Abstract Extrusion-based 3D Concrete Printing (3DCP) offers transformative potential for the construction industry, yet its widespread adoption is hindered by mechanical anisotropy and the lack of tools to accurately predict material performance in design. This research addresses this critical gap by developing and validating a novel Physics-Guided Neural Network (PGNN) framework capable of producing reliable and interpretable anisotropic strength predictions. The PGNN combines a physics-based branch, which models compressive strength as a function of age and water-to-binder ratio (W/B) via a modified Abram’s Law, with a shallow neural network branch that captures residual nonlinear effects from loading orientation and mix design. Training and validation were performed on the largest and most comprehensive international dataset available, generated through the RILEM TC-304 ADC interlaboratory study. The PGNN demonstrated high predictive accuracy on unseen RILEM test datasets (R 2 = 0.86–0.89). In addition, the model was applied to six external datasets in a blind test which confirmed reasonable generalisation for typical printing mortars within the compositional range of the training dataset, but was less conclusive for atypical mix designs such as ultra-high-performance and steel-fibre-reinforced composites. The framework yields a transparent formulation, making the relationship between variables and predicted strength directly understandable and allows the application of response surface analyses to gain insight into the influence of mix design variables such as aggregate size, aggregate-to-binder ratio, and W/B on anisotropy. The presented work provides a positive outlook for the transformative potential of physics-guided machine learning in 3D concrete printing to underpin reliable deployment of 3DCP technologies.
Jamhiri et al. (Mon,) studied this question.