Research on the Performance Prediction of Fly Ash Slurry by Neural Network Based on Information Bottleneck Theory and Attention Mechanism

In grouting projects, the relationship between slurry properties and mix proportions is multivariate and nonlinear. Researchers have widely employed mathematical and computational methods to investigate this relationship, and the use of machine learning to predict the properties of fly ash slurries has gradually become a hot topic. To this end, this paper uses a convolutional neural network (CNN) to establish predictive models relating grout mix proportions to the properties of the grout and the resulting stone bodies. It investigates the effects of the water-to-binder ratio, fly ash dosage, and water glass volumetric dosage on the slurry density, viscosity, initial setting time, and final setting time of fly ash grout, as well as on the compressive strength and stone rate of the corresponding stone bodies. The CNN is optimized using the information bottleneck theory and attention mechanisms, and finally, based on nine sets of self-designed experiments, compares four neural network models for predicting fly ash performance. The results show that the information bottleneck–attention dual-optimized convolutional neural network had an average prediction error of 10.26%, which represents a 44% improvement over the average prediction error of 18.38% for a single convolutional neural network. It also achieved multi-objective performance prediction for fly ash slurry. Finally, the reliability of this prediction model was verified through a combination of laboratory and field tests.