Application of Machine Learning on the Uncertainty Quantification of Fresh Concrete Properties and Reliability-Based Design Optimization of the 3D Printing Process
Please login to view abstract download link
This study aims at developing an efficient method that allows characterizing and quantifying the uncertainties of concrete properties as well as their consideration on the design optimization of the 3D printing process. The well-known Bayesian inference is chosen to characterize the uncertainties of the elastic and plastic properties of fresh concrete using the results of direct tests (e.g., uniaxial compression and direct shear) in laboratory. These characterized mechanical properties and their associated uncertainty are then taken as input parameters for the stability analysis of concrete structure during printing. On the one hand, the Gaussian process (i.e., Kriging metamodeling technique) is chosen as the surrogate to estimate the failure probability of structure due to elastic buckling and plastic collapse. On the other hand, the combined Kriging surrogate with the quantile-based optimization approach is conducted to determine the optimal solution of the printing parameters (e.g., print speed, size of deposited layer). Finally, to overcome the intrinsic limit of the direct tests which are quite heavy and long, the simple and rapid indentation test can be considered as a highly potential method to characterize concrete properties. In this work, using the Artificial Neural Network and synthetic data provided from the numerical simulation of indentation test, the applicability of this method to determine the evolution in time of elastic and plastic properties of concrete at fresh state will be highlighted.
