Acoustic Impedance Prediction Using an Attention-Based Dual-Branch Double-Inversion Network

TL;DR

This paper introduces ADDIN-I, an attention-based dual-branch deep learning network that improves seismic impedance inversion accuracy and resolution by effectively capturing high-frequency signals and adapting to wavelet uncertainties.

Contribution

The paper presents a novel dual-branch architecture with attention mechanisms and wavelet adaptation for enhanced seismic impedance inversion.

Findings

Outperforms traditional methods on synthetic data

Achieves high-resolution impedance predictions on real data

Effectively captures weak high-frequency signals

Abstract

Seismic impedance inversion is a widely used technique for reservoir characterization. Accurate, high-resolution seismic impedance data form the foundation for subsequent reservoir interpretation. Deep learning methods have demonstrated significant potential in seismic impedance inversion. Traditional single semi-supervised networks, which directly input original seismic logging data, struggle to capture high-frequency weak signals. This limitation leads to low-resolution inversion results with inadequate accuracy and stability. Moreover, seismic wavelet uncertainty further constrains the application of these methods to real seismic data. To address these challenges, we propose ADDIN-I: an Attention-based Dual-branch Double-Inversion Network for Impedance prediction. ADDIN-I's dual-branch architecture overcomes the limitations of single-branch semi-supervised networks and improves the extraction of high-frequency weak signal features in sequence modeling. The network incorporates an attention mechanism to further enhance its feature extraction capabilities. To adapt the method for real seismic data applications, a deep learning forward operator is employed to fit the wavelet adaptively. ADDIN-I demonstrates excellent performance in both synthetic and real data applications.