Machine Learning Models with a GUI for Predicting Compressive Strength of Nano-Modified Concrete Exposed to High Temperatures

Nanoparticle-modified concrete can exhibit improved mechanical performance, yet its residual compressive strength (Fc) after fire-like thermal exposure is difficult to predict because the response depends on both mixture design and heating conditions. Building on recent advances in explainable machine learning (ML) for cementitious materials, this study compiles 218 literature datapoints of post-heating Fc from 100 mm concrete cubes incorporating carbon nanotubes (CNTs) and nano-alumina (NA), exposed to 20–800 °C for up to 2 h. Seven input variables are used: cement-to-total aggregate ratio, CNT-to-cement ratio, NA-to-cement ratio, coarse-to-fine aggregate ratio, water-to-cement ratio, peak temperature, and exposure duration at temperature. Two particle-swarm-optimized ensemble regression models, Extreme Gradient Boosting (XGB-PSO) and Random Forest (RF-PSO), were developed and evaluated using a 70/30 train–test split with K-fold cross-validation on the training set. SHAP, individual conditional expectation (ICE), and partial dependence plots (PDPs) were employed to study the individual and combined effects of each input parameter on Fc prediction. The results demonstrated that the XGB-PSO model provides the best predictive performance (training R2 = 0.9983; testing R2 = 0.9434; testing MAE = 1.3168 MPa). Model interpretability was assessed using SHAP, ICE, and PDP analyses, revealing that temperature and exposure duration dominate strength loss, while CNTs and NA contribute positively within dose-dependent regimes. The highest predicted strengths occur for CNTs of 0.05% to 0.15% and NA of 0.65 to 2.71% (by cement mass) under moderate temperature exposure. A Python-based graphical user interface is provided to support rapid what-if assessment of CNT–NA mixtures under elevated-temperature scenarios.