Two stage AI framework for strength prediction and generative LLM for geopolymer concrete

Area of Science:

• Materials Science
• Artificial Intelligence
• Civil Engineering

Background:

• Geopolymer concrete offers a sustainable, low-carbon alternative to Portland cement.
• Current geopolymer concrete mix design relies heavily on time-consuming laboratory testing.
• Accelerating the development of high-performance geopolymer concrete is crucial for wider adoption.

Purpose of the Study:

• To develop a novel two-stage artificial intelligence (AI) system for rapid identification of high-performance geopolymer concrete formulations.
• To reduce the extensive laboratory testing currently required for geopolymer concrete mix design.
• To enhance the accuracy and efficiency of geopolymer concrete mix design using AI.

Main Methods:

• Training machine learning models, including Genetic Algorithm optimized XGBoost (GA XGBoost), TabTransformer, and ANN LM, on 820 geopolymer concrete mixes for compressive strength prediction.
• Employing a hybrid generative model combining a fine-tuned Large Language Model (LLM) with XGBoost for mix design generation.
• Refining LLM-generated textual mix designs by replacing original numerical values with accurate predictions from the XGBoost model.

Main Results:

• GA XGBoost demonstrated superior predictive accuracy for compressive strength (R²=0.9648, RMSE=2.8823, MAE=1.9053).
• The hybrid LLM-generative model achieved high performance with BERTScore (0.9754) and ROUGE L (0.8794).
• Numerical predictions for most target features in the generative model exceeded an R² value of 0.90.

Conclusions:

• The developed AI system significantly speeds up the identification of high-performance geopolymer concrete formulations.
• The hybrid LLM-XGBoost model effectively generates geopolymer concrete mix designs with improved semantic accuracy and numerical precision.
• This AI approach offers a promising solution to overcome the limitations of traditional laboratory-based mix design methods.