AI-Enabled Digital Twin for Crack Evolution in Concrete Beams

Crack formation and propagation in concrete members remain major concerns in structural health monitoring because conventional inspection is periodic, manual, and mostly reactive. In practice, engineers often identify visible damage only after cracks have already developed to a worrying level, while the internal stress state and the likely direction of future deterioration remain difficult to interpret in real time. This project addresses that gap by developing an AI-enabled digital twin for crack evolution in concrete beams. The main goal is to build a computationally efficient monitoring environment that preserves physical relevance while reducing the latency associated with traditional high-cost simulations. Instead of relying on manual observation alone or on purely data-driven prediction, the work combines a physics-informed neural network with an interactive Unity 3D visualization environment so that structural response can be inspected in a way that is both faster and easier to understand.

The proposed workflow is organized into connected layers. At the physical layer, the concrete beam is observed through sensors and camera inputs. At the perception and data layer, captured frames and structural parameters are preprocessed, crack-related features are extracted, and the data are packaged for downstream inference. The digital twin then evaluates a PINN model using key inputs such as spatial coordinates, applied load magnitude, global deflection, concrete compressive strength, and steel yield strength. The model produces stress, strain, and a damage index, which are then mapped back into the Unity environment. In the current workflow, the beam model is segmented into 100 sections, three representative locations are evaluated, and the predicted response is distributed visually as a heat map. Regions whose damage index exceeds the threshold value of 0.21 are marked as probable crack locations, enabling intuitive inspection and quicker interpretation by the user.

Experimental evaluation focuses on comparing the digital twin output against Abaqus finite element simulations so that the visual and analytical behaviour of the system can be checked against a trusted engineering baseline. The project demonstrates that a physics-guided AI pipeline can support visual crack monitoring while remaining more responsive than a simulation-only loop. It also shows the practical value of linking AI inference with a digital twin interface for structural health monitoring education, analysis, and future field deployment. At the same time, the team identified several important limitations: future crack propagation prediction is not yet integrated into the deployed twin, the trained LSTM module is still disconnected from the Unity environment, and automatic crack width and crack length measurement are not yet implemented. Even with those limitations, the project reaches a meaningful conclusion: it establishes a credible foundation for real-time AI-assisted damage visualization in concrete structures and creates a strong path toward future predictive monitoring, richer datasets, and a more complete digital twin for civil infrastructure.