Existing implicit three-dimensional (3D) geological modelling methods predominantly rely on techniques such as Kriging interpolation and radial basis functions (RBFs); these approaches struggle to capture the nonlinear characteristics of complex geological structures and are limited in their ability to integrate multi-source heterogeneous modeling data. To address these limitations, we propose a 3D geological modelling framework based on a dual-task geological-aware attention network (Geo-GAA). The framework begins with the development of a multi-scale graph neighborhood aggregation mechanism designed to capture geological spatial correlations across different scales. Next, a geological-aware attention mechanism is introduced to explicitly incorporate domain-specific geological knowledge into the framework. Finally, a dual-task prediction head is constructed to jointly perform lithological classification and scalar field regression. Comprehensive ablation studies further validate the contributions of the three core components—graph neighborhood aggregation, geological-aware attention, and dual-task learning. Experimental evaluations in the Lingnian-Ningping area of Guangxi Zhuang Autonomous Region (GZAR), China, demonstrate that the proposed Geo-GAA framework achieves superior performance, yielding 92.1% accuracy for lithological classification and a coefficient of determination (R²) of 0.96 for scalar field prediction. Overall, the proposed framework represents a significant methodological advancement for intelligent modelling of complex geological structures, offering promising applications in deep mineral exploration, geohazard early warning, and digital earth twin technologies.