Back to overview

Lecture

WEB Data-driven prediction of structure-property relationships using bonded-particle model and artificial neural networks



Determination of mechanical properties of materials based on their known microstructure plays an important role in different areas of process engineering and material science. On the one hand, knowledge of such structure-property relationships allows one to predict the macro-properties of materials. On the other hand, this makes it possible to solve inverse problems – for desired product properties find an optimal microstructure. In this contribution, we present a simulation framework based on a combination of bonded-particle model (BPM) and artificial neural networks (ANN), which can be used to obtain desired dependencies.

The bonded-particle model is an extension of the discrete element method, where the modeled material is represented as a set of smaller primary particles (nodes) connected by solid bonds. Each particle and bond have unique geometric and material parameters. Thus, the application of this meshfree approach leads to high flexibility during generation of structural model and allows one to define complex structured materials, such as highly porous aerogels [1], multicomponent agglomerates [2], composite materials, etc.

In this contribution, the BPM was used to perform microscale simulation of the mechanical behavior of multicomponent agglomerates under uniaxial compression tests. Here, the microstructure of the material, as well as the properties of individual components, have been varied. Based on the simulation results, a database containing macroscopic properties of agglomerates, such as stiffness, strength, and breakage energy, has been generated. This database was further used for training the artificial neural network. Finally, the obtained ANN was employed to predict structure-property relationships.

The BPM method has been implemented in the simulation framework MUSEN, and the calculations were parallelized for GPU architectures. In the same time, the pre-processing algorithms for generation of initial structures were optimized for multicore systems. This made it possible to efficiently utilize all the computing resources of hybrid CPU-GPU architectures.


[1] Dosta M., Jarolin K., Gurikov P. (2019). Modelling of mechanical behavior of biopolymer alginate aerogels using the bonded-particle model, Molecules 24.


[2] Spettl A., Dosta M., Antonyuk S., Heinrich S., Schmidt V. (2015) Statistical investigation of agglomerate breakage based on combined stochastic microstructure modeling and DEM simulations, Adv. Powd. Techn. 26.

 

Speaker:
Prof. Maksym Dosta
Hamburg University of Technology
Additional Authors:
  • Vasyl Skorych
    Hamburg University of Technology
  • Tsz Tung Chan
    Hamburg University of Technology