The Solution

CoMModO is intended to introduce some new approaches to the modeling of complex materials, enabling to overcome the complexity trap for new high tech materials need in advanced light weight designs.

For the solution it is proposed to use and combine the following technologies as alternatives accompanying the traditional approaches:

  • use of Soft Computing methods for improved flexibility and functionality,
  • interpretation as a time series prediction and classification problem,
  • incorporating a requirement driven process model inspired from the field of software engineering,

Soft Computing Methods

One of the core components within CoMModO is the use of Soft Computing methods. Especially Artificial Neural Networks and Evolutionary Strategies are emphasized here.

Artificial Neural Networks and corresponding Machine Learning models posess a maximum of functionality. They allow the straightforward treatment of fuzzy and uncertain parameters as well as categorical variables and can therefore be used to train material models to behave like the targeted material tests.
Nevertheless they need to be applied with great care, which is the reason why they need to be combined with appropriate process models.

Evolutionary strategies are flexible and powerful tools for global optimization. Within CoMModO they are not only used to tune conventional material models but also for the training of machine learning models. A novelty is the derivation of the fitness function based on requirement driven process schemes.

Interpretation as Time Series Problem

One of the major counter-arguments to Artificial Neural Networks in the field of material models is the absence or the cost of sufficient training data. The interpretation as a time series classification or prediction problem allows to overcome this issue.

Many tests for the mechanical behavior of materials result in force vs. deflection or stress vs.strain curves that pertain to very restricted conditions (e.g., only uniaxial loading). Typically some characteristic values of the curve, e.g., the maximum force or the dissipated energy, are extracted and used to tune a few parameters of the analytical material model. However, this way only a part of the information is retrieved from the test. Much more data is contained in such tests, which can be used for training when interpreting the complete time series of the force-deflection curves.

Software Process Models

Development process models from the field of software engineering nearly always tell the same: begin with the requirements as well as the concept design and start coding afterwards and not at the beginning! Translating this to the field of material modeling implies the following: first think about the load cases and conditions the material model needs to fullfill and do not start with the preconception of the material model. Design the tests according to the requirements (necessary load conditions) and not according to the preassumed material model (the hypothesis). The interpretation as time series problem in combination with machine learning approches allows this approach. The material model should then follow the required tests and not vice versa (sounds trivial and self-evident but isn't in practice, believe it or just be honest to yourself...).

The process models provide guidelines for a requirement driven process instead of a preconceived solution driven one.