Cementing Quality Evaluation Based on Physics-Driven Wavefield Separation and Multimodal Transformer

Summary

Accurate evaluation of cementing quality is critical for maintaining wellbore integrity throughout the life cycle of oil and gas wells. However, strong acoustic (AC) interference in complex environments, such as fast formations, causes severe feature aliasing in multimodal logging data. This severely limits the interpretation accuracy and physical reliability of existing deep learning models. To overcome this bottleneck, we propose a physics-driven “decoupling-fusion-evaluation” multimodal deep learning framework. Unlike conventional feature concatenation, we introduce a novel integration strategy guided by AC mechanisms. The front end uses a complementary soft mask mechanism to achieve explicit, input-preserving decomposition of complex wavefields. The back end utilizes a cross-modal attention mechanism, guided by AC-related curve priors, to accurately capture the physical coevolution between 2D wavefields and 1D AC priors. Using a field data set of 1,252,756 depth samples from 37 production wells in the Changqing Oilfield, we systematically benchmarked the proposed CemFormer network against an industrial baseline representing traditional physical interpretation workflows and seven state-of-the-art (SOTA) deep models (including iTransformer, TimesNet, Mamba, etc.). CemFormer achieves an overall accuracy of 89.20% and a macro-F1 score of 86.85%, improving upon this industrial baseline by 5.42% and 6.97%, respectively. Notably, it eliminates evaluation blind spots in fast formations, increasing the recall rate for the highly ambiguous “medium” bonding category at the second interface to 94.69%. Furthermore, under Gaussian noise at an SNR of 15 dB, CemFormer’s macro-F1 decreased by only 2.96 percentage points, from 86.85% to 83.89%, representing the smallest degradation among the compared models. Beyond model-specific gains, ablation experiments indicate that our wavefield decoupling strategy serves as a reusable physical preprocessing mechanism, improving the reported performance of the tested SOTA architectures when processing complex AC features. Furthermore, CemFormer overcomes the computational bottleneck of high-precision physical modeling with extremely low overhead (0.38 M parameters for the CemFormer back end and a complete two-stage single-point latency of 3.34 ms including front-end decoupling). This demonstrates the potential to meet the real-time industrial requirements for “interpretation-while-logging” on field edge devices.