Prediction of natural carbonation depths in concretes with ensemble metamodel based on artificial neural networks from time series analysis with 20 years of exposure

Abstract

Considering the growing use of carbonation depth prediction models based on artificial neural networks, the ensemble architecture stands out due to its ability to combine different predictive models into a single metamodel, increasing the accuracy of predictions. However, applying these cybernetic models requires greater rigor on the completeness and robustness of the databases employed in the training and validation phases of neural networks. Treating the carbonation depth databases as time series can be a favorable strategy to guarantee completeness and robustness. Thus, this article aims to predict the carbonation depths of concrete structures using an AVR-SARIMA-LSTM-MLP ensemble metamodel with hybrid architecture for neural networks associated with time series analysis. The metamodel was based on several individual SARIMA-LSTM-MLP predictor models trained and validated with information from 36 concretes with different water/binder ratios (0.40, 0.55, and 0.77), types of mineral additions (rice husk ash, fly ash, blast furnace slag, metakaolin, silica fume, and reference – no mineral addition), and curing conditions (wet and dry). The concrete database was made available by the GEDur group and has 2313 depths of natural carbonation measured over 20 years of exposure in a controlled environment. The results of the AVR-SARIMA-LSTM-MLP ensemble metamodel predicted values for about 67 years after the concrete was produced, recording an average correlation coefficient of 0.93 and RMSE between 0.05 and 4.69 mm. These results demonstrate that the ensemble predictor metamodel has high predictive capacity, excellent precision, and accuracy, regardless of the characteristics and properties of the concretes, curing, and exposure conditions.