⚡ electrical engineering

A physics-guided graph neural network for interpretable multiscale modelling of concrete properties

The paper introduces HiCon-GNN, a physics-guided graph neural network that models the hierarchical structure of concrete to achieve competitive multi-property strength predictions while providing interpretable, physically regularized insights into multiscale material behaviors.

Original authors: Ye Li, Fangying Wang

Published 2026-07-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ye Li, Fangying Wang

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

Imagine you are trying to predict how strong a cake will be just by looking at a list of ingredients: flour, sugar, eggs, and baking powder. Most computer programs try to do this by treating the cake as a giant, flat soup of numbers. They say, "Okay, more sugar usually means sweeter, so maybe stronger?" But real life isn't a soup. A cake is a layered masterpiece. The batter (paste) holds the crumbs (sand) together, and the whole thing is baked into a structure where the edges between the crumbs and the batter are the weak spots. If you ignore those layers and just look at the ingredient list, you miss the magic of how the parts actually fit together.

This is the problem scientists face with concrete. Concrete isn't just a mix of cement and rocks; it's a hierarchy. It starts with a sticky paste, gets mixed with sand to become mortar, and finally gets big rocks added to become the final concrete. The places where the rocks touch the mortar are called "interfacial transition zones" (ITZs), and they are often the weakest links where cracks start. For a long time, computers trying to predict concrete strength have ignored this layering, treating the material as a flat, messy vector of numbers. But what if we could teach a computer to see the layers, just like a baker sees the cake? That's the big question this research tackles: Can we build a smart computer model that understands the "onion layers" of concrete to predict its strength better and explain why it's strong?

Enter HiCon-GNN, a new kind of "smart brain" for concrete designed by researchers Ye Li and Fangying Wang. Think of this model not as a simple calculator, but as a digital architect that builds a miniature, invisible city of concrete inside its computer. Instead of just reading a list of ingredients, this model constructs a "graph"—a map of dots and lines. The dots represent the different parts (paste, sand, rocks), and the lines represent how they touch each other.

Here is the cool part: this model is "physics-guided." It doesn't just guess; it follows the rules of how concrete actually works. It knows that the paste is the glue, the sand makes it stiffer, and the rocks add bulk but also create weak edges. The model has special "sub-models" for each layer. First, it figures out the strength of the paste. Then, it uses that to figure out the mortar. Finally, it looks at how the rocks sit in the mortar and calculates the strength of the whole concrete block. It's like building a tower of knowledge: you can't know how strong the top floor is until you know how strong the foundation is.

The researchers trained this digital architect on a massive library of 13,170 real-world concrete records. These records were messy and incomplete—some had data on the paste, some on the rocks, some on the curing temperature, and many were missing pieces. But HiCon-GNN is clever; it has "mask-aware" eyes that can say, "Ah, you didn't tell me the rock size? No problem, I'll learn from the other clues," instead of throwing the data away.

The results are impressive. When tested on new, unseen concrete recipes, HiCon-GNN predicted the compressive strength (how much weight it can hold before crushing) with an accuracy score ( $R^2$ ) of 0.953, which is incredibly high. It also predicted tensile strength (how hard it is to pull apart) and flexural strength (how hard it is to bend) with scores of 0.887 and 0.926, respectively. Compared to a standard computer model that ignores the layers (a "structure-blind" neural network), HiCon-GNN reduced its prediction errors by 19% for compression, 5% for tension, and 16% for bending.

But the real magic isn't just the score; it's the "X-ray vision" the model gives us. Because the model builds the layers, it can show us what it "thinks" is happening inside. It revealed that the paste is indeed the strongest part, the mortar is slightly weaker because of the sand, and the final concrete is the weakest because of the rocky edges. It even estimated the thickness of those weak edges (the ITZs), finding they are usually around 10.83 micrometers thick and act like a porous, weak zone where cracks love to start.

The model also figured out that for very high-strength concrete (like Ultra-High-Performance Concrete, or UHPC), which often has no big rocks at all, the "weak edge" penalty disappears, and the concrete strength becomes almost identical to the mortar strength. This matches real-world physics perfectly.

The researchers are careful to say that while these internal numbers (like the estimated porosity or edge thickness) make perfect physical sense and follow the rules of concrete science, they are "hypotheses" generated by the model, not direct measurements from a microscope. They are like a very smart detective's theory of what happened inside the crime scene, based on all the clues. To be 100% sure, you'd still need to go look with a microscope. However, the fact that the model's internal logic lines up so perfectly with what we know about how concrete works suggests it has truly learned the "language" of the material.

In short, this paper shows that when you teach a computer to see the layers of concrete rather than just the ingredients, it doesn't just get better at guessing numbers; it starts to understand the story of the material. It bridges the gap between raw data and physical reality, offering a tool that could help engineers design stronger, safer, and more efficient concrete without needing to mix and test thousands of batches in a lab. It's a step toward a future where we can design the perfect concrete mix on a computer, knowing exactly how the paste, sand, and rocks will dance together to hold up our bridges and buildings.

Technical Summary: A Physics-Guided Graph Neural Network for Interpretable Multiscale Modelling of Concrete Properties

Problem Statement
Concrete is a hierarchical composite material where macroscopic performance emerges from coupled chemical, microstructural, and mesoscale processes. Existing machine learning approaches for predicting concrete properties typically represent the material as a "flat" vector of mixture descriptors. While statistically effective within training distributions, these flat models are mechanically incomplete: they fail to explicitly represent the paste, mortar, interfacial transition zone (ITZ), and aggregate topology, nor do they provide access to scale-resolved material fields. Furthermore, most physics-informed approaches in this domain attach physical constraints to single outputs or scales, leaving the inherent hierarchy of concrete outside the model. There is a need for a representation that retains this hierarchy and heterogeneity while remaining compatible with data-driven learning to support interpretable, multi-property prediction.

Methodology: HiCon-GNN
The authors propose HiCon-GNN (Hierarchical Multiscale Concrete Graph Neural Network), a physics-guided architecture that mirrors the paste–mortar–ITZ–concrete hierarchy using coupled proxy graphs.

Graph Representation: The model utilizes two coupled Representative Volume Element (RVE) proxy graphs generated from mixture and particle-size descriptors:
- Mortar-scale graph: Links sand and paste regions via contact edges.
- Concrete-scale graph: Links coarse aggregate and mortar regions via aggregate–mortar, mortar–mortar, and aggregate–aggregate edges. Aggregate–mortar edges carry ITZ information.
Architecture and Flow: Information flows through four coupled submodels in the order of material assembly:
1. Paste Submodel: Maps binder chemistry, water content, and curing to paste-level properties, which populate the paste nodes of the mortar graph.
2. Mortar SubGNN: Performs message passing over the sand–paste graph to predict mortar properties, integrating paste state and sand–paste topology.
3. ITZ Submodel: Predicts interface properties (thickness, porosity, strength, modulus) for each aggregate–mortar edge based on mixture and edge geometry.
4. Concrete Mesoscale GNN: Propagates information over the concrete graph to form a global representation of packing, mortar distribution, and ITZ connectivity.
Physics-Guided Mechanisms:
- Mask-Aware Encoders: Retain incomplete literature records by distinguishing between structurally absent features (e.g., coarse aggregate in mortar) and missing data, using learned embeddings for missing values.
- Soft Physical Priors: Instead of fixed empirical coefficients, the model employs soft regularisation in the loss function. This includes Abrams-type water-to-binder (w/b) strength trends, ITZ weak-link behavior (ITZ strength $\le$ mortar strength), and cross-property consistency (tensile/flexural vs. compressive).
- Capacity-Based Heads: Strength outputs are not unconstrained regressions but are derived from a shared multiscale representation. Predicted mortar strength serves as the base capacity, modified by learned efficiency factors ( $\eta_{ITZ}$ , $\eta_{matrix}$ ) that account for interface quality and mesoscale topology.
Training: The model was trained on 13,170 records using a weighted multi-task loss that combines supervised losses for available labels (concrete, mortar, paste, ITZ) with physical regularisation terms.

Key Contributions

Hierarchical Representation: The first application of a coupled graph neural network to explicitly encode the paste–mortar–ITZ–concrete hierarchy, allowing intermediate latent states to be inspected as physically regularised hypotheses.
Multi-Property Prediction: A single coupled framework that jointly predicts compressive, tensile, and flexural strength, maintaining consistency across these modes without requiring separate models.
Interpretability via Latent States: The model outputs intermediate quantities (paste/mortar strength, porosity, ITZ thickness/strength) that recover expected physical hierarchies (e.g., paste > mortar > concrete strength; ITZ > paste > mortar porosity), providing internal consistency checks unavailable to structure-blind models.
Capacity Decomposition: The framework decomposes final strength into a base mortar capacity and multiplicative efficiency factors, isolating the specific contributions of interface weakness and mesoscale packing.

Results
Trained on a heterogeneous database of 13,170 records, HiCon-GNN demonstrated competitive performance against strong structure-blind baselines (Tabular ANN, Random Forest, Gradient Boosting, SVM, Decision Tree, and GEP):

Predictive Accuracy: On held-out test data, HiCon-GNN achieved $R^2$ values of 0.953 (compressive), 0.887 (tensile), and 0.926 (flexural).
Error Reduction: Compared to a matched tabular ANN, HiCon-GNN reduced RMSE by 19% (compressive), 5% (tensile), and 16% (flexural). While target-specific ensemble models (e.g., Random Forest) remained slightly more accurate for individual targets, HiCon-GNN maintained near-specialist accuracy across all three outputs within a single model.
Stability: Performance remained stable across five independent random data partitions, with compressive strength $R^2$ varying by only $\pm 0.004$ .
Physical Consistency:
- The model correctly recovered the strength hierarchy: Median predicted compressive strength increased from concrete (48.75 MPa) to mortar (59.24 MPa) to paste (96.97 MPa).
- Porosity followed the inverse hierarchy (paste > mortar > concrete), and ITZ porosity was consistently higher than the surrounding matrix.
- The ITZ efficiency factor ( $\eta_{ITZ}$ ) correlated strongly with interface thickness, confirming the model learned the "weak-link" behavior of the ITZ.
- Tensile and flexural predictions adhered to empirical reference trends (CEB-FIP, fib MC2010, ACI) without being forced to follow fixed equations.

Significance and Claims
The paper argues that the primary value of HiCon-GNN extends beyond predictive accuracy. By imposing the material hierarchy as an inductive bias, the model learns a representation that is consistent with concrete mechanics.

Mechanistic Inspection: The intermediate latent states serve as "physically regularised internal audits." They convert the learned representation into testable hypotheses regarding paste porosity, ITZ thickness, and bond strength, which can be validated against future experimental data.
Design Utility: The framework offers a pathway for data-driven concrete design that is not merely a "black box" predictor but a structured surrogate. It identifies that mortar strength is the dominant reference scale, while interface-controlled reduction and mesoscale topology act as limiting factors.
Limitations and Scope: The authors modestly note that the generated RVEs are topology proxies, not measured microstructures, and that the latent quantities are consistency checks rather than calibrated measurements. The model's generalization to unseen chemistries or rare regimes (e.g., recycled aggregates) remains uncertain without specific data or stronger priors. However, the framework is designed to be extensible, capable of incorporating measured microstructural graphs or coupling with hydration models in the future.

In summary, HiCon-GNN demonstrates that a hierarchical, physics-guided graph model can achieve competitive multi-property prediction while exposing a structured, interpretable pathway from binder chemistry to macroscopic strength, bridging the gap between data-driven efficiency and physical interpretability.

More like this