🤖 AI

Soil Compaction Parameters Prediction Based on Automated Machine Learning Approach

This study demonstrates that an automated machine learning (AutoML) approach, specifically utilizing the XGBoost algorithm, effectively predicts soil compaction parameters (OMC and MDD) with high accuracy and generalizability across diverse soil types, offering a superior alternative to traditional laboratory methods and empirical models.

Original authors: Caner Erden, Alparslan Serhat Demir, Abdullah Hulusi Kokcam, Talas Fikret Kurnaz, Ugur Dagdeviren

Published 2026-06-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Caner Erden, Alparslan Serhat Demir, Abdullah Hulusi Kokcam, Talas Fikret Kurnaz, Ugur Dagdeviren

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

Imagine you are building a house or a road. Before you can lay the foundation, you need to pack the dirt down tight so it doesn't sink later. This process is called soil compaction.

To do this right, engineers need to know two secret recipes for the dirt:

How much water to add: Too dry, and it won't stick; too wet, and it turns to mud. This is the "Optimum Moisture Content" (OMC).
How hard to pack it: This determines the "Maximum Dry Density" (MDD), or how heavy and solid the dirt can get.

The Old Way: The Slow, Messy Kitchen

Traditionally, finding these recipes is like trying to bake the perfect cake by hand in a dark kitchen. Engineers have to go to a lab, take a soil sample, mix it with different amounts of water, pack it down with heavy hammers, and measure the results. They have to do this at least six times for every single spot to get it right. It's slow, expensive, and requires a lot of human labor.

The New Way: The Smart Chef (AutoML)

This paper introduces a new method using Artificial Intelligence (AI), specifically a tool called AutoML (Automated Machine Learning).

Think of AutoML as a super-smart, automated chef. Instead of a human chef trying to guess which spices to use or how long to bake, you give the chef a huge cookbook of past recipes (data from 126 different soil samples). The chef then:

Tries out thousands of different cooking strategies automatically.
Figures out which strategy works best without needing a human to tweak the settings.
Learns from the mistakes of previous attempts.

The Experiment: Testing the Chef

The researchers gave this "chef" a massive dataset containing information about different types of soil (like sand, clay, and gravel) and their properties (like liquid limit and plastic limit). They asked the chef to predict the two secret recipes (water amount and packing density) just by looking at the soil's basic ingredients.

They tested the chef using four different "kitchen setups" (configurations):

The Default Setup: Letting the chef use its standard settings.
The "Best Quality" Setup: Telling the chef, "Take your time, try everything, and get the best result possible."
The "Multimodal" Setup: A specific setting for handling different types of data.
The Hybrid Setup: Combining the "Best Quality" approach with the specific data handling.

The Results: Who Won the Cooking Contest?

After running the tests, the researchers found that one specific algorithm (a type of math recipe) performed the best. It was called XGBoost.

For the "Water Amount" (OMC): The XGBoost chef got it right 89.1% of the time.
For the "Packing Density" (MDD): It got it right 80.4% of the time.

These scores are much better than many previous attempts using older AI methods. The study also found that the most important ingredient the chef looked at was the Liquid Limit (LL)—basically, how much water the soil can hold before it turns into a liquid.

Why This Matters

The paper claims that this approach is a game-changer because:

It's Fast: It skips the long, manual lab tests.
It's Smart: It works well even when the soil is a mix of different types (heterogeneous), which usually confuses older models.
It's Reliable: By using the "AutoML" method, the system automatically finds the best way to solve the problem without human guesswork.

The Future Vision (According to the Paper)

The authors suggest that this isn't just a lab experiment. They envision a future where this "smart chef" is part of a smart construction site. Imagine a road roller (the machine that packs the dirt) equipped with sensors. As it rolls, it could instantly send soil data to the AI, which would then tell the machine exactly how much water to spray or how hard to pack, all in real-time. This turns a slow, manual guessing game into a fast, precise, automated process.

In short: This paper shows that by letting a computer automatically learn from past soil data, we can predict exactly how to pack dirt for construction much faster and more accurately than before.

Technical Summary: Soil Compaction Parameters Prediction Based on Automated Machine Learning Approach

Problem Statement
Soil compaction is a fundamental process in construction engineering, critical for ensuring the stability of structures such as road embankments and earth dams. The key parameters governing this process are the Optimum Moisture Content (OMC) and Maximum Dry Density (MDD). Traditionally, these values are determined through labor-intensive and time-consuming laboratory experiments, specifically Standard and Modified Proctor tests, which require extensive sample preparation and multiple iterations to generate accurate compaction curves. While empirical regression models and various Artificial Intelligence (AI) techniques (including Artificial Neural Networks, Support Vector Regression, and Genetic Programming) have been proposed as alternatives, they often face challenges regarding prediction accuracy and generalizability, particularly when applied to heterogeneous datasets representing diverse soil types. Furthermore, the manual tuning of hyperparameters in traditional machine learning (ML) workflows can limit model performance.

Methodology
This study proposes an Automated Machine Learning (AutoML) approach to predict OMC and MDD, utilizing the AutoGluon Tabular Prediction library. The methodology involves the following steps:

Data Acquisition and Preprocessing: The primary dataset comprises 115 soil samples (after removing 11 duplicates from an initial 126) sourced from Günaydın [4]. The dataset includes nine distinct soil classes (SC, GM, GC, ML, MH, MI, CL, CH, CI). Input features consist of liquid limit (LL), plastic limit (PL), gravel content (G%), sand content (S%), and fine content (F%). The target variables are OMC and MDD. To validate robustness, the study also references three external datasets representing heterogeneous, fine-grained, and coarse-grained soil types.
AutoML Framework: The study employs AutoGluon, which automates algorithm selection, hyperparameter optimization, and feature engineering. Unlike traditional methods requiring intensive manual intervention, AutoGluon utilizes presets and automated stacking to optimize model performance.
Experimental Configurations: Four distinct training configurations were defined to evaluate performance:
- Configuration 1: Default AutoGluon settings.
- Configuration 2: Utilization of the "Best Quality" preset to prioritize accuracy.
- Configuration 3: Specification of the 'multimodal' hyperparameter.
- Configuration 4: A combination of the "Best Quality" preset and the 'multimodal' hyperparameter.
Model Evaluation: Various ML algorithms (including XGBoost, LightGBM, CatBoost, Random Forest, and Neural Networks) were trained and evaluated using R-squared ( $R^2$ ) scores on test and validation sets. Feature importance was also analyzed to determine the influence of input variables.

Key Results
The study compared the performance of the four configurations across both OMC and MDD prediction tasks:

Optimal Configuration: Configuration 4 (combining "Best Quality" preset and 'multimodal' hyperparameter) yielded the highest performance for OMC prediction, while Configuration 2 ("Best Quality" preset) was highly effective for MDD.
Best Performing Algorithm: The Extreme Gradient Boosting (XGBoost) algorithm consistently demonstrated superior performance.
- OMC Prediction: The XGBoost model in Configuration 4 achieved a test $R^2$ score of 0.891.
- MDD Prediction: The XGBoost model in Configuration 2 achieved a test $R^2$ score of 0.804.
Feature Importance:
- For OMC, Liquid Limit (LL) was identified as the most significant feature (Importance: 0.473), followed by Fine Content (F%) and Sand Content (S%). Gravel Content (G%) showed negligible importance.
- For MDD, Liquid Limit (LL) was again the most influential feature (Importance: 0.356), followed by Plastic Limit (PL) and Fine Content (F%).
Generalization: The models demonstrated effectiveness across different soil types, reinforcing the value of using heterogeneous datasets to improve model generalization.

Significance and Claims
The paper claims that this research represents the first application of AutoML specifically for soil compaction prediction, distinguishing it from prior studies that focused on soil liquefaction, landslide susceptibility, or slope stability. The primary contributions are:

Demonstration of AutoML Efficacy: The study validates that AutoML can automate the selection of optimal algorithms and hyperparameters, leading to higher accuracy and scalability compared to manual ML approaches.
Improved Prediction Accuracy: By leveraging AutoML, the study achieved high $R^2$ values (89.1% for OMC and 80.4% for MDD), suggesting a reliable alternative to traditional laboratory testing.
Foundation for Intelligent Construction: The authors posit that this approach serves as a foundational component for integrating intelligent manufacturing and service systems into the construction industry. They suggest that such models could be embedded into smart infrastructure workflows, potentially enabling real-time decision support and automated equipment adjustments to optimize compaction efforts.
Data-Driven Efficiency: The research contributes to more efficient and reliable construction practices by transforming the prediction of soil compaction parameters from a manual, time-consuming process into a data-driven, automated one.