⚡ electrical engineering

An Interpretable and Comparative Uncertainty-Aware Machine Learning Approach for UHPC Mix Design Optimization

This study proposes an interpretable and uncertainty-aware machine learning framework that evaluates seven models for predicting Ultra-High-Performance Concrete (UHPC) compressive strength, identifies ensemble methods as superior through rigorous accuracy and confidence interval analysis, and delivers a practical GUI tool to optimize mix designs.

Original authors: Biswajit Pal

Published 2026-08-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Biswajit Pal

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 a master chef trying to create the ultimate, indestructible cake. But this isn't just any cake; it's a "super-cake" so strong it could hold up a skyscraper. In the real world, this super-cake is called Ultra-High-Performance Concrete (UHPC). It's a special kind of building material that is incredibly tough, lasts forever, and can support massive weights. However, making it is a nightmare for engineers. It's like trying to bake a cake where the recipe changes every time you open the oven door. You have to mix cement, sand, tiny powders, fibers, and chemicals in exact amounts. If you get it wrong, the cake collapses. Traditionally, engineers had to guess and check, baking hundreds of trial batches, waiting for them to dry, and crushing them to see if they were strong enough. It was slow, expensive, and frustrating.

To solve this, scientists have started using "Machine Learning" (ML). Think of ML as a super-smart digital apprentice that has read every recipe book ever written. Instead of baking a thousand cakes to learn, the apprentice reads the data from thousands of past experiments and learns the secret patterns of what makes a cake strong. But here's the catch: sometimes these digital apprentices are "black boxes." They give you an answer, but they can't explain why they think it's right, and they might be confidently wrong. This paper is about teaching these digital apprentices to not only guess the strength of the concrete but to admit how sure they are, explain their reasoning, and actually help a real engineer in the field.

The Digital Apprentice's Big Test

In this study, a researcher named Biswajit Pal decided to put seven different types of digital apprentices to the test to see which one could best predict the strength of UHPC. Imagine these seven models as seven different students taking a massive math exam. Some students are simple and stick to straight lines (Linear Regression), while others are complex, using trees and forests to find patterns (Random Forest, Gradient Boosting, Decision Trees). There are also students who learn by looking at their neighbors (k-Nearest Neighbors) and others who use complex neural networks (Backpropagation Neural Networks).

The researcher fed all seven students the same massive dataset: 820 real-world examples of concrete mixes. Each example had 13 ingredients listed, like how much cement, water, silica fume, or fiber was used, along with the temperature and how long the concrete had been curing. The goal was simple: look at the ingredients and guess the final strength.

The Results: Who Passed the Exam?

The results were clear. The simple students who tried to draw straight lines (Linear Regression) struggled. They couldn't understand that the relationship between ingredients and strength is messy and curved, not straight. They made big mistakes, sometimes guessing the strength was off by more than 20%.

The winners were the "ensemble" students—specifically Random Forest and Gradient Boosting. These models work like a committee of experts. Instead of one person making a decision, they combine the opinions of many "decision trees." The Random Forest model was the star of the show. It predicted the strength with incredible accuracy, getting it right about 93% of the time (an R² value of 0.93). Its guesses were so close to the real values that the error was less than 10%, and for most samples, it was within a tiny 5% margin. This is a huge deal because in construction, being off by even a little bit can be dangerous or wasteful.

The "Uncertainty" Safety Net

Here is where the paper gets really clever. Usually, people just look at who got the highest score. But this researcher asked, "What if a model is confident but wrong?" To fix this, they added an "Uncertainty Analysis." Imagine a weather forecaster who says, "It will rain tomorrow," versus one who says, "It will rain tomorrow, and I'm 95% sure it will happen between 2 PM and 4 PM."

The study measured how "wide" the safety net was for each model. They found that the winning models (Random Forest and Gradient Boosting) had the narrowest safety nets. This means they weren't just accurate; they were reliably accurate. The simple models had wide, shaky safety nets, meaning their predictions could swing wildly. This uncertainty check proved that the complex, committee-style models were the only ones trustworthy enough for real-world use.

Cracking the Black Box

One of the biggest complaints about AI is that it's a "black box"—you put data in, and a number comes out, but you don't know why. To fix this, the researcher used a tool called SHAP (Shapley Additive exPlanations). Think of SHAP as a magnifying glass that lets you see exactly which ingredient contributed the most to the strength.

The analysis revealed some fascinating truths. The most important factor wasn't just the amount of cement; it was Age. The longer the concrete sits and cures, the stronger it gets. This makes sense, as the chemical reactions take time. Other big winners were Silica Fume (a super-fine powder), Fiber content, and Cement. Interestingly, the tool showed that if you add too much aggregate (the gravel) or too much superplasticizer, the strength actually goes down. The AI learned that there is a "sweet spot" for every ingredient, and adding more isn't always better.

From Lab to Construction Site

Finally, the researcher didn't just leave these findings in a computer. They built a Graphical User Interface (GUI)—basically, a friendly software tool that engineers can use on their computers. Imagine an engineer standing on a construction site. Instead of calling a lab and waiting days for a test, they can open this software, type in their mix recipe (e.g., "I have 500kg of cement and 20kg of fiber"), and instantly get a prediction of how strong the concrete will be. They can easily adjust the input numbers to see what happens if they change the amount of water or the curing temperature. The software even shows them the "confidence score" and explains which ingredients are doing the heavy lifting.

The Bottom Line

This paper suggests that by using smart, committee-style machine learning models (like Random Forest) and checking their confidence levels, we can stop guessing and start designing Ultra-High-Performance Concrete with precision. It proves that these digital tools aren't just fancy calculators; they are reliable partners that can explain their logic and help engineers save time, money, and materials while building stronger, more sustainable structures. The study concludes that while simple models are too shaky for this job, the advanced, interpretable AI models are ready to be the new standard for designing the concrete of the future.

Technical Summary: An Interpretable and Comparative Uncertainty-Aware Machine Learning Approach for UHPC Mix Design Optimization

Problem Statement
Ultra-High-Performance Concrete (UHPC) offers superior mechanical and durability properties but presents a complex mix design challenge due to the nonlinear interactions among numerous constituents (e.g., binders, fibers, admixtures, and curing conditions). Traditional trial-and-error experimental approaches are time-consuming, labor-intensive, and economically inefficient. While Machine Learning (ML) has emerged as a promising tool for predicting concrete properties, existing literature often suffers from three critical gaps: (1) a lack of unified comparative studies evaluating both linear and nonlinear ML models on a single dataset; (2) insufficient uncertainty analysis, where models are selected based solely on accuracy metrics without assessing prediction reliability; and (3) limited practical usability, as many studies fail to provide tools for onsite engineers to implement these models. Furthermore, the "black-box" nature of advanced ML models hinders their acceptance in engineering practice due to a lack of interpretability regarding how specific mix constituents influence compressive strength.

Methodology
This study proposes a comprehensive framework to address these gaps using a dataset of 820 UHPC compressive strength samples compiled from peer-reviewed literature. The methodology involves four key stages:

Data Preparation: Thirteen input variables were selected, including cement content, supplementary cementitious materials (slag, silica fume, limestone powder, quartz powder, fly ash, nano-silica), aggregate content, water content, fiber content, superplasticizer dosage, curing temperature, and age. The dataset was split into 80% for training and 20% for testing.
Model Development and Comparison: Seven distinct ML algorithms were systematically implemented and evaluated using identical datasets and performance metrics:
- Linear Regression (LR)
- Support Vector Regression (SVR)
- Backpropagation Neural Network (BPNN)
- k-Nearest Neighbors (KNN)
- Decision Tree (DT)
- Random Forest (RF)
- Gradient Boosting (GB)
Performance Evaluation: Models were assessed using standard error metrics (MSE, RMSE, MAE, MAPE) and the coefficient of determination ( $R^2$ ). Crucially, the study incorporated uncertainty analysis by calculating the 95% confidence interval (based on the mean and standard deviation of normalized prediction errors) to rank models by reliability rather than accuracy alone.
Interpretability and Application: The optimal model was interpreted using Shapley Additive exPlanations (SHAP) to quantify global feature importance and local interaction effects. Finally, a Python-based Graphical User Interface (GUI) was developed to allow engineers to input mix parameters and predict compressive strength interactively.

Key Results

Model Performance: Ensemble-based models (Random Forest, Gradient Boosting, and Decision Tree) consistently outperformed linear and single-learner models. The Random Forest (RF) model achieved the highest predictive accuracy with an $R^2$ of approximately 0.93 and the lowest RMSE. In contrast, Linear Regression (LR) demonstrated the poorest performance, with RMSE values exceeding 20%, confirming that UHPC behavior is too complex for linear assumptions.
Uncertainty Analysis: The uncertainty analysis revealed significant differences in model reliability that accuracy metrics alone failed to capture. Ensemble models exhibited the narrowest 95% uncertainty bands, indicating superior stability and generalization. LR showed the widest uncertainty bands, confirming its unsuitability for high-performance concrete systems.
Feature Importance (SHAP): SHAP analysis identified curing age as the most dominant factor influencing compressive strength, followed by superplasticizer dosage, fiber content, cement content, and silica fume. The analysis confirmed that these factors generally have a positive influence on strength, while excessive aggregate content or superplasticizer dosage can lead to negative impacts due to porosity or loss of cohesion. The RF and GB models provided smoother, more physically interpretable sensitivity patterns compared to the more erratic responses of the single Decision Tree.
Practical Tool: The developed GUI successfully integrated the best-performing models, allowing users to visualize the impact of input parameters on the 28-day compressive strength and view model performance metrics in real-time.

Significance and Claims
The paper claims to provide a robust, data-driven framework that bridges the gap between theoretical ML prediction and practical engineering application. Its primary significance lies in:

Unified Comparison: Establishing a fair, side-by-side evaluation of diverse ML algorithms, demonstrating that ensemble methods are superior for UHPC mix design.
Uncertainty-Aware Selection: Introducing uncertainty quantification as a critical criterion for model selection, ensuring that chosen models are not only accurate but also reliable and robust for engineering decision-making.
Interpretability: Utilizing SHAP to demystify the "black box" of ML, thereby increasing engineer confidence by explicitly linking input mix constituents to strength outcomes.
Accessibility: Providing a ready-to-use GUI tool that reduces reliance on extensive experimental trials, potentially lowering costs and accelerating the adoption of sustainable, optimized UHPC mix designs.

The author concludes that while the current framework is effective for compressive strength prediction, future work could extend to durability properties and the integration of physics-based relationships to further enhance model generalization.

More like this