Integrating Artificial Intelligence, Physics, and Internet of Things: A Framework for Cultural Heritage Conservation

This paper proposes a novel, open-access framework for cultural heritage conservation that integrates IoT, 3D digital replicas, and Scientific Machine Learning—specifically combining Physics-Informed Neural Networks (PINNs) with Reduced Order Methods (ROMs)—to enable efficient, physics-aware monitoring and predictive maintenance of cultural assets.

The Gist

Imagine you have a precious, ancient stone statue in a museum. Over time, the wind, rain, and temperature changes slowly wear it down. Traditionally, experts would have to guess when to fix it, often reacting after damage appears.

This paper proposes a "Super-Brain" system to protect these treasures before they break. Think of it as building a Digital Twin (a perfect virtual copy) of the statue that doesn't just look like the real thing, but thinks and feels like it too.

Here is how the system works, broken down into simple concepts and analogies:

1. The Three Pillars of the System

The authors combine three powerful tools to create this guardian:

• The Internet of Things (IoT): Imagine tiny, invisible sensors glued to the statue, constantly whispering data to the computer about temperature, humidity, and vibration. It's like the statue having a nervous system.
• Artificial Intelligence (AI): This is the "student" that learns patterns from the data.
• Physics: This is the "teacher." It provides the unbreakable laws of nature (like how heat moves or how metal rusts).

The Innovation: Usually, AI just guesses based on data, and Physics just calculates based on theory. This paper teaches the AI to listen to the Physics teacher while it studies. This ensures the AI's predictions are not just statistically likely, but physically possible.

2. The Four-Layer "Factory"

The authors built a four-story factory to process this information:

• Floor 1: The Acquisition Layer (The Eyes and Ears)

  • What it does: It takes a 3D scan of the statue (like a high-tech photo) and gathers data from the sensors.
  • The Magic Trick: They use a tool called Blender (a popular 3D animation software) as a translator. It takes the messy, complex 3D shape of a statue and automatically turns it into a neat grid of points (a mesh) that the computer can understand. It's like turning a pile of Lego bricks into a perfect instruction manual.

• Floor 2: The Knowledge-Base Layer (The Library)

  • What it does: This is the storage room. It keeps the 3D models, the sensor data, and the "rules" of physics organized so the computer can find them instantly.

• Floor 3: The Inference Engine (The Brain)

  • What it does: This is where the magic happens. It uses two special techniques:
    • PINNs (Physics-Informed Neural Networks): Imagine a student taking a test. If they get an answer wrong, the teacher (Physics) corrects them immediately. PINNs do this constantly, ensuring the simulation follows the laws of nature.
    • ROMs (Reduced Order Models): Solving complex physics equations is like trying to count every grain of sand on a beach—it takes forever. ROMs are like a "smart summary." They figure out the most important patterns so the computer can solve the problem in seconds instead of days. It's like reading a movie summary instead of watching the whole 3-hour film to know the ending.

• Floor 4: The Application Layer (The Dashboard)

  • What it does: This is the screen where the museum curator sees the results. They can see a 3D map of the statue showing exactly where it might crack next year, or where the temperature is too high.

3. How They Tested It (The "Video Game" Scenarios)

To prove this works, they didn't just talk about it; they ran simulations on digital versions of real objects:

• A Rock: They simulated how heat moves through a weirdly shaped rock. They used the system to "guess" the hidden properties of the rock based on temperature data, and the system got it right.
• A Column: They simulated a column getting hot and cold over time. The system predicted how the heat would spread and identified the exact rate of change.
• A Temple: They tested it on a complex temple made of many parts to ensure the system could handle complicated shapes.

4. Why This Matters

• Predictive Maintenance: Instead of waiting for a statue to crack, this system says, "Hey, in three years, this specific spot will get too hot and crack. Let's fix it now."
• Speed: Because of the "smart summary" (ROMs), the computer can run these complex simulations in real-time.
• Open Source: The authors put all their code on GitHub (a public code library). This means other scientists and museums can use their tools for free to protect their own treasures.

The Bottom Line

This paper presents a universal toolkit for saving history. It combines the "gut feeling" of data with the "hard facts" of physics, wrapped in a system that can automatically understand any 3D shape. It turns the slow, reactive work of conservation into a fast, proactive science, ensuring our cultural heritage survives for future generations.

Technical Summary

1. Problem Statement

The conservation of cultural heritage (CH) requires effective monitoring and predictive maintenance to preserve historical assets. Current approaches often suffer from fragmentation:

• Lack of Integration: Existing workflows typically rely on either numerical simulations (Finite Element Methods - FEM), data-driven AI (Machine Learning), or Internet of Things (IoT) sensor data in isolation.
• Computational Cost: High-fidelity simulations (FEM) are computationally expensive, making real-time or "many-query" analysis difficult for complex 3D geometries.
• Data-Physics Gap: Purely data-driven AI models lack physical interpretability and generalizability, while purely physics-based models struggle with sparse real-world sensor data and complex, irregular geometries common in CH.
• Workflow Discontinuity: There is no unified framework that seamlessly processes raw 3D digital replicas (from laser scanning/photogrammetry) into simulation-ready formats while integrating IoT data and physical laws.

2. Methodology

The authors propose a four-layer modular framework that integrates IoT, AI (specifically Scientific Machine Learning), and Physics. The core innovation lies in combining Physics-Informed Neural Networks (PINNs) with Reduced Order Models (ROMs) based on Proper Orthogonal Decomposition (POD).

A. Framework Architecture

The system is structured into four functional layers:

1. Acquisition Layer:
   • Sensor Module: Collects environmental data (temperature, corrosion, pollution) via IoT.
   • API & Open Data: Integrates external environmental data.
   • 3D Model Module: The core innovation for geometry processing. It uses Blender APIs to automatically process raw 3D models (e.g., .blend files). It triangulates meshes, extracts boundary/collocation points, and generates structured data (JSON, .msh, .xdmf) required for simulations.
2. Knowledge-Base Layer: A central database storing sensor data, 3D model metadata, mesh information, and pre-processed training data.
3. Inference Engine Layer: The computational core containing:
   • Offline Module: Uses FEM to generate high-fidelity "snapshots" for a range of parameters. It applies Proper Orthogonal Decomposition (POD) to construct a low-dimensional reduced basis (ROM).
   • Online Module: Uses the ROM to solve problems in real-time (Online Solver) for new parameter sets.
   • PINN Module: Solves Direct Problems (predicting fields given known parameters) and Inverse Problems (identifying unknown physical parameters from sparse sensor data) by embedding physical laws (PDEs) into the neural network loss function.
   • msh2xdmf: Converts simulation results into visualization formats.
4. Application Layer: Provides a dashboard for experts to upload models, visualize simulations, and download results.

B. Key Technical Components

• Physics-Informed Neural Networks (PINNs): These networks minimize a loss function comprising the PDE residual, boundary conditions, initial conditions, and data mismatch. This allows the model to learn solutions that satisfy physical laws even with limited data.
• Reduced Order Models (ROMs): Specifically using POD, the framework compresses high-dimensional FEM solutions into a low-dimensional subspace. This enables rapid "many-query" evaluations (e.g., real-time monitoring) by solving the problem in the reduced space and lifting the solution back to the full space.
• Hybrid Workflow: The framework can use ROM for fast forward simulations and PINNs for parameter identification (inverse problems), or use PINNs directly for direct problems where parameters are known but data integration is needed.

3. Key Contributions

1. Automated 3D Model Processing Pipeline: A novel methodology using Blender APIs to automatically convert complex, irregular 3D cultural asset models into simulation-ready meshes and sampling points, bridging the gap between digital twins and numerical solvers.
2. Unified SciML Framework: The first framework to simultaneously integrate IoT sensor data, PINNs, and ROMs for cultural heritage. It combines the interpretability of physics with the adaptability of data-driven approaches.
3. Inverse and Direct Problem Solving: Demonstrates the ability to identify unknown physical parameters (e.g., material properties, source terms) from sparse sensor data using PINNs, and to perform rapid real-time simulations using ROMs.
4. Open-Source Reproducibility: The authors provide a public GitHub repository containing the code, tools (e.g., bl2pina, bl2msh), and experimental setups, ensuring the framework is accessible and extensible.

4. Experimental Results

The framework was validated on four benchmark test cases involving realistic 3D geometries (a rock, a column, and a temple):

• Test Problem 1 (Poisson on a Rock - Inverse):
  • Goal: Identify parameters $(\lambda, \alpha, \beta)$ for a heat distribution problem using simulated sensor data.
  • Result: PINNs successfully identified parameters with relative errors $< 1.3\%$. The subsequent ROM simulation using these identified parameters achieved a total relative error of $7.77 \times 10^{-3}$ compared to the analytical solution.
• Test Problem 2 (Parabolic System on a Column - Inverse):
  • Goal: Solve a time-dependent coupled heat equation to identify parameters.
  • Result: Despite the added complexity of time-dependence, PINNs identified parameters with errors in the range of $10^{-2}$ to $10^{-3}$. The ROM surrogate achieved a maximum point-wise error of $10^{-8}$ compared to the full-order solution.
• Test Problem 3 (Temperature Monitoring on a Rock - Direct):
  • Goal: Monitor temperature distribution using PINNs with known physics but data-driven boundary conditions (simulating IoT sensors).
  • Result: The framework successfully integrated hard constraints for initial conditions and weighted loss functions. The PINN solution matched the FEM benchmark with an $L^2$ relative error of $8.70 \times 10^{-3}$.
• Test Problem 4 (Diffusion-Reaction on a Column - Direct):
  • Goal: Compare a "physics-only" PINN approach against a "data-integration" approach.
  • Result: The data-integration approach yielded lower errors (e.g., $4.10 \times 10^{-3}$ for magnitude) compared to the physics-only approach, validating the benefit of incorporating sensor data into the training process.

5. Significance and Impact

• Predictive Maintenance: The framework enables real-time monitoring and prediction of degradation (e.g., corrosion, thermal stress) on complex historical structures, moving from reactive to proactive conservation.
• Handling Complexity: By automating the meshing and sampling of irregular 3D geometries, the system makes advanced numerical simulations accessible for non-standard cultural assets that are difficult to model manually.
• Efficiency: The integration of ROMs reduces computational costs by orders of magnitude, allowing for "many-query" scenarios (e.g., testing multiple weather conditions) that would be infeasible with standard FEM.
• Robustness: The hybrid approach ensures that predictions remain physically consistent even when sensor data is sparse or noisy, addressing a critical limitation of pure AI models in safety-critical applications.
• Open Science: By releasing the code and tools, the authors lower the barrier to entry for researchers and practitioners, fostering a community-driven approach to digital heritage conservation.

In conclusion, this paper presents a robust, scalable, and open framework that successfully unifies IoT, AI, and Physics to create a "Digital Twin" capable of reliable, real-time predictive maintenance for cultural heritage assets.