Machine-learning models in physical sciences often assume that complex emergent properties can be predicted using a single homogeneous inference task. However, many physical phenomena are intrinsically conditional, involving distinct processes governing phase feasibility and physical-property optimization. Here we show that superconductivity prediction is more naturally formulated as a hierarchical inference problem in which superconducting feasibility and critical-temperature $\left(T_{\mathrm{c}}\right)$ optimization are represented by a shared descriptor space with task-dependent reweighting. Using a balanced dataset of more than 26 000 inorganic materials comprising both superconductors and nonsuperconductors, we demonstrate that superconducting feasibility is primarily controlled by an electronic–compositional backbone, whereas achievable $T_{\mathrm{c}}$ depends on additional structural and magnetic/orbital effects. Guided by this decomposition, we construct a physics-aligned hierarchical learning framework that first predicts superconducting feasibility using the electronic–compositional backbone and subsequently predicts $T_{\mathrm{c}}$ using an expanded descriptor set. Feature clustering and Shapley additive explanation-based attribution further show that the electronic–compositional backbone provides a minimal dominant representation of feasibility, while additional descriptor groups modulate achievable $T_{\mathrm{c}}$ through constraint and context-dependent effects. The two-stage framework reduces end-to-end $T_{\mathrm{c}}$ prediction error by explicitly accounting for the conditional nature of superconductivity. These results establish superconductivity prediction as a hierarchically structured problem and provide a physically interpretable bridge between data-driven discovery and first-principles modeling, defining a sequential design principle for superconductor discovery. More broadly, the results show that physically interpretable machine-learning behavior can emerge from hierarchical organization and conditional reweighting within descriptor space rather than from increasingly complex model architectures.

Close

PrevNext

View figure in article

• Electrical conductivity
• Electrical properties
• Superconductivity
• Data mining
• Machine learning