2D magnetotelluric forward modeling based on multitask deep learning

Abstract

The accuracy and efficiency in 2D magnetotelluric (MT) forward modeling determine inversion quality. Traditional numerical methods, while achieving reliable results on high-performance computing clusters, face challenges of heavy computational burden and inefficiency when implemented on personal computers due to their computationally intensive nature, this study proposes a novel 2D MT forward modeling method based on a Transformer U-Net (T-Unet) multitask network. Through end-to-end training, the network establishes a mapping relationship between geoelectric models and apparent resistivity as well as phase, generates corresponding datasets, and obtains a neural network weight model capable of directly predicting MT forward modeling results after training. Experiments show that, after model establishment, the T-Unet model significantly shortens the computation time compared with traditional numerical simulations while maintaining high computational accuracy. This research reveals the potential of deep learning neural networks to accelerate MT forward calculations and provides a new pathway for the deep integration and application of artificial intelligence in geophysical exploration.