Transfer learning framework for modelling the compressive strength of ultra-high performance geopolymer concrete

Ultra-high performance geopolymer concrete (UHPGC) presents a sustainable alternative to traditional ultrahigh performance concrete (UHPC). Accurate prediction of UHPGC compressive strength is crucial for its wider adoption in both industry and academia. However, the complex interplay of factors influencing UHPGC compressive strength, coupled with limited available data, makes this task challenging. Therefore, this study aims to establish a transfer learning (TL) framework for UHPGC compressive strength prediction. This work explored the ability of TL to leverage the knowledge acquired from normal-strength geopolymer concrete datasets to develop models for predicting the compressive strength of UHPGC. The results demonstrate that TL models (transferred Convolutional Neural Network (transferred CNN), transferred Tabnet, and Two-stage TrAdaboost.R2) outperform traditional machine learning (ML) models (CNN, Tabnet, Adaboost.R2), with the corresponding R-square score of 0.93, 0.93, and 0.94 (for the TL models) and 0.86, 0.89, and 0.90 (for the traditional ML models). Additionally, TL models exhibit 10%-30% lower RMSE than their traditional counterparts. Furthermore, the findings indicate that a minimum of 40 data samples in the target domain is necessary for reliable predictions. In conclusion, the study captures the effectiveness of the TL approach in overcoming data scarcity and the robustness of TL models to variations of features in the target domain. This research highlights the potential of knowledge transfer from well-researched geopolymers to develop predictive models for specialised geopolymer types with limited data.