Journal of Geo-Energy and Environment | Volume 3, Issue 1: 72-83, 2027 | DOI: 10.62762/JGEE.2026.158449
Abstract
Geo-energy systems involve heterogeneous subsurface properties, irregular spatial structures, and complex interactions. Conventional data-driven models either treat observations independently or rely on graph connectivity without explicitly encoding relative geometry, which limits their ability to distinguish spatially different configurations. This study develops a Geometry-Aware Graph Transformer (GAGT) that integrates node attributes, graph structure, and relative geometric information through a learnable geometric bias in the attention mechanism. A controlled synthetic experiment with 300 irregularly distributed three-dimensional nodes compares a multilayer perceptron (MLP), a Graph Tran... More >
Graphical Abstract