Scientific literature reports fine-grained entity properties, but such knowledge is scattered across papers, expressed with heterogeneous names, and scoped by experimental conditions. Closed-schema extraction requires predefined attribute vocabularies that are hard to enumerate for property-rich scientific entities, while open-schema extraction often yields duplicate and condition-specific attribute labels. We introduce entity-centric attribute schema induction, which constructs a hierarchical type-level attribute schema from a scientific corpus. We propose ObSchema an observation-to-schema framework that extracts local entity-property observations, aggregates corpus-level evidence to identify recurring attributes, and refines them through value-aware merging and adaptive grouping. On two physical science domains, ObSchema consistently outperforms strong LLM-based schema-induction baselines across different backbone models, producing more compact and coherent schemas.