Abstract: Faults act as critical migration pathways and reservoir space for hydrocarbons. Therefore, accurate fault identification is an important prerequisite for breakthroughs in hydrocarbon exploration. However, conventional manual interpretation suffers from low efficiency, while prediction based solely on artificial intelligence (AI) depends on massive data annotations and lacks geological constraints. In this context, this study developed a geologically constrained semi-supervised AI technique for fault prediction. For this technique, geological insights were transformed into quality control constraints, and a closed-loop process consisting of machine prediction, manual supervision, and model optimization was designed. With a small number of annotations as anchors, this technique allowed for the efficient extraction of multi-scale fault features by combining an improved ResUnet-19A network. Through few-shot and transfer learning, the resulting model can rapidly adapt to a new survey area, significantly improving prediction accuracy and efficiency while substantially reducing labor costs. The practical application in the Chishui area of the southeastern Sichuan Basin indicates that the proposed technique can effectively overcome the limitations of both conventional and pure AI approaches, offering an efficient and reliable solution for fault identification in hydrocarbon exploration within complex areas.