Enhancing Graph Neural Network for Boolean Satisfiability Solving via Data Augmentation

Boolean Satisfiability (SAT) is a well-known NP-complete problem that lies at the core of many applications in formal verification, planning, and artificial intelligence. While classical SAT solvers have achieved impressive results on both synthetic and industrial benchmarks, they solve each instance independently without leveraging prior experience. Graph Neural Network (GNN)-based SAT solvers offer a learning-driven approach with the potential to transfer knowledge across problem instances and achieve better generalization performance. However, most existing work in this area focuses heavily on model architecture designs, with limited attention paid to data augmentation techniques that could improve out-of-domain generalization. In this work, we explore two simple data augmentation strategies applied during training and analyze their impact on both in-domain accuracy and out-of-domain generalization. Our findings suggest new directions for enhancing the performance and generalization ability of GNN-based SAT solvers.