⚡ electrical engineering

(FEM-ML) Hybrid models to predict the lateral pile cap displacement under quasi-static load

This study proposes a novel hybrid framework combining Finite Element Modeling (FEM) and Artificial Intelligence, specifically utilizing a Particle Swarm Optimization-tuned Gradient Boosting Regressor and Evolutionary Polynomial Regression, to accurately predict the lateral displacement of single and grouped piles under quasi-static loads, thereby overcoming the limitations of traditional methods in capturing complex non-linear soil-structure interactions.

Original authors: Yossef E. Teleb, Mahmoud S. Hammad, Ahmed M. Ebid, Abdel Salaam Ahmed Mokhtar, Mohamed Gamaleldeen Elsayem

Published 2026-07-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yossef E. Teleb, Mahmoud S. Hammad, Ahmed M. Ebid, Abdel Salaam Ahmed Mokhtar, Mohamed Gamaleldeen Elsayem

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

The Big Problem: Guessing How Much a Building Will Wiggle

Imagine you are building a skyscraper or a bridge. These structures need deep roots, called piles, to stay standing. Sometimes, strong winds or earthquakes push these buildings sideways. Engineers need to know exactly how much the top of the building (the "pile cap") will wiggle or shift before it becomes dangerous.

Traditionally, engineers have two ways to figure this out:

The "Rule of Thumb" Way: Fast and easy, but often too simple. It's like guessing the weather by looking at the sky; it works sometimes, but it misses the complex details.
The "Super-Computer" Way (FEM): This uses powerful physics simulations to model every grain of soil and every crack in the concrete. It's incredibly accurate, but it's like trying to bake a cake by measuring every single molecule of flour. It takes forever and costs a lot of computing power.

The authors of this paper wanted to find a "Goldilocks" solution: a method that is as fast as the rule of thumb but as accurate as the super-computer.

The Solution: A Hybrid "Chef and Sous-Chef" Team

The researchers created a hybrid model that combines two worlds: Finite Element Modeling (FEM) and Artificial Intelligence (AI).

Think of it like a high-end restaurant kitchen:

The FEM (The Master Chef): This is the expert who knows the physics perfectly. They cook the "perfect" meals (simulations) to understand exactly how the soil and concrete react. However, the Master Chef is slow and expensive to hire.
The AI (The Sous-Chef): This is a fast learner. The researchers let the Master Chef cook 450 different "meals" (simulations) with different ingredients (soil types, pile sizes, spacing). The AI watches these 450 meals, learns the patterns, and then creates its own recipe book.

Once the AI learns the patterns, it can predict the result of a new meal instantly, without needing the Master Chef to cook it again.

How They Did It (The Recipe)

1. The Training Phase (Cooking the 450 Meals)
The team used a computer program (PLAXIS 3D) to simulate 450 different scenarios. They changed the ingredients:

Soil: Soft clay vs. hard sand.
Pile Groups: A single pile vs. a big grid of piles (like a 2x2 or 6x6 arrangement).
Loads: How heavy the building is and how hard the wind pushes.
Concrete Cracks: They specifically modeled how concrete cracks under pressure, which makes it weaker (like a dry twig snapping).

2. The Magic Transformation
The raw data from these simulations was messy, like a giant pile of unsorted ingredients. To make it easy for the AI to learn, they turned the complex curves of "load vs. wiggle" into two simple numbers:

Initial Stiffness: How hard it is to start moving the pile (like pushing a heavy car from a stop).
Ultimate Capacity: How much force it takes to break the system (like pushing the car until the tires blow out).

3. The AI Learning (The Sous-Chef Gets Smart)
They fed these two numbers into different types of AI "brains" to see which one learned best:

Linear Regression: A simple, straight-line guess (too simple).
Neural Networks & Support Vector Machines: Complex black boxes that are hard to understand.
Gradient Boosting (GBR): This turned out to be the star student. It was like a detective that looked at the clues one by one, correcting its mistakes until it got the answer almost perfectly.

The Result: The AI model (GBR) was incredibly accurate. It predicted the wiggle with a correlation of nearly 99%. It was so good that it could predict the answer in a fraction of a second, whereas the physics simulation took hours.

Making Sense of the "Black Box"

One problem with AI is that it's often a "black box"—it gives an answer, but you don't know why. Engineers don't trust things they can't explain.

To fix this, the researchers used a tool called SHAP. Think of SHAP as a translator that explains the AI's thought process. It told them:

Geometry is King: The number of piles and how far apart they are spaced matters the most for how much weight the group can hold.
Soil is Queen: The type of soil (how stiff or sticky it is) matters the most for how much the pile wiggles at first.
Vertical Load: Putting more weight down on the pile actually helps it resist being pushed sideways (like pressing down on a spring makes it harder to push over).

The "Magic Formula" (EPR)

While the AI was great at predicting, the researchers also wanted a formula that engineers could write down on paper. They used a method called Evolutionary Polynomial Regression (EPR).

Imagine the AI trying to write a poem. EPR is like a tool that takes the AI's complex thoughts and turns them into a simple, readable sentence (a mathematical equation).

Why it matters: Engineers can now use this specific equation to calculate the wiggle of a pile group without needing a supercomputer or a black-box AI. It bridges the gap between high-tech AI and traditional engineering math.

What They Found (The Takeaway)

Cracking Matters: For long piles, the fact that concrete cracks is a huge deal. If you ignore the cracks, your model will think the pile is stiffer than it really is.
Group Effect: Piles in a group don't just act like individual piles; they interact. If they are too close, they "shadow" each other, making the group less efficient. The AI learned exactly how to calculate this.
Speed vs. Accuracy: The hybrid model is the best of both worlds. It gives the accuracy of the slow physics simulation but runs as fast as a simple calculator.

Limitations (The Fine Print)

The authors are honest about what their model can't do yet:

The "Plateau" Problem: If the force pushing the pile is extremely high (beyond what they simulated), the AI stops predicting accurately and just gives a flat answer. It's like a GPS that stops giving directions once you leave the map area.
Soft Soil: The model is very good for medium-to-hard soils but struggles a bit with very soft, squishy clay.
Real World vs. Simulation: They tested their model against real-world data (like the Cairo Monorail project) and it worked well, but they admit they need more real-world data from different places to make it perfect.

In short: The authors built a digital "crystal ball" that uses physics simulations to train an AI. This AI can now instantly tell engineers how much a group of concrete piles will wiggle under a sideways push, accounting for soil types, pile spacing, and concrete cracks, all while providing a clear mathematical formula they can trust.

Technical Summary: (FEM-ML) Hybrid Models to Predict Lateral Pile Cap Displacement Under Quasi-Static Load

1. Problem Statement

The accurate prediction of lateral pile cap displacement under quasi-static loads remains a significant challenge in geotechnical engineering due to the complex, non-linear nature of Soil-Structure Interaction (SSI). Traditional analytical methods, such as Ultimate Limit State approaches (e.g., Broms) and Subgrade Reaction methods (e.g., p-y curves), often rely on simplifications that fail to capture intricate soil deformations, group effects, and the non-linear structural degradation of reinforced concrete (RC) piles, particularly cracking. Conversely, sophisticated 3D Finite Element Modeling (FEM) offers high fidelity but is computationally expensive and time-consuming, making it unsuitable for rapid design iterations. Furthermore, existing Machine Learning (ML) applications in this domain often suffer from a lack of transparency ("black-box" nature) and a scarcity of high-quality field datasets for large-diameter piles. This study addresses the trade-off between accuracy and efficiency by proposing a hybrid framework that leverages FEM to generate synthetic data for training interpretable and accurate AI models.

2. Methodology

The research employs a four-phase hybrid methodology combining physics-based simulation with data-driven modeling:

Phase 1: FEM Development and Validation

A robust 3D numerical model was developed using PLAXIS 3D (v24.2).

Constitutive Models: The Hardening Soil (HS) model simulated soil behavior, while the Concrete Model (CM) was used for RC piles to explicitly simulate cracking and the reduction of flexural rigidity. Steel reinforcement was modeled as embedded plate elements with interface elements to simulate bond-slip behavior.
Validation: The FEM framework was benchmarked against three full-scale experimental load tests: two from the Cairo Monorail Project (Piles 60C00 and NCC19) and one from the Kiev Pile Load Test. The inclusion of the CM significantly improved accuracy for long piles compared to Linear Elastic (LE) models, with errors remaining within an acceptable engineering range (e.g., -3.71% to 2.54%).

Phase 2: Parametric Study and Synthetic Data Generation

Following validation, a comprehensive parametric study generated 450 unique 3D FEM simulations covering:

Configurations: Single piles and square groups (2×2, 4×4, 6×6).
Geometry: Spacing-to-diameter (S/D) ratios of 2–8 and Length-to-diameter (L/D) ratios of 10, 20, and 30.
Soil Profiles: Six types (3 cohesive clays, 3 cohesionless sands) using the HS model.
Loading: Quasi-static lateral loads up to 25 mm displacement, with varying vertical loads to isolate axial influence.
Data Transformation: Raw load-displacement curves were fitted to a hyperbolic function to extract two physically meaningful targets: Initial Stiffness ( $K_{init}$ ) and Ultimate Lateral Capacity ( $P_{ult}$ ). These targets were log-transformed to normalize data variance.

Phase 3: AI Model Development

The synthetic dataset was split (80/20) into training and testing sets for sand and clay soils. Several ML algorithms were trained to map seven input features (spacing, pile count, length, vertical load, Young's modulus, unit weight, and shear strength parameters) to the target outputs:

Algorithms: Linear Regression (baseline), Gradient Boosting Regressor (GBR), Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP).
Optimization: Hyperparameters for non-linear models were tuned using Particle Swarm Optimization (PSO) to minimize 5-fold Cross-Validation RMSE.
Interpretability: Evolutionary Polynomial Regression (EPR) was employed to derive transparent, symbolic mathematical equations. SHAP (SHapley Additive exPlanations) analysis was applied to the best-performing GBR models to quantify feature importance.

3. Key Results

Model Performance

GBR Superiority: The PSO-tuned GBR model achieved the highest predictive accuracy. For clay stiffness prediction, it reached a correlation coefficient approaching 0.99 with an RMSE as low as 0.015. It consistently outperformed SVR and MLP architectures across all datasets.
EPR Transparency: The EPR models successfully generated closed-form symbolic equations (e.g., Eq. 3–6) that relate geotechnical inputs to pile responses, offering a balance between accuracy and engineering interpretability.

Parametric Insights

Group Effects: Increasing the number of piles significantly enhanced both initial stiffness and ultimate capacity, with the gain in capacity being more dominant in sandy soils.
Spacing: Wider spacing reduced the "shadowing" effect, generally increasing capacity, though group size was the primary governing factor.
Vertical Loading: Vertical loads significantly increased ultimate lateral capacity (due to confinement and delayed cracking) but had a negligible effect on initial elastic stiffness.
Soil Stiffness: The soil elastic modulus was the primary governing parameter for initial stiffness.

Validation and Limitations

Field Validation: The GBR model was validated against a field case (Raj et al., 2024) and showed high correlation in the initial elastic range ( $y < 0.02$ m). However, at higher displacements ( $y > 0.05$ m), the model tended to underpredict ultimate capacity, likely due to the training data's focus on the initial 0.025 m displacement and the limitations of equivalent soil parameters in stratified deposits.
Benchmarks: Against classical methods (Broms, LPile), the GBR model performed well for serviceability limits (±20% error) but exhibited a "plateauing effect" for ultimate capacity in high-capacity regimes, a known limitation of tree-based ensembles that cannot extrapolate beyond training data maxima.

4. Significance and Contributions

The paper claims that this hybrid FEM-AI approach provides a highly accurate, data-driven tool that advances the reliability and efficiency of lateral pile foundation design. Its primary contributions include:

Bridging the Accuracy-Efficiency Gap: The study demonstrates that hybrid models can provide the rigorous accuracy of 3D numerical modeling in a fraction of the time required for direct simulation, making them suitable for preliminary design iterations.
Addressing the "Black-Box" Problem: By integrating EPR and SHAP, the study transforms opaque AI predictions into explainable tools. The SHAP analysis confirmed that the AI learned physical principles, such as the dominance of geometry in capacity and soil properties in stiffness.
Explicit Modeling of RC Cracking: The validation highlighted the critical necessity of using advanced constitutive models (CM) to capture concrete cracking, which significantly affects the flexural rigidity and displacement of long piles.
Symbolic Design Equations: The EPR-derived equations offer engineers transparent, closed-form relationships for predicting pile group responses, facilitating direct integration into design workflows without relying on complex software.

The authors conclude that while the current model is a robust tool for medium-to-hard soils under operational loads, future work must focus on expanding datasets with diverse field measurements, integrating complex multi-layered soil profiles, and advancing Physics-Informed Neural Networks (PINNs) to handle large-scale 3D SSI problems more effectively.