⚡ electrical engineering

Development of Machine Learning Algorithms to Predict Damage Level of RC Structures Exposed to Extreme Dynamic Loads

This study proposes and evaluates five machine learning algorithms trained on finite element data to predict the blast-induced damage of reinforced concrete columns, identifying CatBoost as the most effective model and highlighting structural depth, charge weight, and reinforcement yield strength as the primary influencing factors.

Original authors: Masoud Abedini, Chunwei Zhang, Linnan Zhang, Ardashir Mouhamadzadeh, Masoomeh Barmaki

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Masoud Abedini, Chunwei Zhang, Linnan Zhang, Ardashir Mouhamadzadeh, Masoomeh Barmaki

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

Imagine a reinforced concrete (RC) building column as a sturdy, steel-reinforced tree trunk. Now, imagine a massive explosion happening right next to it. The shockwave hits the tree, and the big question is: How much damage will that tree suffer? Will it get a few scratches, crack a branch, or will it snap completely in half?

Traditionally, engineers have to answer this by building expensive, dangerous physical models in labs or running incredibly complex, slow computer simulations that take hours or days to finish. It's like trying to predict the weather by building a new, full-sized atmosphere in a room every time you want to know if it will rain.

This paper introduces a smarter, faster way: Machine Learning (ML). Think of ML as a super-smart student who has studied thousands of "what-if" scenarios and learned the patterns of how these concrete columns react to explosions. Instead of building a new model every time, we just ask the student, "Here are the details of the explosion and the column; what's the damage?"

Here is a breakdown of how they did it and what they found, using simple analogies:

1. The Training Ground (The Data)

Before the "students" (the AI models) could learn, the researchers had to create a massive library of practice problems. They didn't use real explosions (which are dangerous and expensive); instead, they used a sophisticated computer program called LS-DYNA to simulate hundreds of explosions on virtual columns.

They created a dataset with 10 different "ingredients" for each scenario, such as:

How big the column is (depth and width).
How strong the concrete and steel are.
How heavy the bomb is.
How far away the bomb is.

2. The Five Students (The Algorithms)

The researchers didn't just pick one AI; they hired five different "students" with different learning styles to see who was the best at predicting the damage.

Random Forest (RF): Like a committee of experts. They each look at the problem from a slightly different angle and vote on the answer.
LightGBM, CatBoost, and XGBoost: These are like a team of detectives who work in a sequence. The first one makes a guess, the second one looks at the first one's mistakes and tries to fix them, the third one fixes the second one's mistakes, and so on. They get smarter with every step.
SVM (Support Vector Machine): This is like a strict referee trying to draw a straight line (or a smooth curve) to separate "safe" from "damaged."

3. The Test (Who Won?)

The researchers gave all five students a new set of problems they had never seen before to see who could predict the damage most accurately.

The Winner: CatBoost. This model was the clear champion. It was the most accurate and reliable. It learned the complex, messy patterns of explosions better than the others.
The Runner-Up: Random Forest. It was very consistent and stable, though slightly less accurate than the winner.
The Strugglers: XGBoost and LightGBM were okay but had more ups and downs in their performance.
The Loser: SVM. This model completely failed. It was like trying to use a ruler to measure a squiggly line; the explosion damage is too messy and unpredictable for this specific type of "straight-line" thinking.

4. What Actually Matters? (The Key Ingredients)

The winning model (CatBoost) also told the researchers which "ingredients" mattered the most. It's like a chef telling you which spices make the dish taste good.

The Depth of the Column (The Big Winner): The most important factor was how "deep" or thick the column was. Making a column deeper is like making a tree trunk thicker; it makes it incredibly hard to break. This single factor accounted for nearly 38% of the damage prediction.
The Weight of the Bomb: Obviously, a bigger bomb causes more damage, but the relationship isn't a simple "double the bomb, double the damage." It's a bit more complex, but still very important (about 27%).
The Strength of the Steel: The strength of the steel bars running inside the concrete was the third most important factor (14%). Stronger steel helps the column hold together even after the concrete cracks.

Surprising Findings:

Concrete Strength: Surprisingly, making the concrete stronger didn't help as much as expected once it passed a certain point. It's like using a super-hard rock for a doorstop; once it's hard enough, making it harder doesn't stop the door from swinging.
Transverse Reinforcement (The "Stirrups"): The little steel ties that hold the main bars together didn't matter much if the column was already built to code standards. It's like adding extra seatbelts to a car that already has good ones; it doesn't change the outcome much in a crash.

5. Why This Matters

The paper concludes that we don't need to run slow, expensive computer simulations for every single design anymore. We can use the CatBoost model to instantly predict how much damage a column will take from an explosion.

Speed: It's instant.
Accuracy: It's reliable enough to help engineers design safer buildings.
Safety: It helps identify which buildings are vulnerable and which design changes (like making columns deeper or using stronger steel) will save the most lives.

In short, the researchers built a "crystal ball" for engineers. Instead of guessing or waiting days for a simulation, they can now ask the AI, "If we build a column this size with this much steel, and a bomb of this size goes off here, how bad will the damage be?" and get a trustworthy answer in seconds.

Technical Summary: Development of Machine Learning Algorithms to Predict Damage Level of RC Structures Exposed to Extreme Dynamic Loads

1. Problem Statement

Reinforced concrete (RC) structures are increasingly vulnerable to explosive loads due to the rising frequency of terrorist attacks and potential industrial accidents. Traditional methods for assessing blast damage—experimental testing and numerical simulations—face significant limitations. Experimental analysis is costly, time-consuming, and poses severe safety risks. Conversely, while numerical simulations (e.g., Finite Element Analysis using LS-DYNA) offer detailed insights, they require extensive computational resources and specialized expertise, making them impractical for rapid preliminary design or large-scale vulnerability screening.

There is a critical need for a simpler, faster, and accurate method to predict the damage levels of RC columns under blast loading. This study addresses this gap by investigating the efficacy of Machine Learning (ML) algorithms as data-driven alternatives to traditional simulation methods.

2. Methodology

2.1 Data Generation and Numerical Simulation

The research utilized a comprehensive dataset generated through numerical simulations rather than physical experiments.

Software: LS-DYNA was used to model the dynamic response of RC columns under explosive loading.
Model Details: The models represented rectangular RC columns (500×700 mm cross-section, 4400 mm height) with specific longitudinal and transverse reinforcement.
Material Models:
- Concrete: Modeled using MAT_CONCRETE_DAMAGE_REL3 (MAT 72Rel3) to capture nonlinear behavior, damage accumulation, and strain-rate effects.
- Steel: Modeled using MAT_PIECEWISE_LINEAR_PLASTICITY (MAT 24).
- Strain Rate: Dynamic Increase Factors (DIF) based on CEB-FIB Model Code were applied to both concrete and steel to account for the high strain rates typical of blast events.
Loading: Blast loads were simulated using a simplified triangular pressure-time history applied to the column surface, defined by peak overpressure and positive phase duration.
Validation: The numerical model was validated against experimental data from Baylot and Benvis (2007), showing close agreement in mid-height deflection and residual displacement.

2.2 Damage Assessment

A damage index ( $D$ ) was defined based on the residual axial load-carrying capacity of the column:
$D = 1 - \frac{P_{residual}}{P_d}$
Where $P_{residual}$ is the capacity of the damaged column and $P_d$ is the capacity of the undamaged column. Damage levels were classified into four categories: Low ( $0 < D < 0.2$ ), Medium ( $0.2 < D < 0.5$ ), High ( $0.5 < D < 0.8$ ), and Collapse ( $0.8 < D < 1$ ).

2.3 Machine Learning Framework

Five distinct ML algorithms were trained and tested on the dataset derived from the finite element models:

Random Forest (RF)
Light Gradient Boosting Machine (LightGBM)
Categorical Boosting (CatBoost)
Extreme Gradient Boosting (XGBoost)
Support Vector Machine (SVM)

Input Features (10 parameters):

Geometric: Column depth ( $D$ ), width ( $w$ ), height ( $H$ ).
Material: Concrete compressive strength ( $f_c$ ), longitudinal steel yield strength ( $f_y$ ), transverse steel yield strength ( $f_{yt}$ ), longitudinal reinforcement ratio ( $\rho$ ), transverse reinforcement ratio ( $\rho_s$ ).
Blast: Charge weight ( $W$ ), Scaled distance ( $Z$ ).

3. Key Contributions

Comparative Analysis: The study provides a rigorous comparison of five state-of-the-art ML algorithms specifically for the nonlinear, threshold-rich problem of blast damage prediction in RC columns.
Identification of Optimal Algorithm: It identifies CatBoost as the superior model for this specific application, outperforming RF, XGBoost, LightGBM, and SVM in terms of accuracy and reliability.
Feature Importance Quantification: The research quantifies the relative influence of geometric, material, and blast parameters on damage severity, offering data-driven insights for design optimization.
Sensitivity Analysis: The study maps the nonlinear relationships between input variables (e.g., charge weight, column depth) and the resulting damage index, identifying critical thresholds in structural behavior.

4. Results

4.1 Model Performance

Performance was evaluated using $R^2$ , Mean Squared Error (MSE), Mean Absolute Error (MAE), Explained Variance Score (EVS), Mean Absolute Percentage Error (MAPE), and Mean Squared Logarithmic Error (MSLE).

CatBoost (Tier 1): Achieved the highest performance with an $R^2$ of 0.7125, an MAE of 1,061.09, and a MAPE of 17.63%. It demonstrated the best balance of accuracy and reliability.
Random Forest (Tier 2): Showed good performance ( $R^2 = 0.6274$ ) with the highest stability (lowest standard deviation in $R^2$ ), making it a robust alternative.
XGBoost (Tier 2): Achieved competitive mean accuracy ( $R^2 = 0.6109$ ) but exhibited high variability (standard deviation of 0.2313), suggesting sensitivity to data splits.
LightGBM (Tier 3): Delivered marginal performance ( $R^2 = 0.5043$ ), likely due to its leaf-wise growth strategy overfitting localized patterns.
SVM (Tier 4): Failed completely with an $R^2$ of -0.0170, indicating an inability to capture the complex nonlinearities and discontinuous transitions (e.g., from localized damage to collapse) inherent in blast loading.

4.2 Feature Importance

Based on the CatBoost model, the input parameters were ranked by their influence on damage prediction:

Depth of Structural Element ( $D$ ): 38.45% (Dominant factor).
Charge Weight: 27.25%.
Yield Strength of Longitudinal Reinforcement ( $f_y$ ): 14.55%.
Scaled Distance ( $Z$ ): 6.42%.
Transverse Reinforcement Ratio ( $\rho_s$ ): 3.04%.
Yield Strength of Transverse Reinforcement ( $f_{yt}$ ): 2.71%.
Concrete Compressive Strength ( $f_c$ ): 2.66%.
Column Height ( $H$ ): 2.21%.
Column Width ( $w$ ): 1.78%.
Longitudinal Reinforcement Ratio ( $\rho$ ): 0.93%.

4.3 Sensitivity and Thresholds

Depth: A distinct threshold exists between 550 mm and 650 mm. Increasing depth beyond 600 mm yields diminishing returns in damage reduction.
Charge Weight: Damage increases logarithmically with charge weight, plateauing as the structure approaches total collapse.
Concrete Strength: Increasing $f_c$ beyond 45 MPa offers minimal additional benefit due to the strain-rate effect (DIF) which normalizes static strength differences.
Transverse Reinforcement: Within the studied range, variations in transverse reinforcement had negligible impact on the residual axial capacity damage metric, provided minimum code requirements were met.

5. Significance and Claims

The paper claims that the developed ML models, particularly CatBoost, provide a rapid and accurate tool for predicting the structural damage and failure modes of RC columns under blast loading.

Practical Utility: The models can replace computationally expensive finite element simulations for preliminary design and vulnerability screening, provided appropriate safety factors are applied.
Design Guidance: The study offers specific, data-backed design recommendations:
- Prioritize increasing column depth over increasing concrete strength or adding transverse reinforcement for blast resistance.
- Utilize high-strength longitudinal steel ( $f_y \ge 500$ MPa) to leverage strain-rate benefits.
- Avoid specifying concrete strengths significantly above 45 MPa for blast resistance, as the marginal gain is low.
- Focus on scaled distance (standoff) as a primary protective measure during site planning.

The authors conclude that while the models are limited to the parameter ranges and numerical simulation data used, they represent a significant step toward efficient, data-driven blast-resistant design. The study explicitly avoids claiming universal applicability beyond the tested ranges and suggests future work should incorporate experimental data and extend to other structural elements.