Prediction of mechanical properties of basalt fiber concrete using hybrid recurrent neural networks based on freeze-thaw damage quantification

In the domain of contemporary building materials, the freeze durability of concrete is paramount for ensuring structural safety and longevity. However, traditional assessment methods have hindered research advancement due to their time-intensive and intricate nature. This study presents a novel model that integrates physical damage mechanisms with machine learning to analyze freeze-thaw damage in basalt fiber concrete (BFC). Experimental data on the performance of BFC with various admixtures subjected to freeze-thaw cycles (FTC) were collected. Due to the limited data availability, a hybrid recurrent neural network (HRNN) model was developed by combining recurrent neural network (RNN) and long-short-term memory network (LSTM) techniques to address the challenge of training an effective neural network. Furthermore, irreversible thermodynamics and continuous damage mechanics were incorporated into the loss function of the original model, resulting in a novel hybrid damage function that served to train and constrain the model. The data flow and structural parameters were meticulously adjusted to enhance prediction accuracy and model stability. The results indicate that basalt fibers significantly enhance the frost resistance of concrete, particularly at a dosage of 0.1 %, where the model achieves its lowest prediction error of less than 10 %. Stability and robustness were evidenced under the influence of data with identical parameters. Sensitivity analysis underscored the model's precision in capturing the effects of various parameters. By synergizing thermodynamic principles with data-driven methodologies, the model exhibited remarkable adaptability across FTC damage scenarios. This study offers a novel solution for enhancing the durability of concrete in cold regions and opens up new avenues for the engineering applications of multi-material systems.