Bayesian stochastic multi-scale analysis via energy considerations

TL;DR

This paper introduces a Bayesian multi-scale analysis framework that couples physical processes across scales through energy considerations, enabling probabilistic modeling of material behavior from meso to macro levels, including nonlinear regimes.

Contribution

It develops a hierarchical Bayesian approach for multi-scale modeling that incorporates energy matching and probabilistic upscaling, including model reduction techniques for efficiency.

Findings

Probabilistic macro-scale models reflect meso-scale uncertainty.

Energy-based coupling effectively captures irreversible behaviors.

Model reduction reduces computational cost significantly.

Abstract

In this paper physical multi-scale processes governed by their own principles for evolution or equilibrium on each scale are coupled by matching the stored and dissipated energy, in line with the Hill-Mandel principle. In our view the correct representations of stored and dissipated energy is essential to the representation irreversible material behaviour, and this matching is also used for upscaling. The small scales, here the meso-scale, is assumed to be described probabilistically, and so on the macroscale also a probabilistic model is identified in a Bayesian setting, reflecting the randomness of the meso-scale, the loss of resolution due to upscaling, and the uncertainty involved in the Bayesian process. In this way multi-scale processes become hierarchical systems in which the information is transferred across the scales by Bayesian identification on coarser levels. The quantities to be matched on the coarse-scale model are the stored and dissipated energies. In this way probability distributions of macro-scale material parameters are determined, and not only in the elastic region, but also for the irreversible and highly nonlinear elasto-damage regimes, refelcting the aleatory uncetainty at the meso-scale level. For this purpose high dimensional meso-scale stochastic simulations in a non-intrusive functional approximation forms are mapped to the macro-scale models in an approximative manner by employing a generalised version of the Kalman filter. To reduce the overall computational cost, a model reduction of the meso-scale simulation is achieved by combining the unsupervised learning techniques based on the Bayesian copula variartional inference with the classical functional approximation forms from the field of uncertainty quantification.