What question did this study set out to answer?

The aim is to develop an AI framework to improve the prediction of compressive strength and the design process of geopolymer concrete.

April 28, 2026Open Access

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

Key Points

The aim is to develop an AI framework to improve the prediction of compressive strength and the design process of geopolymer concrete.
Trained multiple machine learning models including GA XGBoost on a dataset of 820 geopolymer concrete mixes.
Implemented a hybrid generative model combining a fine-tuned LLM with XGBoost to refine mix design outputs.
Evaluated models based on predictive accuracy and generation scores.
GA XGBoost achieved R² score of 0.9648, RMSE of 2.8823, and MAE of 1.9053 on the test set.
The generative mix design had a BERTScore of 0.9754 and a ROUGE L score of 0.8794.
Numerical predictions for target features showed R² values greater than 0.90.

Abstract

Geopolymer concrete is a low carbon alternative promising replacement for the conventional Portland cement, yet its mix design process still depends to a great extent on lengthy laboratory testing. This article presents a new two stage artificial intelligence system aimed at speeding up the identification of high performance geopolymer concrete formulations through the combination of generative and predictive models. The initial phase includes training a number of machine learning models such as a Genetic Algorithm optimized XGBoost (GA XGBoost), TabTransformer and Levenberg Marquardt optimised Artificial Neural Network (ANN LM) on a set of 820 Geopolymer concrete mixes drawn from scientific literature to make predictions for compressive strength. Among the compressive strength predicting models GA XGBoost performed better in terms of predictive accuracy with an R^2 score of 0. 9648, a Root Mean Square Error (RMSE) of 2. 8823 and a Mean Absolute Error (MAE) of 1. 9053 on the test set. The second phase involves using a hybrid generative model were a fine tuned Large Language Model (LLM) along with XGBoost which enhances the numerical accuracy of the generated text. LLM creates the textual framework of a geopolymer concrete mix design. This output is then refined by substituting its original numeric values with more accurate forecasts from XGBoost model. This generative mix design performed well with a BERTScore of 0. 9754 and a ROUGE L score of 0. 8794 and the numerical prediction of almost all the target features had R^2 value greater than 0. 90. This concludes that the hybrid LLM has the capability to generate semantically and structurally with better numerical accuracy.

AI से पूछें

Bookmark

View Full Paper