Valentino and colleagues propose a layered conservation framework in which cultural heritage becomes a continuously sensed, simulated and predictive environment. IoT acquisition, three-dimensional models, physics-informed neural networks, reduced-order methods and finite-element analysis are integrated so that degradation can be studied through both data and physical law. The work’s conceptual strength lies in refusing the usual opposition between empirical learning and mechanistic modelling. Scientific machine learning operates here as a hybrid epistemology: observed signals constrain simulation, while material knowledge disciplines statistical inference. Automated processing of complex geometries extends the framework from abstract prediction to operational digital replicas. Heritage conservation is consequently reframed from episodic inspection toward anticipatory maintenance. The broader intellectual bridge concerns the ontology of preservation. A monument is not a static object protected from time, but a changing material system whose future states can be modelled, compared and acted upon. Conservation becomes an infrastructure of temporal care, joining sensing, interpretation and intervention within a recursive technical ecology.