Autoencoder‑Augmented Graph Neural Networks for Accurate and Scalable Structure Recognition in Analog/Mixed‑Signal Schematics
Auteur(s): Kyamakya KyandoghereNom de la revue/Journal: IEEE ACCESS
Mois: juillet
Volume: 13
Numéro: DOI: 10.1109/ACCESS.2025.3591720
Année: 2025
pages: 129721 - 129740
Résumé
<p><span style="background-color:rgb(255,255,255);color:rgb(34,34,34);">The increasing complexity of Analog/Mixed-Signal (AMS) schematics has been posing significant challenges in structure recognition, particularly in the intellectual property (IP) industry, where data scarcity and confidentiality constraints limit model training. In this work, a novel framework has been proposed that combines the generative augmentation capabilities of convolutional Autoencoders with the structural analysis power of Graph Convolutional Networks (GCNs). Realistic schematic variants have been synthesized from limited proprietary data to enhance model generalization, while the GCN has been used to capture topological features critical to substructure recognition. The method has been validated on a curated AMS dataset, where it surpassed a GCN-only baseline by reducing reconstruction error and achieving a balanced classification accuracy of 96.7%, thereby exceeding the long-standing 95 … </span></p>