Machine learning and RSM models for prediction of compressive strength of smart bio-concrete
In recent years, bacteria-based self-healing concrete has been widely exploited to improve the compressive strength of concrete using different bacterial species. However, both the identification of the optimal involved reaction parameters and theoretical framework information are still limited. In...
Saved in:
| Main Authors: | , , , , , , , , |
|---|---|
| Format: | Article |
| Published: |
Techno-Press
2021
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/96740/ http://dx.doi.org/10.12989/sss.2021.28.4.535 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In recent years, bacteria-based self-healing concrete has been widely exploited to improve the compressive strength of concrete using different bacterial species. However, both the identification of the optimal involved reaction parameters and theoretical framework information are still limited. In the present study, both experimentally and numerical modelling using machine learning (ANN and ANFIS) and response surface methodology (RSM) were implemented to evaluate and optimse the evolution of bacterial concrete strength. Therefore, a total of 58 compressive strength tests of the concrete incorporating new bacterial species were designed using different concentrations of urea, cells concentration, calcium, nutrient and time. Based on the results, the compressive strength of the bacterial concrete improved by 16% due to the decrement of the pore percentage in the concrete skin; specifically, 5 mm from the concrete surface, compared to that of the control concrete. In the same context, both machine the learning and RSM models indicated that the optimal range of urea, calcium, nutrient and bacterial cells were (18-23 g/L), (150-350 mM), (1-3 g/L) and 2×107 cells/mL, respectively. Based on the statistical analysis, RMSE, R2, MPE, RAE and RRSE were (0.793, 0.785), (0.985, 0.986), (1.508, 1.1), (0.11, 0.09) and (0.121, 0.12) from both the ANN and ANFIS models, respectively, while; the following values (0.839, 0.972, 1.678, 0.131 and 0.165) was obtained from RSM model, respectively. As such, it can be concluded that a high correlation and minimum error were obtained, however, machine learning models provided more accurate results compared to that of the RSM model. |
|---|