An anti-noise block and frequency-aware framework in deep learning for formation resistivity prediction by transient electromagnetic data

Formation resistivity prediction is critical for understanding subsurface fluid behavior. Traditional methods struggle to accurately measure subsurface fluid parameters from cased wells, where a transient electromagnetic method can be effectively applied. However, resistivity prediction using transient electromagnetic method data faces two significant challenges: high-frequency disaster and environmental noise. These factors collectively hinder the ability of neural networks to capture high-frequency features and to handle noise interference, diminishing prediction accuracy. To address these challenges, this study proposes a frequency-aware framework and a temporal anti-noise block. The frequency-aware framework addresses high-frequency disaster by employing a dual-stream structure with wavelet transformation to isolate and learn high-frequency components. Meanwhile, the temporal anti-noise block mitigates environmental noise by denoising temporal features through a soft threshold attention mechanism. Using a long short-term memory model as a baseline, these enhancements are integrated to conduct two experiments. The ablation experiment demonstrates that the proposed block and framework significantly improve prediction accuracy, achieving an R2 of 0.91 (0.22 higher than the baseline) with significant gains in handling high-frequency features (an improvement of 0.30 in R2 at high-frequency part). The robustness experiment shows that the temporal anti-noise block reduces the impact of Gaussian and impulse noise by 1/8 compared to the baseline, confirming its strong noise resistance. This study achieves accurate formation resistivity prediction using transient electromagnetic method data, paving the way for advanced subsurface fluid behavior analysis in complex geological settings.