Fraunhofer Institute for Mechanics of Materials IWM
Fraunhofer Institute for Mechanics of Materials IWM

Machine Learning

© Fraunhofer IWM

Nowadays, machine learning is used in almost all areas of research and thereby heavily changes the approach to many existing problems. Machine learning therefore has a high potential to create new innovations. This is also true for material science in general and especially for forming processes. As experts in characterizing materials, material modeling and numerical simulation of forming processes, the experts at the Fraunhofer IWM investigate how machine learning can be applied to solve problems concerning materials sciences in the forming industry. We are active in the following areas:

  • Integration of domain knowledge and numerical simulations in machine learning and how to bring this all together in grey-box models
  • Parameter identification for material models using neural networks
  • Softsensors: accessing information about the state of material in components through indirect online-measurements
  • Application of machine learning in production processes and development of intelligent process control
  • Development of methods to map and invert microstructure-property linkages focusing on metallic materials and their usage within industry

Parameter identification via neural networks

Fig. 1.1: Procedure to create a machine learning-based transfer function for material model parameter identification
© Fraunhofer IWM
Fig. 1.1: Procedure to create a machine learning-based transfer function for material model parameter identification

Application

The aim is to identify material model parameters by fitting the material model response to experimental measurements. Machine learning enables setting up a transfer function to directly map experimental measurements on model parameters.

Procedure

(see Fig. 1.1)

  • Creating a physical model that describes the experimental setup. The physical model includes the material model for the parameters that are to be identified.
  • Defining the parameter space within which the model should work
  • Evaluating the physical model for several sets of material model parameters. The data points are distributed in parameter space depending on the sampling strategy that is employed.
  • Setting up an inverse model via machine learning (i.e. neural networks). The input for the machine learning model comes from experimental measurements, the output consists of material parameters.
Fig. 1.2: Experimentally measured stress-strain curves of three metallic materials and the reproduced curves using an elasto-plastic material model and the identified model parameters via only one neural network model
© Fraunhofer IWM
Fig. 1.2: Experimentally measured stress-strain curves of three metallic materials and the reproduced curves using an elasto-plastic material model and the identified model parameters via only one neural network model

Advantages compared to an optimizer-based strategy

The successfully trained machine learning model directly estimates material model parameters. In contrast to optimization methods, no iterative (and time consuming) procedure is necessary.

It is possible to apply the machine learning model on different materials tested as well as to identify model parameters for which the machine learning model is trained within the same experimental setup (see Fig. 1.2).

Sufficient sampling in parameter space leads to a greater likelihood offinding a global solution for the problem of parameter identification. In contrast to optimizers, the machine learning model does not tend to get stuck in local optima, especially when the problem is ambiguous.

to top

Intelligent process control based on digital materials and machine learning

Fig. 2.1: Schematic illustration of a digitalized process chain in sheet metal forming, using an intelligent process control to optimize the internal states of the material
© Fraunhofer IWM
Fig. 2.1: Schematic illustration of a digitalized process chain in sheet metal forming, using an intelligent process control to optimize the internal states of the material

Vision

  • There is a huge potential to optimize complex production processes using process data combined with numerical simulations and machine learning
  • From adaptive process control towards an autonomous production, for example in sheet metal forming (see Fig. 2.1)

Benefit

  • Processes adjust their behavior autonomously to changing conditions (i.e. states of a material, disturbances, wear & tear)
  • Avoiding time-consuming start-up time and enabling sustainable production by directly estimating an optimal set of process parameters
  • Example: during the Fraunhofer lighthouse project Machine Learning for Production we aim to upgrade the glass bending machine developed and owned by the Fraunhofer IWM

to top

Publications

 

  • Heese, R.; Walczak, M.;  Morand L.; Helm D.; Bortz, M., The Good, the Bad and the Ugly: augmenting a black-box model with expert knowledge, Artificial Neural Networks and Machine Learning – ICANN 2019: Workshop and Special Sessions, Lecture Notes in Computer Science Vol. 11731; Tetko, I.V.; Kůrková, V.; Karpov, P.; Theis, F. (Eds.); Springer, Cham (2019) 391-395 Link
  • Morand, L.; Helm, D., A mixture of experts approach to handle ambiguities in parameter identification problems in material modeling, Computational Materials Science 167 (2019) 85-91 Link
  • Morand, L.; Pagenkopf, J.; Helm, D., Material-based process-chain optimization in metal forming, PAMM 17/1 Special Issue: 88th Annual Meeting of the International Association of Applied Mathematics and Mechanics (GAMM); Könke, C.;Trunk, C. (Eds.); Wiley-VCH Verlag & Co. KGaA, Weinheim (2017) 709-710 Link

to top