Menü des aktuellen Bereichs:




Gravels, corn seeds, pharmaceutical powders, sands and ores can not be easily characterized as for instance a steel beam.

A set of experimental and numerical solutions, together with artificial neural networks, can improve the accuracy and the range of applicability of the characterization of particles properties, and reduce the computational costs.

After the data collection (thanks to a set of characterization devices) physical key parameters (bulk-macro behavior) of particles are identified. Meanwhile, combinations of DEM-microscopic parameters have been simulated, each leading to a different numerical bulkmacroscopic behavior, through DEM (LIGGGHTS) and CFDEM (CFDEMcoupling).

The predictive capability of any simulation strongly depends on the validity of the particle based simulation parameters. To ensure it is not necessary to evaluate a huge number of parameter sets, but rather to evaluate the sensitivity of the bulk behavior with respect to individual particle based parameters. This can be realized efficiently by Artificial Neural Networks (NN)! Inside the NN neurons are linked to particle based input parameters. By matching the output of the artificial neural network to DEM simulation results the network is trained (i.e. individual neurons are weighted), with excellent regression results (Fig.3).

Fig. 3: Artificial neural network

Later, the trained neural network can sets of particle based simulation parameters: we compare the macrobulk behavior of these sets against the experimental data collected, gaining averages and validity range (Fig. 5). In fact, we obtain valuable information about the dependence of bulk solid behavior on individual particle properties and eventually mutual dependencies (Fig. 4).

Fig. 4: Regression plot

Fig. 5: Cloud plot

Fig. 6: Radar plot

(Luca Benvenuti)