🤖 AI

Predicting California Bearing Ratio with Ensemble and Neural Network Models: A Case Study from Turkiye

This study demonstrates that a Random Forest model, trained on 382 soil samples from Türkiye, effectively predicts California Bearing Ratio with high accuracy (R²=0.83), offering a rapid and reliable machine learning alternative to traditional, time-consuming laboratory tests for geotechnical engineering.

Original authors: Abdullah Hulusi Kökçam, Uğur Dağdeviren, Talas Fikret Kurnaz, Alparslan Serhat Demir, Caner Erden

Published 2026-09-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Abdullah Hulusi Kökçam, Uğur Dağdeviren, Talas Fikret Kurnaz, Alparslan Serhat Demir, Caner Erden

Original paper licensed under CC BY 4.0 (http://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

Before a road can be built, engineers must know if the ground beneath it is strong enough to hold the weight of traffic. This is not a guess; it is a calculation based on how much pressure the soil can take before it gives way. For nearly a century, the standard way to measure this strength has been a test called the California Bearing Ratio. In a laboratory, a piston is pushed into a sample of soil at a steady, slow speed. The force required to push the piston in is compared to the force needed to push it into a standard block of crushed stone. The result is a single number that tells engineers how well that soil will perform as a foundation for highways or airport runways. While this method is reliable, it is also slow, expensive, and requires sending samples to a lab, which can delay projects and increase costs. As the demand for faster, more efficient construction grows, researchers have begun asking if computers can learn to predict this number just by looking at other, easier-to-measure properties of the soil.

A team of researchers in Türkiye set out to answer this question by teaching computers to recognize the hidden patterns in soil data. They gathered a large collection of information from 382 different soil samples taken from various regions across the country. For each sample, they already knew the final California Bearing Ratio from past lab tests. They also had a list of other characteristics for every sample, such as how much water the soil could hold, how fine or coarse the particles were, and how much the soil would shrink or swell. The researchers treated this information like a puzzle, feeding these known characteristics into twelve different computer programs designed to find mathematical relationships. These programs ranged from simple decision-making trees to complex networks that mimic the way human brains process information. The goal was to see if the computers could learn the rules connecting the soil's basic traits to its final strength rating without ever seeing the actual strength test results during their training.

The researchers split their data into two groups. They used most of the samples to teach the computer programs, letting them practice making predictions and correcting their own mistakes. They saved the remaining samples to act as a final exam, testing the programs on soil they had never seen before. To ensure the results were fair and not just a lucky guess, they ran the tests multiple times and adjusted the settings of each computer program to find the best possible performance. After running this rigorous comparison, one type of program stood out as the most accurate. It was an ensemble method called Random Forest, which works by combining the opinions of many smaller decision-making models to reach a single, reliable conclusion. This model proved to be the most skilled at the task, correctly predicting the strength of the soil in the final exam with a high degree of accuracy.

The results showed that the Random Forest model could predict the soil strength with remarkable precision, matching the actual lab results much better than the other eleven programs tested. When the researchers looked at the difference between what the computer predicted and what the lab actually measured, the errors were small and balanced, meaning the model did not consistently guess too high or too low. While a few extreme cases were harder to predict, the overall performance was strong enough to suggest that this approach is viable. The study did not claim that this method replaces the need for physical testing entirely, but it demonstrated that a computer could provide a very close estimate using data that is often already available. This finding offers a practical path forward for engineers, suggesting that they could use these digital tools to quickly screen soil conditions and make faster, more informed decisions during the early stages of designing roads and foundations.

Technical Summary: Predicting California Bearing Ratio with Ensemble and Neural Network Models

Problem Statement
The California Bearing Ratio (CBR) is a fundamental geotechnical indicator used to assess the load-bearing capacity of subgrade soils for transportation infrastructure and foundation design. Traditionally, CBR values are determined through laboratory penetration tests (wet or dry) involving specific piston insertion speeds and penetration depths (2.5 mm and 5 mm). While accurate, these conventional methods are characterized as time-consuming, labor-intensive, and costly, particularly when applied to large-scale projects or diverse soil profiles. Consequently, there is a need for indirect, rapid, and reliable estimation methods that utilize readily available soil index properties.

Methodology
This study proposes a data-driven framework to predict CBR values using machine learning (ML) algorithms. The research utilized a comprehensive dataset comprising 382 soil samples collected from various geoclimatic regions across Türkiye, sourced from the laboratory archives of the General Directorate of Türkiye Highways.

Input Features: The model inputs included seven soil properties: Gravel Content (GC), Sand Content (SC), Fines Content (FC), Liquid Limit (LL), Plasticity Index (PI), Maximum Dry Density (MDD), and Optimum Moisture Content (OMC).
Data Partitioning: The dataset was randomly split into a training set (80%, n=305) and a test set (20%, n=77).
Algorithm Selection: Twelve distinct regression-based models were implemented and compared:
- Ensemble Methods: Random Forest, Bagging, Extra Trees, Voting, Stacking, Gradient Boosting, AdaBoost, and XGBoost.
- Tree-Based: Decision Tree.
- Kernel-Based: Support Vector Regression (SVR).
- Instance-Based: K-Nearest Neighbors (K-NN).
- Neural Networks: Multi-Layer Perceptron (MLP).
Optimization and Validation: To ensure robustness and prevent overfitting, hyperparameter tuning was conducted using Grid Search with 5-fold cross-validation. Each model was trained and evaluated five times with different random states to account for variability, with average performance metrics reported.

Key Results
The performance of the models was evaluated using the coefficient of determination ( $R^2$ ), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE).

Top Performer: The Random Forest Regressor emerged as the most effective model. It achieved an $R^2$ score of 0.832 on the test set, with an MAE of 6.263 and an RMSE of 11.823.
Comparative Performance: While other ensemble models such as Bagging, Extra Trees, and Voting Regressors demonstrated strong performance, their results were slightly lower than Random Forest on the final test set. Notably, the Voting Regressor showed a high training $R^2$ (0.994) but a lower validation $R^2$ (0.683), indicating potential overfitting compared to the more stable Random Forest.
Model Behavior: Visual analysis of the Random Forest predictions (Actual vs. Predicted) revealed that most data points aligned closely with the ideal diagonal line. The error distribution was near-normal and centered around zero, indicating unbiased predictions, though a slight tendency for overestimation was observed in specific cases.
Underperformers: Models such as Stacking, MLP, and Decision Tree yielded lower test set $R^2$ scores (0.519, 0.748, and 0.683, respectively), suggesting they were less effective at capturing the nonlinear relationships in this specific dataset compared to the ensemble tree-based methods.

Key Contributions

Comprehensive Comparative Analysis: The study provides a systematic evaluation of twelve diverse ML algorithms, ranging from simple decision trees to complex stacking ensembles and neural networks, specifically applied to CBR prediction in the context of Turkish geoclimatic conditions.
Robust Validation Framework: By employing 5-fold cross-validation, multiple random state iterations, and a strict train/test split, the study establishes a rigorous protocol for assessing model generalization in geotechnical applications.
Identification of Optimal Architecture: The research identifies Random Forest as the superior architecture for this specific problem, highlighting its ability to map complex, nonlinear relationships between soil index properties and bearing capacity more effectively than the other tested methods.

Significance and Claims
The paper claims that integrating machine learning into geotechnical engineering offers a viable alternative to traditional laboratory testing. The primary significance lies in the potential to reduce testing time and costs while maintaining predictive accuracy. The study supports the adoption of data-centric models for indirect CBR estimation, promoting a shift toward digital transformation in infrastructure analysis and design. The authors conclude that ensemble-based learning techniques, particularly Random Forest, are robust tools for predictive geotechnical tasks, capable of handling the complexity of soil behavior without the logistical burdens of physical penetration tests.

More like this