The accelerated design of next-generation semiconductor interconnects faces a critical “applicability gap”. Purely data-driven models effectively navigate vast chemical spaces, but they often yield candidates that are theoretically performant yet violate practical manufacturing constraints. To bridge this disconnect, this study proposes a neuro-symbolic decision support framework that systematically integrates inductive graph learning with deductive engineering logic for Safe-by-Design material screening. The framework operates through a hierarchical dual-stream architecture. First, an inductive Graph Neural Network (GNN) engine transforms 3D crystal structures into topological graph representations to predict thermodynamic stability and metallicity with high discriminative power (AUC = 0.868). Second, a deductive safety layer enforces explicit domain ontology, including toxicity thresholds, raw material costs, and reactivity limits, to preemptively prune high-risk candidates. Operationally, this hybrid approach reduces the candidate search space of over 20,000 compounds by approximately 96% within minutes, demonstrating orders-of-magnitude computational efficiency gains over traditional ab initio high-throughput screening. The system’s reliability is further validated through structural perturbation analysis and high-fidelity physics simulations, identifying robust binary compounds such as HfB and NbAl3 that exhibit cohesive energies up to 2.1 times that of copper. These results demonstrate the efficacy of integrating symbolic reasoning with deep learning to create transparent, reliability-aware computational tools for early-stage engineering decision-making.
Tang et al. (2026) studied this question.