CARE-GNN: A Causal-Aware and Rule-Enhanced Graph Neural Network on Domain Knowledge Graphs for Cementing Quality Diagnosis and Optimization

Summary Accurate cementing quality diagnosis is frequently impeded by the inherent “black-box” nature of conventional models and the fragmented semantic reasoning in multisource engineering data. To address these limitations, we propose a Causal-Aware Rule-Enhanced (CARE) Graph Neural Network (GNN)(CARE-GNN) framework, which specifically aims to enhance the representation and reasoning performance of domain-specific knowledge graphs (KGs) by integrating causal reasoning with the rule-based physical constraints. The core innovation features a dual-channel encoder that mathematically disentangles data-driven topological features from rule-based constraints via an attention gating unit, enabling an adaptive balance between statistical patterns and engineering logic. Unlike parallel multitask paradigms, a sequential causal chain is implemented to model the intrinsic dependencies among root cause diagnosis, quality prediction, and optimization recommendation, ensuring that identified physical precursors directly inform downstream decision-making. Integration of a physics-informed regularization loss further constrains the reasoning space to honor fundamental engineering boundaries. Comparative evaluations yield a quality prediction accuracy of 91.07% and a 6.63% improvement in the multitask hit rate (QRM@1) over state-of-the-art baselines. Field validation on Well [Bozi-X] confirms the system’s capability to provide interpretable, expert-aligned diagnostic insights, offering a theoretically grounded solution for intelligent cementing engineering.