Accelerating Superconductor Discovery with Disorder-Aware Multi-Fidelity Graph Transformers

Discovering new superconductors with high critical temperature Tc remains challenging because reliable experimental data are scarce and real materials frequently exhibit crystallographic disorder that is difficult to represent faithfully in machine-learning models. We present FiDeTGraph, a disorder-aware, multi-fidelity SE(3)-equivariant graph transformer for predicting superconducting Tc from 3D crystal structures. FiDeTGraph addresses two practical bottlenecks in data-driven superconductivity: (i) the tendency of prior disorder handling to collapse substitutional/positional disorder into occupancy-weighted averages that obscure local variations, and (ii) the limited coverage of high-fidelity superconducting datasets. To capture disorder explicitly, we introduce virtual nodes placed at the centroid of each disordered cluster and connect them bidirectionally to the constituent atomic instances, enabling the network to learn higher-order disorder effects without linear mixing. To mitigate data scarcity, we propose a multi-fidelity learning strategy that combines high-fidelity SuperCon3D structures with low-fidelity 3DSCMP and supercell-permuted variants that sample plausible disorder realizations, and we condition predictions on a learned graph-level fidelity embedding. Across 10-fold cross-validation on SuperCon3D, virtual-node modeling improves performance over occupancy-summation baselines, and FiDeTGraph further benefits from low-fidelity augmentation, yielding consistent gains on both the full set and the disorder-only subset. Finally, we integrate FiDeTGraph into a high-throughput screening pipeline over the ICSD and Materials Project databases, applying stability and metallicity filters to prioritize synthesizable candidates for downstream first-principles verification. Our results show that jointly modeling disorder and exploiting multi-fidelity data can substantially improve Tc prediction and accelerate the identification of promising superconducting materials.