💻 computer science

An Explainable Android-Based Decision Support System for High-Performance Concrete Strength Prediction Using XGBoost and Monte Carlo SHAP

This study presents an explainable Android-based decision support system that utilizes an optimized XGBoost model and a computationally efficient Monte Carlo SHAP algorithm to provide real-time, interpretable predictions of high-performance concrete compressive strength directly on mobile devices.

Original authors: mohsen beigi

Published 2026-08-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: mohsen beigi

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 skeleton of our cities, bridges, and homes. While ordinary concrete is a mixture of cement, water, sand, and stone, high-performance concrete is a more sophisticated version engineered to be stronger, more durable, and more workable. Creating this material requires a delicate balance; engineers must mix specific amounts of ingredients like cement, water, and various additives to achieve a target strength. Traditionally, determining the right mix involves time-consuming laboratory tests where samples are cured and crushed to see how much force they can withstand. This process is slow, expensive, and often leaves engineers guessing about the final result until after the concrete has already been poured.

In recent years, scientists have turned to computer programs that learn from past data to predict these outcomes. These programs, known as machine learning models, can analyze thousands of previous concrete recipes and their resulting strengths to find complex patterns that human intuition might miss. However, a significant problem has emerged: the most accurate computer models often act like "black boxes." They can predict the strength of a new mixture with great precision, but they cannot explain why they arrived at that number. For an engineer standing on a construction site, knowing the predicted strength is useful, but understanding which ingredient is pushing the strength up or pulling it down is essential for making safe, informed decisions. Furthermore, these powerful models usually require massive computers or internet connections to run, making them impractical for use in the field where power and connectivity are often scarce.

A researcher has now bridged this gap by creating a tool that brings both high-accuracy prediction and clear explanation directly to a smartphone. Their work focuses on a specific type of advanced computer model called XGBoost, which they found to be the most effective at guessing the strength of high-performance concrete. The researcher tested this model against five other popular computer learning methods using a database of over two thousand different concrete mixtures. The results showed that the XGBoost model was the most accurate, with its predictions deviating from the actual measured strength by an average of only 3.512 megapascals, a unit of pressure. This performance was superior to the other models, which had larger margins of error.

The true innovation of this study, however, lies in how they made this powerful model work on a mobile device and made it understandable. Running a complex model on a phone is difficult, but the researcher solved this by translating the trained model into a simple text file format that the phone could read and process without needing any external internet connection. This allowed the application to run entirely offline, meaning an engineer could use it anywhere, even in remote locations. To solve the "black box" problem, the researcher developed a new way to explain the model's thinking that is fast enough to run on a smartphone. Standard methods for explaining these models are so computationally heavy that they would take too long and drain a phone's battery. The researcher instead used a technique that samples random variations of the input data to estimate how much each ingredient contributed to the final prediction.

The result is a decision-support system that does more than just give a number. When a user enters the details of a concrete mixture—such as the amount of cement, water, and the age of the mix—the app instantly calculates the predicted strength. Simultaneously, it generates a visual chart that breaks down the prediction. This chart shows a starting point based on an average mixture and then adds or subtracts values for each specific ingredient the user entered. For example, it might show that a higher amount of cement pushed the predicted strength up, while a specific type of additive pulled it down. This visual explanation allows the user to see exactly how the computer reached its conclusion, turning a complex calculation into a transparent story about the mixture's composition.

The researcher demonstrated that this approach works effectively by testing it on their dataset and confirming that the explanations matched the model's behavior. They found that the simplified method for generating these explanations was hundreds of times faster than the traditional, exact method, making it feasible to use in real-time on a mobile device. By combining a highly accurate prediction engine with a clear, visual explanation that runs offline, this system offers a practical tool for engineers. It transforms the process of designing concrete from a trial-and-error exercise into a data-driven decision, allowing professionals to optimize their mixtures and ensure structural safety without needing a laboratory or a supercomputer.

Technical Summary: An Explainable Android-Based Decision Support System for High-Performance Concrete Strength Prediction

Problem Statement
Accurate prediction of the compressive strength (CS) of High-Performance Concrete (HPC) is critical for optimizing mixture designs and reducing experimental costs. However, HPC strength depends on complex, nonlinear interactions between eight input variables (Cement, Blast Furnace Slag, Fly Ash, Water, Superplasticizer, Coarse Aggregate, Fine Aggregate, and Age), rendering traditional empirical methods insufficient. While Machine Learning (ML) models offer high predictive accuracy, they often function as "black boxes," lacking the interpretability required for engineering decision-making. Furthermore, existing Explainable AI (XAI) techniques, such as SHapley Additive exPlanations (SHAP), are computationally prohibitive for real-time deployment on resource-constrained mobile devices, typically requiring cloud servers or desktop environments. This study addresses the gap in deploying accurate, interpretable, and offline-capable ML systems directly on smartphones for field applications in civil engineering.

Methodology
The study followed a structured workflow involving data analysis, model development, optimization, and mobile deployment:

Dataset and Preprocessing: A dataset comprising 2,171 HPC mixtures with eight input variables and CS as the target was utilized. The data was split into 75% training and 25% testing sets, with 5-fold cross-validation employed to ensure robust performance assessment. Hierarchical clustering and correlation heatmaps were used to analyze feature dependencies.
Model Development and Comparison: Six ML algorithms were implemented and optimized using the Optuna framework: AdaBoost, Random Forest (RF), Support Vector Regression (SVR), Artificial Neural Network (ANN), CatBoost, and XGBoost. Hyperparameter tuning was conducted over 30 trials per model.
Model Selection and Export: XGBoost demonstrated superior predictive performance. The trained model was exported into a JSON format containing the base score and decision tree structures, enabling reconstruction without external ML libraries.
Mobile Deployment: An Android application was developed using the Basic4Android (B4A) platform. It integrates three modules: data input, a lightweight XGBoost inference engine (reconstructing trees from JSON for offline prediction), and an explainability module.
Monte Carlo SHAP Approximation: To overcome the computational intractability of exact SHAP calculations on mobile devices (which would require over 4.4 million model evaluations for the dataset), a Monte Carlo approximation was implemented. This method estimates Shapley values by sampling random feature permutations (e.g., $N=1000$ ) and calculating sequential marginal contributions. This reduces computational cost by approximately 494-fold while preserving interpretability.
Visualization: The application visualizes SHAP values via an interactive waterfall plot, showing the baseline prediction and the positive (green) or negative (red) contribution of each input variable to the final CS prediction.

Key Results

Predictive Performance: Among the six models, XGBoost achieved the highest accuracy with an $R^2$ of 0.969, a correlation coefficient ( $r$ ) of 0.985, and the lowest error metrics (RMSE = 3.512 MPa, MAE = 2.297 MPa). CatBoost performed similarly, while SVR and AdaBoost showed lower performance.
Stability: XGBoost maintained consistent performance across all 5-fold cross-validation splits ( $R^2$ ranging from 0.962 to 0.984).
Mobile Implementation: The JSON-based XGBoost engine successfully reproduced Python-level predictions on Android devices without internet connectivity.
Explainability Efficiency: The Monte Carlo SHAP approach enabled near real-time generation of feature attributions on smartphones. The reconstructed prediction (baseline + sum of SHAP values) closely matched the model's direct output, validating the approximation's consistency.
Feature Insights: The waterfall plots revealed that Cement (C) and Age were primary drivers of strength, while the system effectively visualized how specific mixture parameters (e.g., Water-to-binder ratio components) influenced individual predictions.

Key Contributions

Comparative Analysis: Systematic evaluation of six ML algorithms for HPC strength prediction, identifying XGBoost as the optimal model.
Offline Mobile Deployment: Successful integration of a high-performance XGBoost model into a standalone Android application using B4A, eliminating the need for cloud infrastructure.
Lightweight XAI: Development and implementation of a Monte Carlo SHAP algorithm specifically designed for resource-constrained mobile environments, making real-time model interpretability feasible on-device.
Integrated Framework: Creation of a unified decision-support tool that combines prediction, portability, and explainability, allowing engineers to understand why a specific strength is predicted.
Field Applicability: Demonstration of a practical tool for on-site engineering decisions where internet access is limited or unavailable.

Significance and Claims
The paper claims that the developed framework successfully bridges the gap between advanced ML capabilities and practical engineering needs. By integrating prediction and explainability into a single mobile platform, the system moves beyond "black-box" predictions to provide transparent, data-driven insights directly at construction sites. The author asserts that this approach supports reliable decision-making by allowing practitioners to visualize the specific influence of mixture parameters on strength. The study concludes that deploying explainable AI directly on mobile devices is feasible and offers a valuable tool for preliminary mixture evaluation, education, and field support, with potential for future extension to other concrete properties such as tensile strength and durability.

More like this