Domain-Informed Genetic Superposition Programming: A Case Study on SFRC Beams

Mohammad Sadegh Khorshidi, Navid Yazdanjue, Hassan Gharoun, Mohammad Reza Nikoo, Fang Chen, Amir H. GandomiarXiv 2025 · DOI · arXiv

Introduction

In structural engineering, the ability to predict the failure of components like beams is critical for safety and efficiency. While deep learning models have demonstrated high predictive accuracy across various domains, their "black-box" nature often limits their adoption in high-stakes engineering applications where interpretability and physical consistency are paramount. This paper presents Domain-Informed Genetic Superposition Programming (DIGSP), a symbolic regression framework designed to bridge the gap between high-performance machine learning and physically intuitive modeling.

Problem Statement: Shear Strength in SFRC Beams

Predicting the ultimate shear strength ($V_u$) of steel fiber-reinforced concrete (SFRC) beams is a complex challenge. The shear capacity is influenced by a non-linear interplay between concrete matrix properties, longitudinal reinforcement, and the bridging effect of steel fibers. Traditional empirical models often oversimplify these interactions, while standard symbolic regression (SR) techniques—which evolve mathematical expressions using Genetic Programming (GP)—frequently suffer from "feature entanglement." This results in overly complex, "bloated" equations that lack physical meaning and fail to generalize.

The DIGSP Framework

DIGSP addresses these limitations by incorporating domain knowledge directly into the evolutionary search process. The framework introduces three primary innovations:

  1. Domain-Informed Feature Partitioning: Input variables are grouped into semantically coherent subsets based on physical mechanisms: Concrete + Geometry, Steel + Geometry, and Fiber + Geometry.
  2. Isolated Sub-populations: The algorithm evolves independent populations of equations for each feature group. This encourages the discovery of "pure" expressions for specific physical contributions (e.g., fiber bridging) before they are combined.
  3. Adaptive Hierarchical Symbolic Abstraction Mechanism (AHSAM): To prevent stagnation, DIGSP uses ANOVA to identify statistically significant sub-expressions. These are abstracted into high-level building blocks and shared across populations, enabling the formation of a holistic, multi-mechanism model.
  4. Complexity Control: The final recombination uses Elastic Net regression to weight the evolved "genes," penalizing unnecessary complexity and ensuring a parsimonious final model.

Dataset and Variables

The framework was evaluated using a benchmark dataset of 213 experimental instances of SFRC beams without stirrups. The dataset includes nine input variables covering geometry, material strengths, and fiber characteristics.

Table of variables and distribution of experimental data for SFRC beams.

The target variable for the regression is the ultimate shear strength, $V_u$ (MPa). The data was split into training (65%), validation (10%), and testing (25%) sets to ensure robust evaluation.

Key Results and Discussion

DIGSP was benchmarked against a standard Multi-Gene Genetic Programming (BGP) model. The results highlight significant improvements in both accuracy and model interpretability:

  • Predictive Accuracy: DIGSP consistently outperformed the baseline, achieving lower Root Mean Squared Error (RMSE) on both training and test sets. Statistical significance was confirmed via a Wilcoxon rank-sum test ($p < 0.01$).
  • Model Parsimony: One of the most striking findings was the reduction in model complexity. The median tree size for DIGSP models was approximately 41 nodes, compared to 204 nodes for the BGP baseline—a 5x reduction in size.
  • Physical Consistency: Elasticity analysis demonstrated that DIGSP-evolved formulas correctly capture expected physical trends, such as the positive correlation between fiber volume fraction ($V_f$) and shear capacity.

Conclusion

DIGSP represents a significant step toward "Physics-AI" integration in structural engineering. By leveraging the principle of superposition and domain-informed partitioning, the framework produces accurate, compact, and physically consistent mathematical models. Future work will explore the application of DIGSP to larger datasets and more complex multi-physics problems, such as fluid-structure interactions.

Summary updated

BibTeX
@misc{khorshidi2025domain,
  author = {Mohammad Sadegh Khorshidi and Navid Yazdanjue and Hassan Gharoun and Mohammad Reza Nikoo and Fang Chen and Amir H. Gandomi},
  title = {{Domain-Informed Genetic Superposition Programming: A Case Study on SFRC Beams}},
  year = {2025},
  eprint = {2509.21355},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.21355},
}
Figure