Explainable Levenberg Marquardt trained neural network paradigm for forecasting concrete compressive strength

Area of Science:

• Structural Engineering
• Materials Science
• Computational Mechanics

Background:

• Concrete compressive strength (CS) is a critical performance indicator in structural engineering.
• Existing machine learning (ML) models for CS prediction often lack transparency.
• There is a need for accurate and explainable models for concrete CS estimation.

Purpose of the Study:

• To develop an accurate and explainable ML model for estimating concrete CS.
• To provide a transparent ML model for concrete CS prediction.
• To facilitate the integration of ML models into structural engineering software.

Main Methods:

• Developed an ML model using the Levenberg-Marquardt algorithm.
• Utilized 1030 laboratory-measured concrete CS data points for model development.
• Evaluated model performance using R², RMSE, AAPRE, and APRE metrics.

Main Results:

• The model achieved a high R² of 0.949 on the test dataset.
• Key contributing variables identified: binder proportion (+0.52), superplasticizer (+0.39), and curing age (+0.35).
• The model is presented in an explicit mathematical form with a user-friendly GUI.

Conclusions:

• The developed ML model provides accurate and transparent estimation of concrete CS.
• The model's explicit mathematical form and GUI facilitate practical application in structural engineering.
• Sensitivity analysis confirms the influence of key mix design parameters on concrete strength.