Physics-informed deep learning and finite element integration for mesoscale impact response of concrete targets

Designing resilient and durable protective structures is reliant on the understanding of concrete's behaviour when impacted by high-velocity projectiles. In this study, we proposed an integrated computational framework of three-dimensional finite element simulations and physics-informed deep learning to evaluate the penetration resistance of concrete targets when impacted at high-velocities. We developed a validated mesoscale concrete model by including randomly distributed spherical aggregates and the concrete damaged plasticity formulation and simulated 253 impact scenarios using a robust velocity range of 100–900 m/s. Energy metrics such as kinetic energy (KE), internal energy (IE), and energy dissipated as a result of damage to the specimen were extracted to study the way an impact loads a specimen transfers and absorbs energy. By developing a physics-consistent neural network model that incorporated an energy conservation constraint, we improved our models predictive performance through relating residual velocity to both internal energy and kinetic energy output via the use of experimental and numerical reference data. The model prediction with coefficients of determination (R²) of 0.9781 for KE and 0.9503 for IE across the PINN prediction dataset, while a correlation coefficient of 0.999 was achieved for residual velocity during benchmark validation against experimental and numerical reference data. A Bayesian extension was added to, quantify epistemic uncertainty, revealing that specimens with interfacial transition zone (ITZ) strength of 15–25 MPa exhibited high energy absorption capacity and low variability in their predictions when the aggregate volume fraction contained was between 0.25–0.35. Sensitivity analysis identified critical meso‑structural parameters such as aggregate size, aggregate distribution uniformity, and ITZ toughness which directly influence an impact's energy dissipative capacity and, therefore, its resilience to failures due to impacts. This proposed framework will provide a robust, interpretable and uncertainty-aware mechanism for assessing and optimizing the impact performance of heterogeneous concrete systems, thereby aiding engineers in the design of next-generation protective infrastructure in an empirically supported manner that is data-driven and risk-informed.