Stochastic data-driven inference of mesoscale lattice discrete particle model parameters via multiscale observations

Concrete failure mechanics exhibit significant variability at the macroscopic scale, which is predominantly driven by stochasticity at the spatial scale of the coarse aggregate particles, generally referred to as mesoscopic scale. However, mesoscale material parameters are difficult to estimate, making uncertainty quantification a fundamental challenge. To address this limitation, a data-driven multiscale inverse inference framework is proposed to quantify the stochastic mesoscale behavior by integrating both mesoscale and macroscale observations. In this framework, a stochastic data-driven model using a hybrid Proper Orthogonal Decomposition–Gaussian Process Regression (POD-GPR) algorithm is first developed based on data generated by mesoscale Lattice Discrete Particle Model (LDPM) simulations. Leveraging this efficient data-driven model, a novel multiscale Bayesian inverse inference method is proposed to infer the stochastic distributions of the mesoscale features. When applied to experimental data, the proposed framework successfully captures the stochastic distributions of mesoscale material parameters, reproduces macroscale responses, and outperforms conventional single-scale Bayesian inference approaches. Additionally, SHapley Additive exPlanations (SHAP) are integrated to further interpret the effect of mesoscale stochastic material behavior on macroscale uncertainty, offering valuable insights for the accuracy improvement of LDPM simulations and future mesoscale-level optimization to achieve more robust macroscale performance.