When you drive past a steep highway embankment or retaining wall, you probably don’t give much thought to what’s keeping all that soil in place. Behind many of these structures are remarkable engineering materials called geogrids that work silently to reinforce the soil. But how do engineers know these reinforcements will hold? A fascinating new study from researchers at IIT Hyderabad has applied cutting-edge machine learning techniques to predict how well geogrids perform in different soil types, potentially saving countless hours of laboratory testing and making construction safer and more efficient.
Geogrids are synthetic mesh-like materials used to reinforce soil structures. If you’ve ever seen plastic netting with regular openings, that’s what geogrids basically look like, except they’re specifically engineered to be extremely strong. These materials are typically made from polymers and are designed with openings (called apertures) that allow soil particles to interlock with them.
Engineers use geogrids in various applications:
- Reinforcing highway embankments
- Stabilizing retaining walls
- Supporting bridge abutments
- Reinforcing slopes in challenging terrain
When soil is reinforced with geogrids, it creates what engineers call Mechanically Stabilized Earth (MSE) walls. These structures depend heavily on a critical property called “pullout resistance”, essentially how well the geogrid stays anchored in the soil when forces try to pull it out.
Traditionally, measuring this pullout resistance has been a major headache for engineers. It requires expensive laboratory testing where geogrids are embedded in soil samples and pulled out under controlled conditions. These tests are time-consuming, costly, and need to be repeated for different soil types and geogrid configurations.
A team of researchers from IIT Hyderabad, led by Vaishnavi Bherde, Samay Kumar Attara, and Professor Umashankar Balunaini, recognized an opportunity to use machine learning to predict pullout resistance without expensive testing.
Their study details how they used advanced AI techniques to tackle this problem. What makes this study particularly impressive is the massive dataset they assembled, 759 data points collected from 58 different research papers covering a wide variety of soil types and geogrid properties.
“Conducting pullout tests on geogrids is labor-intensive, time-consuming, and expensive,” the researchers note. Their goal was to develop models that could predict a key design parameter called the pullout resistance factor (F*) using readily available soil and geogrid properties.
The researchers applied several machine learning techniques to predict pullout resistance. But what exactly were they trying to predict?
The pullout resistance factor (F*) is a dimensionless parameter used in design codes like the Federal Highway Administration (FHWA) guidelines. It’s used in this equation:
Pr = C·σ’v·Le·F*·α
Where:
- Pr is the pullout load in kN/m
- C is the reinforcement effective unit perimeter (C=2 for geogrid)
- σ’v is the effective normal stress in kPa
- Le is the embedment length of geogrid in meters
- F* is the pullout resistance factor
- α is the scale correction factor (0.8 for geogrids)
Currently, engineers often use a simplified equation: F* = (2/3)·tan φ, where φ is the peak friction angle of the soil. But this simplification doesn’t account for many factors that affect real-world performance.
The researchers identified eight key parameters that influence pullout resistance:
- Normal stress (σ’v) – The pressure exerted on the geogrid by the soil above it
- Relative compaction (RC) – How densely the soil is packed
- Fines content (FC) – The percentage of tiny particles in the soil
- Embedment length (Le) – How much of the geogrid is buried in the soil
- Pullout displacement rate (PR) – How quickly the geogrid is pulled during testing
- Spacing ratios (St/D50 and Sl/D50) – The ratio of spacing between geogrid ribs to soil particle size
- Ultimate tensile strength (Tult) – How much force the geogrid can withstand before breaking
The team tested six different machine learning models:
- Random Forest (RF): A method that creates multiple decision trees and merges their predictions
- Gradient Boosting Model (GBM): Creates a sequence of models, each trying to correct errors from previous ones
- Extreme Gradient Boosting (XGB): An optimized version of GBM that’s faster and more efficient
- Light Gradient Boosting Model (LGB): Another improved version of GBM that uses less memory
- Decision Tree (DT): A single tree-like model that makes decisions based on features
- Multilinear Regressor (MLR): A traditional statistical method that fits a linear equation
These might sound complicated, but think of them as different approaches to learning patterns in data some more sophisticated than others.
After rigorous testing, the XGB model emerged as the clear winner, achieving an impressive accuracy with an R² value of 0.91 and RMSE of 0.18 on the test dataset. In simple terms, this means the model could predict pullout resistance with about 91% accuracy.
What’s particularly interesting is that the researchers identified which factors matter most in determining pullout resistance:
- Normal stress – The pressure from the soil above had the biggest impact
- Embedment length – How much of the geogrid is buried in the soil
- Relative compaction – How densely the soil is packed
This makes intuitive sense. More pressure from above (higher normal stress) creates more friction to hold the geogrid in place. A longer embedded length gives more surface area for the soil to grip. And more compacted soil provides stronger interlocking with the geogrid.
The researchers found that even with just these three parameters, they could still predict pullout resistance with reasonable accuracy (R² = 0.80). This is a game-changer for practical applications, as these three properties are relatively easy to measure in the field.
The team didn’t stop at theoretical models. They conducted actual laboratory pullout tests using their own experimental setup to validate their findings. Using three different types of geogrids embedded in sand, they performed six different tests varying the parameters they had identified as most important.
The results were remarkable as their XGB model predicted the pullout resistance with an error of less than 3% compared to the laboratory measurements. This kind of accuracy demonstrates the real-world applicability of their approach.
Let’s consider a practical example. Imagine you’re an engineer designing a 20-foot high retaining wall for a new highway project. The wall will be built with locally available sandy soil reinforced with geogrids.
Traditionally, you would need to:
- Send soil samples to a specialized laboratory
- Wait weeks for pullout test results
- Pay thousands of dollars for testing
- Possibly redesign if the results aren’t what you expected
With this new approach, you could:
- Measure the normal stress (based on the height of soil above each geogrid layer)
- Determine your planned embedment length
- Measure the relative compaction of your soil
- Input these values into the model
- Get an immediate prediction of pullout resistance
- Optimize your design on the spot
This could save weeks of time and thousands of dollars in testing costs for a single project. Multiply that across the thousands of infrastructure projects built globally each year, and the impact becomes enormous.
This research represents a significant step forward in applying artificial intelligence to civil engineering problems. Here’s why it matters:
- Cost Reduction: Laboratory pullout tests can cost thousands of dollars each. Eliminating or reducing these tests leads to significant savings.
- Time Savings: Instead of waiting weeks for test results, engineers can get predictions instantly.
- Optimization: Engineers can quickly evaluate multiple design alternatives to find the most efficient solution.
- Sustainability: More efficient designs mean less material usage, contributing to more sustainable construction practices.
- Accessibility: This approach makes advanced design techniques available to smaller projects and in regions where specialized testing facilities might not be readily available.
The researchers have even created a user-friendly graphic interface (GUI) for their model, making it accessible to practicing engineers without specialized knowledge of machine learning.
This research exemplifies a broader trend in engineering using data-driven approaches to solve complex problems that were previously tackled with simplifications or extensive physical testing.
In geotechnical engineering, where soil behavior is notoriously complex and variable, machine learning offers a powerful new tool. The ground beneath our feet contains countless variables such as particle sizes, moisture content, mineral composition, and more making it extremely difficult to model using traditional approaches.
Machine learning excels at finding patterns in this complexity. As more data becomes available and algorithms improve, we can expect to see these techniques applied to increasingly complex geotechnical problems.
In conclusion, the research by Bherde, Attara, and Balunaini demonstrates how machine learning can transform a specific but critical aspect of geotechnical engineering predicting geogrid pullout resistance. Their approach not only saves time and money but potentially enables more optimized and safer designs.
For civil engineers, this research provides a valuable new tool. For the rest of us, it means safer infrastructure built more efficiently. The next time you drive past a reinforced soil structure, remember that artificial intelligence might have helped ensure it stays firmly in place keeping us all safer while saving resources.
The integration of traditional civil engineering with cutting-edge machine learning represents the future of infrastructure design, where computer models and physical understanding work hand in hand to create better, more efficient structures.
Reference
Abdi, & Mirzaeifar, H. (2017). Experimental and PIV evaluation of grain size and distribution on soil–geogrid interactions in pullout test. SOILS AND FOUNDATIONS, 57(6), 1045–1058. https://doi.org/10.1016/j.sandf.2017.08.030
Bergado, D., Bukkanasuta, A., & Balasubramaniam, A. (1987). Laboratory pull-out tests using bamboo and polymer geogrids including a case study. Geotextiles and Geomembranes, 5(3), 153–189. https://doi.org/10.1016/0266-1144(87)90015-x
Bherde, V., Attara, S. K., & Balunaini, U. (2025). Ensemble-based approach for automatic prediction of pullout resistance of geogrids in different soil types. Geotextiles and Geomembranes, 53(4), 1035–1047. https://doi.org/10.1016/j.geotexmem.2025.03.004
Mittal, M., Satapathy, S. C., Pal, V., Agarwal, B., Goyal, L. M., & Parwekar, P. (2021). Prediction of coefficient of consolidation in soil using machine learning techniques. Microprocessors and Microsystems, 82, 103830. https://doi.org/10.1016/j.micpro.2021.103830
Miyata, Y., & Bathurst, R. J. (2012). Reliability analysis of soil-geogrid pullout models in Japan. SOILS AND FOUNDATIONS, 52(4), 620–633. https://doi.org/10.1016/j.sandf.2012.07.004