Research on Hyperparameter Optimization of Concrete Slump Prediction Model Based on Response Surface Method.

Area of Science:

• Civil Engineering
• Materials Science
• Artificial Intelligence

Background:

• Predicting concrete slump is crucial for quality control in construction.
• Traditional methods may lack precision and efficiency.
• Neural networks offer a promising approach for complex material behavior prediction.

Purpose of the Study:

• To develop and optimize a back-propagation (BP) neural network model for accurate concrete slump prediction.
• To identify the key variables influencing concrete slump.
• To enhance the prediction accuracy of neural network models through parameter optimization.

Main Methods:

• Constructed a BP neural network using eight input variables (cement, slag, fly ash, water, superplasticizer, aggregates, flow) and slump as output.
• Tuned hyperparameters including learning rate, momentum factor, hidden nodes, and iterations for 2-layer and 3-layer networks.
• Applied the Response Surface Method (RSM) to optimize the neural network parameters.

Main Results:

• The RSM-optimized network model showed a higher coefficient of determination on the test set compared to the unoptimized model.
• The optimized model achieved superior prediction accuracy for concrete slump.
• Identified flow, water, coarse aggregate, and fine aggregate as the primary influencing factors on slump, with flow having the maximum influence (0.875).

Conclusions:

• The RSM optimization significantly improves the prediction accuracy of BP neural networks for concrete slump.
• The study provides a novel method for efficiently adjusting neural network parameters to enhance concrete slump prediction.
• Understanding the influence of input variables is key to controlling and predicting concrete slump effectively.