A study of the effect of hyperparameters GRU-CNN hybrid deep learning EI inversion
LIANG Li-Feng1(), LIU Xiu-Juan1(), ZHANG Hong-Bing2, CHEN Cheng-Hao1, CHEN Jin-Hua1
1. Department of Geography,Lingnan Normal University,Zhanjiang 524057,China 2. School of Earth Science and Engineering,Hohai University,Nanjing 210098,China
Previous studies have shown that CNN-GRU hybrid deep learning inversion EI has the advantages of strong applicability and strong generalization capability.However,there are many pre-stack inversion parameters based on deep learning,such as internal deep learning network learnable parameters and external hyperparameters.At present,there is still no systematic research on the impact of hyperparameter selection on network performance and computing speed,which will directly affect the further promotion and application of the method.Therefore,based on the hybrid deep learning inversion elastic impedance,this paper discusses the impact of five hyperparameters,i.e.,learning rate,Epoch,batch_size,regularization parameter,and the number of wells participating in network training on network performance and calculation speed,thus providing a basis for studying the selection of seismic inversion hyperparameters.The research results can provide a feasible quality control method for three-dimensional large-area deep learning inversion,which is of certain significance for promoting the wide application of deep learning methods in petroleum geophysical prospecting.
梁立锋, 刘秀娟, 张宏兵, 陈程浩, 陈锦华. 超参数对GRU-CNN混合深度学习弹性阻抗反演影响研究[J]. 物探与化探, 2021, 45(1): 133-139.
LIANG Li-Feng, LIU Xiu-Juan, ZHANG Hong-Bing, CHEN Cheng-Hao, CHEN Jin-Hua. A study of the effect of hyperparameters GRU-CNN hybrid deep learning EI inversion. Geophysical and Geochemical Exploration, 2021, 45(1): 133-139.
Wang Y C, Liu L T, Xu H Z. The identification of gravity anomaly body based on the convolutional neural network[J]. Geophysical and Geochemical Exploration, 2020,44(2):394-400.
[2]
Russakovsky O, Deng J, Su H, et al. Image net large scale visual recognition challenge[J]. Int. J. Comput. Vis., 2015,115(3):211-252.
doi: 10.1007/s11263-015-0816-y
Han W X, Zhou Y T, Chi Y. Deep learning convolutional neural networks for random noise attenuation in seismic data[J]. Geophysical Prospecting for Petroleum, 2018,57(6):862-869,877.
Lin N T, Fu C, Zhang D, et al. Prediction of petroleum reservoirs under unsupervised and supervised learning[J]. Geophysical Prospecting for Petroleum, 2018,57(4):601-610.
[7]
Wu X M, Shi Y Z, Fomel S, et al. Fault net 3D:Predicting fault probabilities,strikes,and dips with a single convolutional neural network[J]. IEEE Transactions on Geoscience and Remote Sensing, 2019,57(11):9138-9155.
doi: 10.1109/TGRS.36
[8]
Alfarraj M, AlRegib G. SEG technical program expanded abstracts 2018[M]. Houston:Society of Exploration Geophysicists, 2018: 2141-2146.
An P, Cao D P, Zhao B Y, et al. Research on prediction method of reservoir physical property parameters based on LSTM recurrent neural network[J]. Progress in Geophysics, 2019,34(5):1849-1858.
Zhang X D, Chen Y T, Meng J. Logging curve generation method based on cyclic neural network[J]. Petroleum Exploration and Development, 2018,45(4):598-607.
Deng S. Hyper-parameter optimization of CNN based on improved Bayesian optimization algorithm[J]. Application Research of Computers, 2019,36(7):1984-1987.
[15]
Smith L N. Cyclical learning rates for training neural networks[C]// 2017 IEEE Winter Conference on Applications of Computer Vision (WACV),IEEE, 2017: 464-472.
[16]
Breuel T M. The effects of hyperparameters on SGD training of neural networks[EB/OL]. Computer Science, 2015. [2015-08-12]. https://arxiv.org/abs/150802788.
[17]
Martin G S, Wiley R, Marfurt K J. Marmousi2:An elastic upgrade for marmousi[J]. The Leading Edge, 2006,25:156-166.
doi: 10.1190/1.2172306
