Transfer (machine) learning approaches coupled with target data augmentation to predict the mechanical properties of concrete

Transfer learning, a machine learning technique which employs prior knowledge from solving a source problem to solve a related target problem, is utilized in this work to predict the compressive strength and modulus of elasticity of different concrete mixtures. The use of data augmentation through empirical models to estimate missing data outputs allows for the use of inductive parameter transfer-learning artificial neural network (ANN) models for fast convergence. The paper considers two distinct cases: one where the domain of the target lies somewhat outside that of the source — termed domain expansion, and another where the target output (e.g., elastic modulus) is different, but related to the source output (e.g., compressive strength) — termed domain adaptation. Transfer learning is found to be most accurate when the source dataset is more complex than the target dataset, since more features could be learned. Data augmentation and the coupling of traditional machine learning with transfer learning are demonstrated to greatly enhance the predictive capability for important concrete properties, from mixture proportions. Limited experimental data can be used to transfer-learn the properties (output) of a new dataset from a reliable source model for a related system.