Resilient maintenance of carbonation-affected concrete infrastructure via physics-informed learning and predictive strategy

Climate change accelerates carbonation in concrete, raising risks of cracking and spalling. However, existing model formulations often oversimplify this impact, inadequately representing the intrinsically non-linear and phase-dependent behavior of the carbonation process. In this paper, we propose a novel framework that integrates Physics-Informed Neural Networks (PINNs) and a Predictive Markov Decision Process with Phase-type Approximation (PMDP-PH), enabling resilient and cost-effective infrastructure maintenance under carbonation risks. PINNs embed governing physical laws within neural architectures, enabling accurate inference of carbonation dynamics even under limited observational data. This framework accommodates non-exponential sojourn time distributions in both the initiation and propagation phases, effectively approximated using hypo-exponential models. The PMDP-PH adaptively updates inspection and maintenance strategies by continuously refining the remaining useful life (RUL) distribution in real time. This decision-making process is formulated as a tractable multi-stage model predictive control (MPC) problem over selected belief states, ensuring a non-decreasing value function throughout the robust horizon. Applied to a representative infrastructure system, our method reduces total maintenance costs by up to 61.9% compared to benchmark strategies under variable deterioration scenarios. These findings highlight the promise of combining physics-informed learning with a new form of predictive control strategy to strengthen infrastructure resilience under the growing threat of carbonation-induced deterioration.