Identification of the relations between synthesis conditions and compressive strength in geopolymer systems via statistical learning methods

Abstract

Geopolymers are inorganic amorphous materials, which can be used in construction industry, and their production is characterized by low CO2 emission. Compressive strength (CS) is a durability measure, which, in the case of geopolymers, depends on various manufacturing factors in a rather complex functional relation. The aim of this study is to develop a decision tree and extract reliable decision rules based on it. A database consisting of 879 data points and 19 predictors is constructed based on available publications. Forward Selection (FS), Backward Elimination (BE) and Rough Set theory are used as feature selection algorithms. The decision tree parameters (maximum number of splits, Smax and minimum parent size Pmin) and number of relevant predictors for FS and BE and threshold parameter (eps), which is introduced for reduct set formation, are optimized via 5-fold CV with 20 repetitions. The reliability of the obtained decision rules is expressed via Stability index (SI) whereas the accuracy of the models via measures, such as RMSE. Minimum test RMSE is decreased up to 10.0 MPa for Smax = 150 BE based tree, but the interpretation of such large trees is demanding. Hence, a smaller sized (Smax = 60) reduct based tree is used for rule extraction. The CV RMSE and SI, test RMSE and R2 of the resulting tree is 13.2 MPa, 0.79, 13.2 MPa and 0.64 respectively. Based on this tree, the most important predictor affecting geopolymer CS is found to be the type of aluminosilicate precursor material. Two material groups are identified corresponding to higher and lower CS. The data space is subsequently divided, and relations (linear regression, local models etc.) regarding smaller subspaces are developed. For higher CS group, water/Al, Na/Al and ambient curing time have a significant effect, whereas for lower CS compounds, purity and heat curing time are also found to be effective. Moreover, the effect of the predictors on CS can vary based on the precursor types, showing the presence of significant predictor interactions.