AIG2PT: A Generative Pre-trained Transformer for Unconditional And-Inverter Graph Synthesis

And-Inverter Graphs (AIGs) underpin modern logic synthesis and optimization, yet existing techniques explore only a narrow portion of the vast design space of valid circuit structures. To address this limitation, this paper introduces AIG2PT, a foundational building block toward structurally diverse, future function-aware, and equivalence-guided AIG generation. Inspired by advances in molecular generation, AIG2PT decouples structural synthesis from functional constraints, enabling the model to learn the syntax of valid, expressive, and structurally novel circuit topologies through unconditional generation. The approach adapts a Graph Generative Pre-trained Transformer (G2PT) with a domain-specific vocabulary and a topology-aware encoding. A systematic study of sampling strategies reveals tunable control over Validity, Uniqueness, and Novelty (V.U.N.) scores: 100 % structural validity and uniqueness with Diverse Beam Search, and over 60 % novelty with Multinomial Sampling. These results substantially outperform established deep learning-based graph-generation baselines, demonstrating that AIG2PT provides a robust foundation for future work in generative circuit design, conditional synthesis, and broad design space exploration.