Preliminary investigation of prompt-based synthetic data generation using large language models for machine learning applications in biochar-modified concrete

Biochar-modified concrete has emerged as a promising material to mitigate carbon emissions. However, limited availability of published experimental datasets (PED), necessitates predictive models. In this study, PED were collected from Web of Science database. Synthetic datasets (SDs) were generated using a large language model (LLM) based on the descriptive statistics of PED training data. Subsequently, SDs were augmented with PED to investigate their influence on predictive modelling. Baseline models were developed using linear and nonlinear regressors, and extreme gradient boosting (XGBoost) with Optuna-based hyperparameter optimization was adopted for reliability and augmentation analysis. Exploratory data analysis revealed the inability of SDs to capture multivariate structure of PED. For PED-only models, test R 2 of 0.911 identified XGBoost as the best-performing model. Despite high training and validation scores, reliability models showed average R² performance of 0.513 on PED test set. Additionally, augmented models failed to outperform PED-only models, despite outperforming reliability models. Effect of augmentation size also revealed that small to moderate SD additions outperformed larger SD additions, although evaluation procedure has an effect on ranking. These findings suggests that LLM-generated SDs can be used for preliminary exploration and controlled augmentation, but should not be used in replacing experimental data without rigorous validation.