⚡ electrical engineering

Understanding the Governing Parameters of UHPC’s Strength Using Explainable Machine Learning

This study demonstrates that an XGBoost machine learning model, trained on 810 experimental mixtures and interpreted via SHAP analysis, achieves high predictive accuracy for Ultra-High-Performance Concrete (UHPC) compressive strength while identifying fiber content and curing age as the most influential governing parameters.

Original authors: Nafis Niaz Chowdhury, Syeda Tabassum, Badhon Singha

Published 2026-06-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nafis Niaz Chowdhury, Syeda Tabassum, Badhon Singha

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 bake the perfect, super-strong loaf of bread. In the world of construction, this "bread" is called Ultra-High-Performance Concrete (UHPC). It's not your average sidewalk concrete; it's a high-tech material strong enough to build skyscrapers and long-span bridges.

The problem is that making this "super-bread" is incredibly tricky. You can't just follow a simple recipe like "add 2 cups of flour." The ingredients (cement, sand, fibers, water, etc.) interact in complex, non-linear ways. If you change one tiny thing, the strength might skyrocket or crash. Traditionally, engineers had to guess and test physically, which is slow, expensive, and destructive.

This paper is like a team of data detectives who decided to teach a computer how to understand this complex recipe using Machine Learning, but with a special twist: they made the computer explain why it made its predictions, so engineers could trust it.

Here is the story of how they did it, broken down into simple steps:

1. The Data Collection: Gathering the Recipe Book

The researchers didn't invent new concrete; they went on a scavenger hunt through scientific literature. They found 810 different recipes (experimental mixtures) that had already been tested in labs.

The Ingredients: They looked at 13 different variables, like how much cement, silica fume (a fine powder), steel fibers, water, and how long the concrete was cured (aged).
The Goal: They wanted to predict the final "strength" of the concrete based on these ingredients.

2. The Contest: Four Digital Chefs

To find the best way to predict the strength, they trained four different "digital chefs" (Machine Learning models) to learn from the 810 recipes:

SVR (Support Vector Regression): A method that tries to draw a smooth line through the data.
ANN (Artificial Neural Network): A model that mimics how human brains connect neurons.
RF (Random Forest): A team of decision trees that vote on the answer.
XGBoost: A powerful "boosting" algorithm that learns from its mistakes sequentially, getting smarter with every step.

The Result:
Think of this as a cooking competition.

ANN was a bit confused and made the most mistakes.
SVR did okay but wasn't perfect.
Random Forest was very good.
XGBoost was the champion. It predicted the strength with incredible accuracy (getting it right 96.4% of the time). It was the most reliable "chef" in the kitchen.

3. The "Black Box" Problem: Why Trust the Computer?

Usually, powerful AI models are like black boxes. You put ingredients in, and a number comes out, but you have no idea why the computer gave that number. Engineers can't use a tool they don't understand.

To fix this, the researchers used a tool called SHAP (SHapley Additive exPlanations).

The Analogy: Imagine the computer predicts a strength of 153 MPa. SHAP acts like a referee that breaks down the score. It says: "Okay, the starting score was 126. The steel fibers added +14.8 points. The curing time added +12.2 points. But the water subtracted a little bit."
This turned the "black box" into a transparent glass box, showing exactly which ingredients were doing the heavy lifting.

4. The Big Discoveries: What Actually Makes it Strong?

Using their champion model (XGBoost) and the explanation tool (SHAP), they found out what really matters:

The Time Factor (Curing Age): Just like bread needs time to rise, UHPC needs time to get strong. The model showed that as the concrete gets older, it gets stronger, but the biggest gains happen in the first 28 days.
The Reinforcement (Fiber Content): Adding steel fibers was the single most important factor for boosting strength. It's like adding steel rebar to a cake; it holds everything together and prevents cracks.
The Secret Powder (Silica Fume): While the computer's "explanation" tool said fibers and time were the top bosses, a simple math check showed that Silica Fume had the strongest direct link to strength. It acts like a super-filler, making the concrete dense and tight.

5. The Conclusion: A New Way to Build

The paper concludes that we don't need to guess anymore. By using this "explainable" AI:

We can predict the strength of UHPC with high accuracy (the XGBoost model).
We can understand why it's strong (thanks to SHAP), confirming that physics (like hydration and fiber bridging) is actually happening inside the model.

In short: The researchers built a super-smart, transparent computer program that learned from 810 past experiments. It can now tell engineers exactly how strong their concrete will be and which ingredients are the most important, saving time and money while ensuring safety.

Technical Summary: Understanding the Governing Parameters of UHPC's Strength Using Explainable Machine Learning

Problem Statement
Ultra-High-Performance Concrete (UHPC) exhibits highly nonlinear compressive strength behavior driven by complex interactions among mixture constituents and curing conditions. Conventional empirical models and linear regression approaches often fail to capture these higher-order feature interactions, multiplicative coupling effects, and threshold-sensitive nonlinearities intrinsic to UHPC hydration and microstructural systems. Furthermore, standard design codes (e.g., ACI 318-19, Eurocode 2) lack explicit constitutive provisions for UHPC, relying instead on simplified relationships calibrated for normal-strength concrete. While machine learning (ML) offers a paradigm for approximating nonlinear input-output mappings without predefined constitutive constraints, high-capacity models often suffer from the "black-box" problem, limiting their acceptance in structural practice due to a lack of physical interpretability.

Methodology
This study employs a data-driven framework integrating predictive modeling with Explainable Artificial Intelligence (XAI) to analyze a curated database of 810 experimental UHPC mixtures compiled from peer-reviewed literature.

Data Curation: The dataset includes 13 input variables (e.g., cement, silica fume, fiber content, curing age, temperature) and one output variable (compressive strength). Data underwent rigorous preprocessing to remove duplicates and physically inconsistent values. The dataset was stratified into early-age (≤28 days) and later-age (>28 days) regimes to account for distinct strength development mechanisms.
Model Development: Four machine learning algorithms were trained and evaluated: Support Vector Regression (SVR), Artificial Neural Network (ANN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost).
Optimization: Systematic hyperparameter optimization was conducted using GridSearchCV and RandomizedSearchCV with five-fold cross-validation to ensure robustness and prevent overfitting.
Evaluation Metrics: Model performance was assessed using the coefficient of determination ( $R^2$ ), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE).
Explainability: The best-performing model was subjected to SHapley Additive exPlanations (SHAP) analysis. This approach decomposes predictions into additive feature contributions, allowing for both global feature importance ranking and local sample-level attribution based on axiomatic cooperative game theory.

Key Results

Predictive Performance: Among the evaluated models, XGBoost achieved the highest predictive accuracy, yielding a test $R^2$ of 0.9641, an RMSE of 7.3237 MPa, and an MAE of 5.0748 MPa. Five-fold cross-validation confirmed strong generalization capability with a mean $R^2$ of 0.970 ± 0.004. In comparison, Random Forest performed well ( $R^2$ = 0.9502), while SVR and ANN showed lower performance ( $R^2$ = 0.9025 and 0.8678, respectively).
Correlation Analysis: Pearson correlation analysis identified silica fume ( $r=0.521$ ), fiber content ( $r=0.440$ ), curing age ( $r=0.415$ ), and superplasticizer dosage ( $r=0.396$ ) as having the strongest positive linear correlations with compressive strength.
Explainability (SHAP): SHAP analysis revealed that while silica fume had the highest linear correlation, fiber content and curing age were the most influential variables in the nonlinear multivariate model.
- Global Importance: Curing age and fiber content ranked highest in SHAP importance values.
- Local Interpretation: For a representative prediction of 153.65 MPa, fiber content contributed +14.82 MPa and curing age contributed +12.26 MPa to the baseline.
Trend Validation: Partial Dependence Plots (PDPs) and sensitivity analyses confirmed physically realistic trends, such as rapid strength gain at early ages followed by a plateau, and a positive but diminishing marginal benefit of fiber content at higher dosages.

Significance and Claims
The paper claims that the integration of explainable machine learning provides a dual benefit: high predictive reliability and engineering interpretability. By utilizing SHAP, the study bridges the gap between "black-box" ML predictions and physical understanding, demonstrating that the model's decisions align with established material science mechanisms, including hydration maturity, matrix densification, and fiber-reinforcement.

The authors assert that this framework offers a transparent tool for performance-based UHPC mix design and optimization. Unlike traditional empirical formulas that impose rigid functional structures, this approach autonomously identifies interaction effects and latent structural patterns. The study concludes that explainable ML can effectively identify the governing parameters controlling UHPC compressive strength, providing meaningful engineering insights into the relative influence of mixture constituents and curing conditions without the need for destructive testing at fixed ages.

1. The Data Collection: Gathering the Recipe Book

2. The Contest: Four Digital Chefs

3. The "Black Box" Problem: Why Trust the Computer?

4. The Big Discoveries: What Actually Makes it Strong?

5. The Conclusion: A New Way to Build

More like this