Strength Prediction Method for Phosphogypsum Concrete Based on Dynamic Weighted Transfer Learning.

Area of Science:

• Materials Science
• Civil Engineering
• Sustainable Construction

Background:

• Phosphogypsum concrete (PGC) offers high-value utilization for industrial solid waste.
• Limited experimental samples hinder accurate compressive strength prediction for PGC.
• Traditional models struggle with PGC characterization due to data scarcity.

Purpose of the Study:

• To develop a dynamic weighted transfer learning method for accurate PGC compressive strength prediction.
• To address the challenge of small-sample characterization in PGC.
• To enable knowledge transfer from conventional concrete to PGC.

Main Methods:

• Component proportion normalization and feature alignment to address differences between conventional concrete and PGC.
• Bootstrap Resampling for data augmentation to expand training samples.
• Error feedback-driven dynamic weight calculation and weighted loss optimization for transfer learning.

Main Results:

• The transfer learning model achieved R² = 0.95 on PGC test samples, a 15.9% improvement over traditional methods.
• Robust performance (R² = 0.97) was maintained on external validation samples.
• Shapley Additive Explanations (SHAP) analysis revealed nonlinear coupling effects of PGC parameters on strength.

Conclusions:

• The proposed method accurately predicts PGC strength under small-sample conditions.
• Transfer learning effectively overcomes data scarcity limitations in PGC research.
• This study provides a scientific foundation for PGC strength prediction and application.