table 5.  Comparison of predicted and calculated values for perovskite materials (in eV).

    化学式 带隙计算值 CGCrabNet RF SVR GBR
    NbTlO3 0.112 0.658 1.458 1.296 1.614
    ZnAgF3 1.585 1.776 1.836 2.194 1.840
    AcAlO3 4.102 3.212 2.881 3.197 2.963
    BeSiO3 0.269 1.116 2.813 2.963 3.777
    TmCrO3 1.929 1.682 1.612 1.987 1.668
    SmCoO3 0.804 0.644 0.821 1.043 0.724
    CdGeO3 0.102 0.586 0.911 1.675 0.196
    CsCaCl3 5.333 4.891 4.918 5.116 5.157
    HfPbO3 2.415 2.724 1.733 2.346 1.967
    SiPbO3 1.185 1.327 1.407 1.543 1.079
    SrHfO3 3.723 3.683 2.821 3.370 3.253
    PrAlO3 2.879 3.139 2.665 2.091 2.984
    BSbO3 1.405 1.123 0.653 -0.025 0.579
    CsEuCl3 0.637 0.388 1.500 4.477 0.949
    LiPaO3 3.195 3.100 2.443 -0.306 2.553
    PmErO3 1.696 1.309 1.550 1.682 1.252
    TlNiF3 3.435 2.806 2.063 3.049 3.255
    MgGeO3 3.677 1.256 0.979 1.623 1.073
    NaVO3 0.217 0.785 0.911 0.180 0.989
    RbVO3 0.250 0.616 1.736 0.290 1.534
    KZnF3 3.695 3.785 2.853 3.203 3.295
    NdInO3 1.647 1.587 1.653 0.889 1.590
    RbCaF3 6.397 6.974 6.482 6.372 6.028
    RbPaO3 3.001 2.952 2.864 -0.234 2.937
    PmInO3 1.618 1.480 1.896 1.222 1.754
    KMnF3 2.656 2.991 2.647 2.428 2.730
    NbAgO3 1.334 1.419 1.369 1.227 1.265
    CsCdF3 3.286 3.078 2.990 2.724 2.879
    KCdF3 3.101 3.125 2.789 2.365 2.990
    CsYbF3 7.060 6.641 6.325 6.523 6.736
    NaTaO3 2.260 1.714 1.680 2.093 1.715
    CsCaF3 6.900 6.874 6.291 6.416 6.379
    RbSrCl3 4.626 4.470 4.966 4.647 4.795
    AcGaO3 2.896 3.199 2.740 2.869 2.981
    BaCeO3 2.299 1.789 3.918 2.696 3.655
    注: CGCrabNet, RF, SVR和GBR分别代表特征融合神经网络、随机森林回归、支持向量回归和梯度提升回归模型计算得到的带隙值.
    下载: 导出CSV
  • [1]

    范晓丽 2015 中国材料进展 34 689Google Scholar

    Fan X L 2015 Mater. China 34 689Google Scholar

    [2]

    万新阳, 章烨辉, 陆帅华, 吴艺蕾, 周跫桦, 王金兰 2022 物理学报 71 177101Google Scholar

    Wan X Y, Zhang Y H, Lu S H, Wu Y L, Zhou Q H, Wang J L 2022 Acta Phys. Sin. 71 177101Google Scholar

    [3]

    Xie T, Grossman J C 2018 Phys. Rev. Lett. 120 145301Google Scholar

    [4]

    Chen C, Ye W K, Zuo Y X, Zheng C, Ong S P 2019 Chem. Mater. 31 3564Google Scholar

    [5]

    Karamad M, Magar R, Shi Y T, Siahrostami S, Gates L D, Farimani A B 2020 Phys. Rev. Materials 4 093801Google Scholar

    [6]

    Jha D, Ward L, Paul A, Liao W K, Choudhary A, Wolverton C, Agrawal A 2018 Sci. Rep. 8 17593Google Scholar

    [7]

    Goodall R E A, Lee A A 2020 Nat. Commun. 11 6280Google Scholar

    [8]

    Wang A Y T, Kauwe S K, Murdock R J, Sparks T D 2021 NPJ Comput. Mater. 7 77Google Scholar

    [9]

    胡扬, 张胜利, 周文瀚, 刘高豫, 徐丽丽, 尹万健, 曾海波 2023 硅酸盐学报 51 452Google Scholar

    Hu Y, Zhang S L, Zhou W H, Liu G Y, Xu L L, Yin W J, Zeng H B 2023 J. Chin. Chem. Soc. 51 452Google Scholar

    [10]

    Guo Z, Lin B 2021 Sol. Energy 228 689Google Scholar

    [11]

    Gao Z Y, Zhang H W, Mao G Y, Ren J N, Chen Z H, Wu C C, Gates I D, Yang W J, Ding X L, Yao J X 2021 Appl. Surf. Sci. 568 150916Google Scholar

    [12]

    Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez A N, Kaiser L, Polosukhin I 2017 arXiv: 1706.03762v5 [cs. CL

    [13]

    Nix D A, Weigend A S 1994 Proceedings of 1994 Ieee International Conference on Neural Networks (ICNN’94) Orlando, FL, USA, 28 June–02 July, 1994 p55

    [14]

    You Y, Li J, Reddi S, et al. 2020 arXiv: 1904.00962v5 [cs. LG

    [15]

    Smith L N 2017 arXiv: 1506.01186v6 [cs. CV

    [16]

    Saal J E, Kirklin S, Aykol M, Meredig B, Wolverton C 2013 JOM 65 1501Google Scholar

    [17]

    Jain A, Ong S P, Hautier G, Chen W, Richards W D, Dacek S, Cholia S, Gunter D, Skinner D, Ceder G, Persson K A 2013 APL Mater. 1 011002Google Scholar

    [18]

    Yamamoto T 2019 Crystal Graph Neural Networks for Data Mining in Materials Science (Yokohama: Research Institute for Mathematical and Computational Sciences, LLC

    [19]

    Kirklin S, Saal J E, Meredig B, Thompson A, Doak J W, Aykol M, Rühl S, Wolverton C 2015 NPJ Comput. Mater. 1 15Google Scholar

    [20]

    Calfa B A, Kitchin J R 2016 AIChE J. 62 2605Google Scholar

    [21]

    Ward L, Agrawal A, Choudhary A, Wolverton C 2016 NPJ Comput. Mater. 2 16028Google Scholar

    [22]

    Tshitoyan V, Dagdelen J, Weston L, Dunn A, Rong Z Q, Kononova O, Persson K A, Ceder G, Jain A 2019 Nature 571 95Google Scholar

    [23]

    Breiman L 2001 Mach. Learn. 45 5Google Scholar

    [24]

    Wu Y R, Li H P, Gan X S 2013 Adv. Mater. Res. 848 122Google Scholar

    [25]

    孙涛, 袁健美 2023 物理学报 72 028901Google Scholar

    Sun T, Yuan J M 2023 Acta Phys. Sin. 72 028901Google Scholar

    [26]

    Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M, Duchesnay É 2011 J. Mach. Learn. Res. 12 2825Google Scholar

  • [1] 欧阳鑫健, 张岩星, 王之龙, 张锋, 陈韦嘉, 庄园, 揭晓, 刘来君, 王大威. 面向铁电相变的机器学习: 基于图卷积神经网络的分子动力学模拟. 物理学报, 2024, 73(8): 086301. doi: 10.7498/aps.73.20240156
    [2] 刘凯越, 李腾耀, 郑娜娥, 田志富, 蔡通, 王彦朝, 曹超华. 基于迁移学习的共形超构表面散射场高效智能计算方法. 物理学报, 2024, 73(23): . doi: 10.7498/aps.73.20241160
    [3] 孙涛, 袁健美. 基于深度学习原子特征表示方法的Janus过渡金属硫化物带隙预测. 物理学报, 2023, 72(2): 028901. doi: 10.7498/aps.72.20221374
    [4] 雷波, 何兆阳, 张瑞. 基于迁移学习的水下目标定位方法仿真研究. 物理学报, 2021, 70(22): 224302. doi: 10.7498/aps.70.20210277
    [5] 王晨阳, 段倩倩, 周凯, 姚静, 苏敏, 傅意超, 纪俊羊, 洪鑫, 刘雪芹, 汪志勇. 基于遗传算法优化卷积长短记忆混合神经网络模型的光伏发电功率预测. 物理学报, 2020, 69(10): 100701. doi: 10.7498/aps.69.20191935
    [6] 彭向凯, 吉经纬, 李琳, 任伟, 项静峰, 刘亢亢, 程鹤楠, 张镇, 屈求智, 李唐, 刘亮, 吕德胜. 基于人工神经网络在线学习方法优化磁屏蔽特性参数. 物理学报, 2019, 68(13): 130701. doi: 10.7498/aps.68.20190234
    [7] 院琳, 杨雪松, 王秉中. 基于经验知识遗传算法优化的神经网络模型实现时间反演信道预测. 物理学报, 2019, 68(17): 170503. doi: 10.7498/aps.68.20190327
    [8] 魏德志, 陈福集, 郑小雪. 基于混沌理论和改进径向基函数神经网络的网络舆情预测方法. 物理学报, 2015, 64(11): 110503. doi: 10.7498/aps.64.110503
    [9] 孟庆芳, 陈月辉, 冯志全, 王枫林, 陈珊珊. 基于局域相关向量机回归模型的小尺度网络流量的非线性预测. 物理学报, 2013, 62(15): 150509. doi: 10.7498/aps.62.150509
    [10] 李盼池, 王海英, 戴庆, 肖红. 量子过程神经网络模型算法及应用. 物理学报, 2012, 61(16): 160303. doi: 10.7498/aps.61.160303
    [11] 王永生, 孙 瑾, 王昌金, 范洪达. 变参数混沌时间序列的神经网络预测研究. 物理学报, 2008, 57(10): 6120-6131. doi: 10.7498/aps.57.6120
    [12] 张军峰, 胡寿松. 基于多重核学习支持向量回归的混沌时间序列预测. 物理学报, 2008, 57(5): 2708-2713. doi: 10.7498/aps.57.2708
    [13] 张军峰, 胡寿松. 基于一种新型聚类算法的RBF神经网络混沌时间序列预测. 物理学报, 2007, 56(2): 713-719. doi: 10.7498/aps.56.713
    [14] 李 军, 刘君华. 一种新型广义RBF神经网络在混沌时间序列预测中的研究. 物理学报, 2005, 54(10): 4569-4577. doi: 10.7498/aps.54.4569
    [15] 胡玉霞, 高金峰. 一种预测混沌时间序列的模糊神经网络方法. 物理学报, 2005, 54(11): 5034-5038. doi: 10.7498/aps.54.5034
    [16] 崔万照, 朱长纯, 刘君华. 薄膜场发射开启电场的小波神经网络预测模型研究. 物理学报, 2004, 53(5): 1583-1587. doi: 10.7498/aps.53.1583
    [17] 谭 文, 王耀南, 周少武, 刘祖润. 混沌时间序列的模糊神经网络预测. 物理学报, 2003, 52(4): 795-801. doi: 10.7498/aps.52.795
    [18] 神经网络的自适应删剪学习算法及其应用. 物理学报, 2001, 50(4): 674-681. doi: 10.7498/aps.50.674
    [19] 于丽娟, 朱长纯. 用人工神经网络预测场发射开启电场. 物理学报, 2000, 49(1): 170-173. doi: 10.7498/aps.49.170
    [20] 马余强, 张玥明, 龚昌德. Hopfield神经网络模型的恢复特性. 物理学报, 1993, 42(8): 1356-1360. doi: 10.7498/aps.42.1356
计量
  • 文章访问数:  3013
  • PDF下载量:  109
  • 被引次数: 0
出版历程
  • 收稿日期:  2023-06-22
  • 修回日期:  2023-07-31
  • 上网日期:  2023-08-24
  • 刊出日期:  2023-11-05

/

返回文章
返回