(113e) Data Driven Method Using Powder Characterisation Tools to Calibrate DEM Simulations

Powder and particle characterisation tools have been essential to improving our understanding of powders since the 1700s. They have become much more advanced and automated in the 21st century but the way measurements are used remains the same, as a number to compare between powders. To take powder and particle characterisation tools to the next level, the measurements in the lab need to be able to directly inform decisions and solve problems in industrial equipment.

Powder characterisation tools are already advanced enough to be able to make a variety of accurate measurements of bulk powder properties such as the angle of repose. The discrete element method (DEM) has started to allow powders to be modelled on large scales in industrial equipment, improving our ability to design powder handling equipment and solve issues. However, to calibrate DEM simulations the microscopic properties of the powder being simulated are required. A method is needed to calculate the microscopic properties from the bulk measurements made using powder characterisation tools.

Finding the relationship between the microscopic properties and measurable bulk macro measurements, a micro-macro relationship mapping, would allow powder characterisation tools to measure the properties of the particles that make up the powder. This would also provide a better calibration method than currently used methods which can be inaccurate and time intensive to conduct.

Method

Bulk measurements from 4 powder characterisation tools: GranuDrum, GranuHeap, GranuFlow and GranuPack will be linked to the microscopic properties: restitution, sliding friction, rolling friction and cohesive energy density, using a data driven method. To generate the data required, DEM simulations of each powder characterisation tool (digital twins) were developed. Thousands of simulations of each digital twin were then conducted on supercomputers and the bulk measurements recorded at various different values of microscopic properties.

To develop the micro-macro relationship, the typically data-driven approach today would be to train a neural-network. However, this has various problems including poor extrapolation and the fact that neural networks could fit anything to anything. Instead, an exploratory equation approach is being taken. This is similar to fitting a simple second order model to a dataset but instead of fitting just the coefficients you also generate a unique equation structure. This has the advantage of having better extrapolation due to trying to fit a physical equation to the data and reduces the black box nature that neural-networks have as you end up with a real equation you can manipulate afterwards.

Results

Apart from the novel calibration approach developed as part of the project that allows the calculation of microscopic properties of particles from quick powder characterisation tools experiments, digital twins have been developed and sensitivity analysis have been conducted. The top half of the figure below shows the digital twins that have been developed. From left to right is the GranuHeap, GranuPack, GranuFlow and GranuDrum.

Using the data generated for the data-driven calibration method there are other analysis techniques that can be used to provide further insight into powder characterisation. For example, high-dimensional model representation sensitivity analysis was used to determine how influential each microscopic particle property is on the bulk measurements made by powder characterisation tools. The results are shown in the lower half of the figure below.