Multi-Modal AI-Driven Electrical and Optical Characterization Framework for Sub-5 nm Semiconductor Defect Localization

Abstract

This research proposes a multi-modal AI framework for characterizing electrical and optical sub-5 nm defects within a polypython–matlab co-simulation framework that electron microscopically integrates and nano electically probes optical interference. A hybrid CNN-Transformer-GNN (CTG) fusion model learns spatial and spectral relationships across heterogeneous data and topological channels. The model achieves precise defect location through contrastive feature alignment and cross-modal embedding fusion. COMSOL and TCAD simulations modeling optical near-field scattering and localized current perturbations were used to construct synthetic training datasets. Experimental evaluation shows AI multi modality models outperform single-modality models by over 40% on false positive and less than 3 nm deviation across the 3 nm and 5 nm technology node with above 3 nm localization deviation. The framework shows seamless integration with fab probes for real-time integration and self correcting inference on electron beam, scatterometry, and SEM systems. These results form an initial step towards ecosystems of augmented metrology by AI. Future work will tackle detection of defects through data-synthesis based on physics, and adaptive reinforcement data fusion EUV stochastic metrology information systems.