新疆大学学报（自然科学版中英文）2025，Vol.42Issue(3)：280-299,20.DOI:10.13568/j.cnki.651094.651316.2025.03.31.0005
遗传算法在瞬变电磁深度学习反演中的优化策略
Optimization Strategy of Genetic Algorithm in Transient Electromagnetic Deep Learning Inversion
摘要
Abstract
Transient electromagnetic(TEM)1D inversion has been widely applied in geological engineering,yet these conventional methods remain constrained by strong dependence on initial models,limited noise resistance,and inefficiency in achieving real-time inversion.To address these challenges,we propose a convolutional neural network-long short-term memory hybrid architecture tailored to the inherent characteristics of TEM inversion.Using loop-source TEM 1D forward modeling,we generate training data comprising sampling time-decay voltage pairs as network inputs.An optimization strategy combining the Adam optimizer with the ReduceLROnPlateau learning rate scheduler is implemented to adaptively adjust learning rates during parameter updates.In view of the problem that the hyper-parameter setting of the current network structure depends on the empirical value and lacks scientificity,which leads to the waste of computing power and time,the genetic algorithm is proposed to optimize the hyper-parameters of the neural network structure in the model training stage to reduce the training cost and improve the model performance.The output layer provides subsurface resistivity-depth profiles corresponding to input electromagnetic responses,enabling deep learning-based TEM inversion.The trained GA-CNN-LSTM network demonstrates robust performance in real-time predictions for randomly generated three-layer and five-layer models,with validation metrics yielding R2＞0.9.Further evaluation using noise-contaminated data reveals that the optimized network achieves an average inversion time of 0.13 s and a structural similarity index of 90.138％across four common models,outperforming both Occam and LSTM inversion methods.Generalization capability is validated through successful inversion of 3D forward modeling data.These results demonstrate the algorithm's reliability,computational efficiency,and practical utility in complex geological scenarios.Finally,Occam inversion and neural network inversion are carried out on the measured data respectively.The trained neural network only take 0.73 s to complete the inversion accuracy,which verifies the practicability of the algorithm in this paper.关键词
瞬变电磁法/深度学习/遗传算法/超参数优化Key words
transient electromagnetic method/deep learning/genetic algorithm/hyper-parameter optimization分类
天文与地球科学引用本文复制引用
吴文宇,张莹莹,吴新宇,谢斌..遗传算法在瞬变电磁深度学习反演中的优化策略[J].新疆大学学报（自然科学版中英文）,2025,42(3):280-299,20.基金项目
新疆维吾尔自治区重点研发计划"新疆天山及邻区早二叠世金镍重大成矿事件与找矿评价关键技术开发"(2024B03005-3). （2024B03005-3）
