⚡ electrical engineering

A Sensor-Integrated, Explainable Machine-Learning Digital-Twin Framework for Field-Validated Service-Life Prediction of Cementitious Materials under Coupled Chloride–Sulfate Carbonation Degradation

This paper proposes a five-layer, sensor-integrated, explainable machine-learning digital-twin framework that couples multi-hazard mechanistic degradation models with field-updatable AI to predict the service life of cementitious materials under combined chloride, sulfate, and carbonation attacks, addressing critical gaps in existing industry standards and data-limited ML studies.

Original authors: Devesh Ojha, Pawan Singh

Published 2026-10-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Devesh Ojha, Pawan Singh

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Concrete is the most widely used building material on Earth, forming the backbone of our bridges, roads, and skyscrapers. Yet, like all materials, it is not immortal. Over decades, invisible chemical forces begin to eat away at the steel reinforcement hidden inside. Salt from the ocean or winter roads seeps in, sulfur compounds from the air or soil react with the cement, and carbon dioxide from the atmosphere slowly neutralizes the concrete's natural defenses. When these processes combine, they can cause the steel to rust and expand, cracking the concrete from the inside out. For engineers tasked with keeping these structures safe, predicting exactly when this failure will happen is a matter of public safety and economic necessity. Currently, the tools used to make these predictions often look at each chemical attack in isolation, treating salt, sulfur, and carbon dioxide as separate problems, or they rely on computer models that are so complex and opaque that no one can explain why they reached a specific conclusion.

A team of researchers at Amity University in India has proposed a new way to solve this puzzle. They are not reporting new laboratory experiments or a single breakthrough discovery, but rather a comprehensive blueprint for a smarter, more transparent system. They call this a "digital twin," which is essentially a virtual, living copy of a real concrete structure that updates itself in real time. This new framework is designed to listen to sensors embedded directly into the concrete, combine that live data with the known laws of chemistry and physics, and use artificial intelligence to predict how long the structure will last. Crucially, unlike many modern AI systems that act as "black boxes" where the reasoning is hidden, this system is built to explain its own logic, telling engineers exactly which factors are driving the prediction.

The core of their proposal is a five-layer framework that mimics how a human expert would think, but with the speed and precision of a computer. The first layer involves the physical sensors. Imagine tiny devices buried inside the concrete that act like a nervous system, constantly measuring the moisture, the electrical resistance, and the chemical levels of chloride and sulfate ions. These sensors send their data to a central hub. The second layer is the "mechanistic core," a set of mathematical rules that describe how these chemicals move and react within the concrete. This part of the system is based on established physics, ensuring that the computer never predicts something that violates the laws of nature, such as chemicals moving faster than they physically can.

The third layer introduces the artificial intelligence, but with a twist. Instead of letting the AI guess based on patterns alone, the researchers force it to work alongside the physics rules. The AI looks at the data from the sensors and the predictions from the physics model, and it learns to correct any small errors or gaps. If the physics model says the salt should be at a certain depth, but the sensor says it is deeper, the AI adjusts the prediction. This creates a hybrid model that is both data-driven and physically grounded. The fourth layer is the "digital twin" itself, which acts as a continuous updater. As new data arrives from the field, the system recalculates its understanding of the structure's health, refining its estimate of the remaining service life. This is similar to how a weather forecast improves as new satellite data comes in, but here the "weather" is the slow, decades-long decay of a bridge or a dam.

The most significant innovation in this work is the fifth layer: explainability. In many advanced computer systems, the final answer is given without a clear reason, leaving engineers hesitant to trust it. This framework uses a specific method to break down the prediction into its component parts. It can tell an engineer, "The structure is likely to fail in 15 years, and the primary reason is the high level of sulfate in the soil, not the salt." This transparency allows human experts to verify the logic against their own experience and physical intuition. The researchers tested their idea by comparing it against six other existing digital twin and machine learning frameworks. They found that while other systems might excel at detecting structural cracks or vibrations, none of them successfully combined the chemical coupling of multiple hazards with real-time sensor data and clear, human-readable explanations.

The authors are careful to note that this is a conceptual framework, a detailed plan for how such a system should be built, rather than a finished product currently running on a bridge. They have not yet built a new physical structure to test it on. Instead, they demonstrated how their system would work using data ranges and scenarios found in existing scientific literature. They showed that their mathematical formulas could handle the complex interactions between chloride, sulfate, and carbonation, and that their algorithm could theoretically update its predictions as new information arrived. They also outlined a rigorous plan for how this system could be validated in the real world, suggesting that future work should involve testing it against long-term monitoring data from actual infrastructure, such as the Itaipu Dam in Brazil or the marine exposure sites used to calibrate current industry standards.

The researchers acknowledge that significant challenges remain before this system can be widely adopted. The sensors themselves must be able to survive for decades inside the harsh environment of concrete without failing or creating new pathways for chemicals to enter. The computational cost of running these complex, coupled models in real time is high, and the system requires a steady stream of high-quality data that is currently difficult to obtain. Furthermore, while the physics of chloride and carbonation are well understood, the behavior of sulfate attack and other chemical reactions is still not fully mapped out, meaning the "mechanistic core" of the digital twin will need to evolve as science progresses.

Despite these hurdles, the paper offers a clear path forward for the future of infrastructure management. By merging the reliability of physical laws with the adaptability of machine learning, and by insisting that the computer must explain its reasoning, the researchers propose a system that could fundamentally change how we maintain the built environment. This approach moves beyond simple guesswork or isolated models, offering a way to see the hidden chemical battles happening inside our concrete and to predict their outcome with a clarity that has been missing until now. The ultimate goal is not just to predict failure, but to provide engineers with the specific, understandable information they need to make better decisions about repair and maintenance, ensuring that the structures we rely on remain safe for generations to come.

Technical Summary: A Sensor-Integrated, Explainable Machine-Learning Digital-Twin Framework for Field-Validated Service-Life Prediction of Cementitious Materials

1. Problem Statement

Reinforced cementitious infrastructure is subjected to multiple, simultaneous degradation mechanisms—specifically chloride ingress, sulfate attack, carbonation, and freeze-thaw or alkali-silica reaction (ASR) cycling. However, current industry-standard service-life models (e.g., Life-365, DuraCrete, fib Model Code) largely treat these mechanisms in isolation or rely on single-mechanism deterministic laws (e.g., Fick's second law). These models often lack the capacity to handle coupled multi-hazard interactions, particularly under unsaturated or cyclic wet-dry conditions where reaction kinetics differ significantly from saturated laboratory tests.

Concurrently, while Machine Learning (ML) and Deep Learning (DL) studies have demonstrated high predictive accuracy (up to 99%) on curated laboratory datasets, they suffer from critical limitations:

Lack of Field Validation: Most models are trained on lab-generated data with limited or no validation against long-term field monitoring.
Poor Interpretability: Models often function as "black boxes," lacking the explainability required for engineering decision-making.
Absence of Mechanistic Constraints: Purely data-driven approaches do not incorporate governing physical laws, leading to poor generalization in real-world scenarios.
Static Nature: Existing frameworks rarely accommodate real-time updates from embedded sensors to refine predictions as new data arrives.

The identified research gap is the absence of an explainable, sensor-integrated digital-twin framework that couples multi-hazard mechanistic degradation models with field-updatable machine learning for cementitious materials.

2. Methodology

The authors propose a five-layer conceptual framework designed to merge embedded sensing, physics-based modeling, and explainable AI within a synchronized digital twin architecture.

Layer 1: Field Sensing
- Utilizes embedded electrochemical sensors (half-cell potential, linear polarization resistance, resistivity) and optical sensors (Fiber Bragg Grating for strain/temperature, pH/chloride optodes).
- Includes periodic ex-situ sampling (core extraction) for microstructural ground-truthing (SEM, XRD, MIP).
Layer 2: Coupled Mechanistic Core
- Serves as the physically constrained backbone, modeling multi-ion reactive transport.
- Governing equations include:
  - Chloride Transport: Fick's second law with binding isotherms (Tang and Nilsson) and Nernst-Planck migration terms.
  - Sulfate Reaction-Diffusion: Kinetics of ettringite formation coupled with moisture-dependent reaction rates ( $k_r(\theta_w)$ ) to reconcile conflicting saturated/unsaturated kinetics.
  - Carbonation: A square-root-of-time propagation model derived from CO₂ diffusion-reaction theory.
- These mechanisms are coupled via common state variables: moisture content, temperature, and competitive ion binding. The framework develops the mathematical formulations for each layer, synthesizing existing coupled Fickian/Nernst-Planck diffusion and sulfate-driven ettringite formation kinetics previously shown for pairwise coupling.
Layer 3: Explainable Machine-Learning Layer
- A hybrid ML model (e.g., Gradient-Boosted Trees or Physics-Informed Neural Networks - PINNs) trained on the combined output of the mechanistic core and sensor streams.
- Physics-Informed Loss: The ML model minimizes a composite loss function ( $L = L_{data} + \lambda_{phys} L_{phys}$ ), where $L_{phys}$ enforces consistency with the governing transport equations (e.g., Equation 8).
- Explainability: Uses SHAP (SHapley Additive exPlanations) values to decompose predictions into contributions from individual exposure and mix-design variables, ensuring the output is interpretable by engineers.
Layer 4: Digital-Twin Synchronisation
- Employs a Bayesian data assimilation scheme (Kalman filter or particle filter) to update the digital twin's internal state.
- Distinguishes between aleatoric uncertainty (measurement/material variability) and epistemic uncertainty (model-form uncertainty).
- The state vector is updated by reconciling the mechanistic-ML prior with new sensor observations ( $y_k$ ), allowing the model to "learn" from field data and reduce uncertainty over time.
Layer 5: Decision-Support Output
- Generates a probabilistic Remaining Service Life (RSL) estimate.
- Provides ranked, interpretable maintenance triggers and identifies the top drivers (via SHAP values) influencing the prediction.

Computational Workflow: The framework operates in a synchronization cycle (e.g., daily/weekly):

Acquisition: Raw sensor data is pre-processed.
Mechanistic Forward Pass: The coupled core predicts the state prior.
ML Correction: The ML layer applies a bias correction based on the prior and sensor data.
Bayesian Assimilation: The state is updated using the Kalman filter equation.
RSL Projection: The posterior state is propagated forward until a critical threshold (e.g., critical chloride content) is reached.
Reporting: Results are visualized for asset managers.

3. Key Contributions

Novel Integration Architecture: The paper presents a conceptual framework that integrates a mechanistic multi-hazard reactive-transport core, field-updatable explainable ML, and Bayesian digital-twin updating within a single synchronized architecture. This integration distinguishes the proposed approach from existing frameworks by simultaneously addressing constraints between the mechanistic model and ML correction terms in terms of dataflow and update logic, rather than claiming the absolute novelty of the individual components.
Mathematical Formulation: The work develops the governing mathematical formulations for each layer, synthesizing existing coupled Fickian/Nernst-Planck diffusion and sulfate-driven ettringite formation kinetics (previously shown for pairwise coupling) and explicitly introducing moisture-dependent reaction rates to address conflicting literature findings on sulfate attack kinetics.
Explainability by Design: Unlike previous "black-box" durability models, this framework mandates SHAP-based attribution as a core output, addressing the interpretability gap identified in recent RILEM reviews.
Uncertainty Quantification: It formally treats uncertainty propagation, separating aleatoric and epistemic components and demonstrating how Bayesian updates reduce epistemic uncertainty as field observations accumulate.
Validation Strategy: The authors outline a staged validation approach using existing, named datasets (e.g., Itaipu Dam core records, FHWA LTBP program, Life-365 field calibration data) rather than proposing new experimental campaigns.

4. Results and Demonstration

The paper does not report new laboratory experimental results. Instead, it provides:

Illustrative Calculations: Figure 2 demonstrates the mechanistic core's output using literature-derived parameter ranges for chloride ingress, showing the model's ability to simulate diffusion profiles.
Qualitative Capability Scoring: A rubric-based scoring system (Figure 3, Table 3a) compares the proposed framework against six existing digital-twin and ML-durability studies. The proposed framework scores highest across all five dimensions: mechanistic coupling, explainable AI, Bayesian updating, embedded field sensing, and multi-hazard capability.
Synthesis of Gaps: The analysis confirms that while individual components (mechanistic models, ML, sensors) exist in literature, their intersection in a single, field-validatable framework is currently empty.

5. Significance and Claims

The authors claim that this work addresses a specific, literature-backed research gap by providing a conceptual and method-oriented framework rather than a new experimental dataset. Its significance lies in:

Bridging the Gap: It connects the "mechanistic-only" and "ML-only" approaches, creating a hybrid system that leverages the physical consistency of mechanistic models and the adaptability of ML.
Field-Readiness: By incorporating Bayesian updating and handling sensor drift/uncertainty, the framework is designed for long-term, real-world deployment where data is noisy and incomplete.
Engineering Utility: The focus on SHAP-based explainability ensures that the output is actionable for asset-management engineers, not just data scientists.
Scalability: The framework is designed to be modular, allowing for the integration of different sensor types, ML architectures, and mechanistic cores without altering the core synchronization logic.

The paper concludes that genuinely field-validated, interpretable, multi-hazard service-life prediction requires this transparent, mathematically rigorous development. Future work is identified as building standardized datasets, testing the physics-informed ML layer with real multi-mechanism data, and conducting pilot field deployments to validate the Bayesian updating scheme and practical implementation considerations (e.g., sensor drift, BIM integration).