A new paradigm for the efficient inclusion of stochasticity in engineering simulations

TL;DR

This paper introduces a novel method that incorporates stochasticity into engineering simulations using extended material models, enabling accurate uncertainty quantification with only a single, computationally efficient simulation.

Contribution

The proposed approach allows inclusion of randomness in simulations through extended deterministic models, reducing computational cost from multiple runs to just one.

Findings

Accurately estimates expectation and variance of internal variables.

Reduces computational effort compared to Monte Carlo simulations.

Demonstrated on three complex, non-linear material models.

Abstract

As a physical fact, randomness is an inherent and ineliminable aspect in all physical measurements and engineering production. As a consequence, material parameters, serving as input data, are only known in a stochastic sense and thus, also output parameters, e.g., stresses, fluctuate. For the estimation of those fluctuations it is imperative to incoporate randomness into engineering simulations. Unfortunately, incorporating uncertain parameters into the modeling and simulation of inelastic materials is often computationally expensive, as many individual simulations may have to be performed. The promise of the proposed method is simple: using extended material models to include stochasticity reduces the number of needed simulations to one. This single computation is cheap, i.e., it has a comparable numerical effort as a single standard simulation. The extended material models are easily derived from standard deterministic material models and account for the effect of uncertainty by an extended set of deterministic material parameters. The time-dependent and stochastic material behavior are separated, such that only the deterministic time-dependent behavior of the extended material model needs to be simulated. The effect of stochasticity is then included during post-processing. The feasibility of this approach is demonstrated for three different and highly non-linear material models: viscous damage, viscous phase transformations and elasto-viscoplasticity. A comparison to the Monte Carlo method showcases that the method is indeed able to provide reliable estimates of the expectation and variance of internal variables and stress at a minimal fraction of the computation cost.