Adaptive Multiscale Modelling Methodology

Highlights

  • In multiscale modelling empirical, simplified models of material properties are replaced by more sophisticated fine scale models
  • Multiscale Modelling provides high predictive capabilities, but they have extremally large computational requirements
  • High reliability and high accuracy are not always required
  • The most of sophisticated and complex fine scale models have simpler surrogates
  • The Adaptive Multiscale Modelling Methodology (AM3) allows to balance between required quality of modelling results and available compuational resources to meet the best results in practical cases

The aim of Adaptive Multiscale Modelling Methodology

Tu powinny być ładne obrazki:)

Multiscale modelling offers high predictive capabilities, but requires extremally large computational resources. There are several approaches to minimize these needs, however most of them are far from optimal and require high skills in numerical modelling and material science. Usually, development of a satisfying multiscale model requires long lasting try-and-error activities. That makes multiscale modelling very difficult tool, especially if expected to be applied in industrial environment. On the other hand, modern and competitive market of metal goods is focused on quick development of new technologies and materials. The time of development could be decreased with multiscale modelling (e.g. with the Integrated Computational Material Engineering approach) – but that requires ability of multiscale modelling.

The Adaptive Multiscale Modelling Methodology (AM3) is a design methodology for structuring complex and multiscale models. It aims to reduce the programming effort required to develop fully coupled models, and limit the computing resources needed by the model. AM3 framework, designed for supporting development of multiscale models simplifies integration of numerous macro-, meso- and microscale submodels, including commercial and in-house solutions. Furthermore, integrated Knowledge Based System controls the execution of a simulation of autonomously choose the best configuration of submodels, balancing available computational resources and required reliability. With AM3, development of multiscale models is faster and easier. From the user’s perspective, AM3 allows to control computational time and reliability of the model easier and more efficient than classical approachs.

Scientific background

The Adaptive Multiscale Modelling Methodology (AM3) is a design methodology for structuring complex multiscale models. It aims to reduce the programming effort required to develop fully coupled models, and limit the computing resources needed by the model. the multiscale model consists in a macroscopic module linked to several lower-scale modules. The state of the model is continuously tracked by a Knowledge Base System and relayed to a control unit that decides on the run whether to use a lower scale model, and which one to use, in order to provide the best compromise between reliability and computing time. The decisions made by the control unit are based on a set of user defined rules that provide, amongst others, the range of applicability of each fine-scale model.

Fine scale models classically feature local variables, averaged over a Representative Volume Element (RVE) or a Statistical Volume Element (SVE) [Rauch2014, Rauch2015]. In a coarse domain, an instance of the fine scale model needs to be ran for each computational point. If the fine scale model is computationally demanding, the overall simulation will require large computational resources and/or a long computational time, even in a parallel or a cloud environment. Instead, the AM3 framework introduces adaptability, meaning that the structure of the multiscale model can adapt to the actual conditions — phenomena expected to be present in a process, required reliability and available computing power. Fine scale models are used only when and where they are needed. In the original AM3 approach, fine scale models could just be turned on or off. In presented research, several models may be available for the same variables, differing mainly by the computational costs and the reliability but also by the range of applicability. The control unit can choose the most relevant one. For example, there could be two models of recrystallization, one based on Internal Variable (IV) approach and a second based on Phase Field method. The first is much less computationally demanding, however reliability of the IV-based model worsens rapidly when process conditions become different from the ones for which the IV model had been calibrated. Hence, in some range of process parameters both models provides the same reliability. In other range, the IV model is clearly non-applicable. In both ranges an optimal choice of the model is clear — IV-based in the first one and Phase Field in the second. However, there is still a range when both  models may be used and their relevancy depends on process parameters but also required reliability and available computing power. In the present state of the AM3 framework, the mentioned above KBS system is able to decide is fine scale model necessary in particular computational point and if yes, which fine scale model is most relevant in particular conditions and requirements.

Contact person:

PhD. Piotr Macioł, pmaciol(at)agh.edu.pl

Connected scientific papers:

Connected projects:

  • Developmentof a methodology for “agile” multiscale modelling , Polish National Science Centre, Sonata, 2012-2016
  • Application of the Knowledge Based Systems for controlling of uncertainity in optimization of metal forming, Polish National Science Centre, Opus, 2015-2018