Application of surrogate models based on neural networks for fast optimization of reinforced concrete frame structures

Abstract

This paper presents the development, implementation, and validation of a comprehensive software pipeline for creating and applying a high-fidelity neural network-based surrogate model, designed to solve the multi-parameter optimization problem of building structures. The design optimization of cast-in-situ reinforced concrete frames is a fundamentally complex computational challenge. Its difficulty arises from the high cost of Finite Element (FE) analysis, which is the primary tool for verifying compliance with regulatory requirements, rendering traditional multi-variant searches for optimal solutions via direct enumeration impractical within realistic design timelines.

The proposed methodology involved the creation of a detailed parametric FE model of a spatial cell of a reinforced concrete frame in the Ansys environment, the automated generation of a representative dataset with 12000 unique design parameter combinations, and their subsequent engineering post-processing. A key stage of the post-processing, which ensured the physical correctness of the training data, was the calculation of the required reinforcement area for each element. This calculation was performed using a specially developed iterative algorithm that fully implements the nonlinear deformation model in accordance with the current Ukrainian State Standard DSTU B V.2.6-156:2010 (harmonized with Eurocode 2: EN 1992-1-1).

Based on this enriched dataset, a Multi-Layer Perceptron (MLP) model was trained and validated. The results of the final testing on a hold-out set demonstrated that the trained surrogate model achieved high predictive capability with a coefficient of determination R² = 0.9946. When applied to a practical optimization problem with a minimum cost criterion, the model was able to analyze 10 million candidate designs in approximately one minute – a task that would require an estimated 8 years of continuous computation using direct FE analysis. The final verification of the top 10 optimal solutions, performed by their full analysis in Ansys, showed a high convergence between predicted and actual values, with the prediction error for the objective function (cost) not exceeding 2.1%.

Thus, this work demonstrates that the proposed approach, which combines modern engineering computational models and machine learning methods, enables the creation of reliable predictive tools that accelerate the process of finding optimal structural solutions by orders of magnitude, opening new possibilities for efficient and economically justified design.