A preprint details a four-layer architecture that processes 3D digital replicas of cultural assets, fuses IoT sensor streams with physics-based simulations, and applies Scientific Machine Learning to forecast degradation (Valentino et al., arXiv:2604.03233, 2026). The central innovation embeds physical laws directly into neural network loss functions via Physics-Informed Neural Networks, enabling solution of both direct and inverse problems on real-world geometries.

This extends Raissi et al. (J. Comput. Phys., 2019) foundational PINN methodology, previously applied to fluid dynamics and material fatigue, into heritage conservation contexts where environmental variables such as humidity and thermal cycling drive crack propagation and erosion. The integration of Proper Orthogonal Decomposition as a Reduced Order Method cuts computational cost for Finite Element-compatible simulations, addressing scalability gaps identified in EU INCEPTION project deployments (2015-2020) that prioritized 3D documentation over predictive physics. Original coverage of the preprint omits how the open-source 3D processing pipeline corrects mesh artifacts that commonly invalidate physics simulations on scanned statues and monuments.

Synthesizing these elements with UNESCO's 2021 AI-for-Heritage guidelines, which emphasized digitization but omitted hybrid scientific ML, reveals an overlooked pathway from reactive sensor logging to proactive parameter identification. Patterns from structural health monitoring of the Leaning Tower of Pisa sensor array (ongoing since 2001) demonstrate IoT data abundance without physics-constrained models; the proposed framework closes that loop, allowing material property inference from sparse readings while maintaining consistency with conservation-domain PDEs.