# Prediction of Impact Resistance of Nano-SiO2 and Hybrid Fiber Modified Geopolymer Gel Concrete in Marine Wet–Thermal and Chloride Salt Environment

## TL;DR

This study improves the impact resistance of geopolymer concrete in harsh marine environments using nano-SiO2 and hybrid fibers, and predicts performance with a neural network model.

## Contribution

A novel RBF-BP neural network model is proposed to predict the impact resistance of nano-SiO2 and hybrid fiber modified geopolymer concrete in marine environments.

## Key findings

- The RBF-BP model outperformed standalone RBF and BP models with R2 values of 0.900 and 0.922.
- NSHFGPC showed improved impact resistance after 60 days of simulated marine exposure.
- Grid search and K-fold cross-validation optimized the neural network's hyperparameters.

## Abstract

The oceanic wet–thermal and chloride salt environment creates extremely harsh service conditions for marine infrastructures. As a green construction material, geopolymer concrete has a promising application prospect in marine engineering due to its excellent durability. The impact resistance of geopolymer concrete subjected to wet–thermal and chloride salt environment is of great significance for the durability and quality of marine engineering structures. This study uses nano-SiO2 (NS) and hybrid fibers (HF) to enhance the impact resistance of geopolymer gel concrete (GPC). Radial basis function (RBF) and back-propagation (BP) composite neural networks are used to predict the impact resistance of NS and HF-reinforced geopolymer gel concrete (NSHFGPC). The impact resistance of NSHFGPC specimens is characterized by two indicators: the cumulative number of repeated impact blows required to initiate the first visible crack (N1) and the cumulative number of impact blows corresponding to ultimate failure (N2). To evaluate the durability of NSHFGPC under oceanic conditions, specimens were exposed to a simulated marine environment within a simulation test chamber for 60 days prior to impact testing. The 60-day duration was selected to achieve a sufficient level of chloride penetration and matrix aging. Based on the resulting experimental database, an RBF-BP neural network was constructed to predict the material’s impact resistance. In this study, grid search and K-fold cross-validation were employed to select the optimal hyperparameters. Compared to standalone RBF and BP models, the RBF-BP network demonstrated superior performance, achieving R2 values of 0.900 and 0.922. These results represent improvements of 20.18% and 11.18% over the standalone RBF model, respectively. Consequently, the RBF-BP algorithm serves as an experimental tool for predicting NSHFGPC impact resistance and guiding future mix design optimization.