CurveNet: Scale-Specific Curvelet-Domain Deep Learning for Acoustic Well-Logging Data Denoising

Abstract

Borehole reflection imaging (BRI) is an important technique for characterizing subsurface geological structures near a borehole using acoustic wave reflections from surrounding formations. In practice, the recorded reflection signals often suffer from a low signal-to-noise ratio (SNR) due to the complexity of borehole wave propagation, strong energy attenuation in the formation, and various noise sources, which significantly degrades imaging quality. To address this challenge, we propose a novel deep learning-based denoising method, termed CurveNet, that combines the curvelet transform with a U-Net architecture. By exploiting the multi-scale and multi-angle properties of the curvelet transform, CurveNet learns representative signal features from the data, reducing reliance on manual feature engineering and supporting automated denoising. In addition, the multi-scale nature of the curvelet transform enables scale-specific denoising networks with lower measured network-operation counts than UnifiedNet under the evaluated configuration. Numerical simulations evaluate the denoising performance of CurveNet, and a field-data application demonstrates its within-well adaptation to Well X. Under the tested synthetic-data conditions, CurveNet improves the SNR by approximately 5.4 dB over the UnifiedNet baseline and by approximately 8.5 dB over the best-performing traditional method on average across the three evaluated noise levels. These results demonstrate the potential of the scale-specific architecture for enhancing borehole reflection data, while the within-well application indicates its practical potential after well-specific adaptation.