AI Mix Design of Fly Ash Admixed Concrete Based on Mechanical and Environmental Impact Considerations

Abstract

It has become very important in the field of concrete technology to develop intelligent models to reduce overdependence on laboratory studies prior to concrete infrastructure designs. In order to achieve this, a database representing the global behavior and performance of concrete mixes is collected and prepared for use. In this research work, an extensive literature search was used to collect 112 concrete mixes corresponding to fly ash and binder ratios (FA/B), coarse aggregate and binder ratios (CAg/B), fine aggregate and binder ratios (FAg/B), 28-day concrete compressive strength (Fc28), and the environmental impact point (P) estimated as a life cycle assessment of greenhouse gas emissions from fly ash- and cement-based concrete. Statistical analysis, linear regression (LNR), and artificial intelligence (AI) studies were conducted on the collected database. The material binder ratios were deployed as input variables to predict Fc28 and P as the response variables. From the collected concrete mix data, it was observed that mixes with a higher cement content produce higher compressive strengths and a higher carbon footprint impact compared to mixes with a lower amount of FA. The results of the LNR and AI modeling showed that LNR performed lower than the AI techniques, with an R2(SSE) of 48.1% (26.5) for Fc and 91.2% (7.9) for P. But ANN, with performance indices of 95.5% (9.4) and 99.1% (2.6) for Fc and P, respectively, outclassed EPR with 90.3% (13.9) and 97.7% (4.2) performance indices for Fc and P, respectively. Taylor’s and variance diagrams were also used to study the behavior of the models for Fc28 and P compared to the measured values. The results show that the ANN and EPR models for Fc28 lie within the RMSE envelop of less than 0.5% and a standard deviation of between 15 MPa and 20 MPa, while the coefficient of determination sector lies between 95% and 99% except for LNR, which lies in the region of less than 80%. In the case of the P models, all the predicted models lie within the RMSE envelop of between 0.5% and 1.0%, a coefficient of determination sector of 95% and above, and a standard deviation between 2.0 and 3.0 points of impact. The variance between measured and modeled values shows that ANN has the best distribution, which agrees with the performance accuracy and fits. Lastly, the ANN learning ability was used to develop a mix design tool used to design sustainable concrete Fc28 based on environmental impact considerations. Doi: 10.28991/CEJ-SP2023-09-03 Full Text: PDF