中南大学学报（自然科学版）2026，Vol.57Issue(4)：1636-1648,13.DOI:10.11817/j.issn.1672-7207.2026.04.015
自适应损失权重的多任务大地电磁深度学习反演
Multi-task deep learning inversion for magnetotellurics using adaptive loss weighting
摘要
Abstract
Data-driven deep learning has great potential in solving the inverse problems of magnetotellurics.To enhance the generalization ability of deep learning in the inversion regression problems of magnetotellurics,a deep learning inversion method was investigated based on an adaptive loss weighting strategy.A"W"-shaped deep learning model MSEAdpNet was constructed,which combined a Transformer and convolutional neural network(CNN)for multi-task learning,guiding the model training through adaptive loss weighting feedback.The model took noisy data as input and consisted of a shared encoder and two independent decoders.One decoder estimated the model error for the inversion task,while the other estimated the response data error for the auxiliary task.By using adaptive error weighting,the model ensured the priority of the inversion task while improved overall task performance.The results demonstrate that the multi-task deep learning approach with an adaptive loss weighting strategy can achieve good inversion results,effectively reduce validation errors,and it is validated with real measurement results.关键词
多任务深度学习/大地电磁反演/自适应损失权重/神经网络Key words
multi-task deep learning/magnetotellurics inversion/adaptive loss weighting/neural network分类
天文与地球科学引用本文复制引用
樊娟,席振铢,王鹤,刘英锋,朱开鹏..自适应损失权重的多任务大地电磁深度学习反演[J].中南大学学报（自然科学版）,2026,57(4):1636-1648,13.基金项目
黔科合重大专项([2024]029) （[2024]029）
国家重点研发计划项目(2024YFC2909203,2022YFC2903404)(Project([2024]029)supported by Guizhou Department of Science and Technology （2024YFC2909203,2022YFC2903404）
Projects(2024YFC2909203,2022YFC2903404)supported by the National Key Research and Development Program of China) （2024YFC2909203,2022YFC2903404）
