Hybrid PINN–GPR Method for Probabilistic Prediction of Concrete Creep Behavior
TL;DR
A novel hybrid method integrating physics-informed neural network (PINN) and Gaussian process regression (GPR) that enforces the B4 creep model as a physics-informed constraint by embedding its governing equations into the loss function, effectively incorporating physical knowledge into data-driven training.
Abstract
Concrete creep, a critical phenomenon influencing long-term structural deformation and durability, poses significant modeling challenges due to its complex, time-dependent nonlinear behavior. Traditional empirical models and numerical simulations often struggle to balance physical interpretability with uncertainty quantification. To address this issue, this study proposes a novel hybrid method integrating physics-informed neural network (PINN) and Gaussian process regression (GPR). The PINN component enforces the B4 creep model as a physics-informed constraint by embedding its governing equations into the loss function, effectively incorporating physical knowledge into data-driven training. Simultaneously, the GPR module provides Bayesian uncertainty quantification for creep predictions and generates a 95% credible interval to characterize prediction reliability. The proposed PINN method is validated against the Northwestern University (NU) creep database and demonstrates 37.08%–76.93% lower RMSE compared to both the conventional model-driven B4 method and three data-driven methods: extreme gradient boosting (XGBoost), genetic algorithm-back propagation neural network (GA-BPNN), and convolutional neural network (CNN). Laboratory experiments further confirm the reliability of the proposed PINN-GPR method, as its 95% credible intervals accurately capture measured creep deformations across varying loading ages, concrete mix proportions, and environmental conditions. This demonstrates robust prediction accuracy and uncertainty characterization for creep analysis.