(144i) Access: Autonomous Characterisation and Calibration Using Evolutionary Simulation Software

The discrete element method (DEM) is a powerful simulation technique that is capable of numerically modelling the behaviour of complex granular media, being used to better understand and optimise the internal dynamics of a large number of systems in both academic fields and industrial sectors, from fundamental research into contact mechanics to improving plant-scale reactors [1]. DEM can offer exceptional accuracy through its lack of approximations over meshes and, if correctly calibrated, simulations can provide results with quantitative precision. It is this "if", however, that also represents DEM's biggest drawback: without choosing appropriate contact models and carefully calibrating multiple DEM parameters, the simulation outputs simply cannot be trusted. This calibration is a time-consuming process, typically involving the measurement of diverse particle properties including size, density, restitution and friction coefficients and, for purely "virtual" parameters such as the cohesive energy density, a great deal of experimentation [2].

To automate DEM calibration against experimental measurements, we have developed ACCESS – Autonomous Characterisation and Calibration using Evolutionary Simulation Software. ACCESS enables a researcher to calibrate virtually any DEM parameters against a user-defined cost function, quantifying and subsequently minimising the disparity between the simulated system and the experimental reality using state-of-the-art evolutionary strategies – in essence, autonomously ‘learning’ the physical properties of the particles within the system, without the need for human input. This cost function is completely general, allowing ACCESS to calibrate DEM against measurements as simple as photographed occupancy plots, or complex system properties captured through e.g. Lagrangian particle tracking. In our tests, ACCESS successfully calibrated up to 12 unknown particle parameters in a matter of hours against the free surface shape of an industry-standard GranuTools GranuDrum™.

ACCESS also represents a solution to the open problem of efficiently simulating large systems – the autonomous determination of particle properties is agnostic to the size of the particles simulated relative to the true size of particles used in experiment. Thus, it can be used to directly determine optimum properties for coarse-grained particles, which by definition cannot be calculated from direct characterisation measurements. A calibrated coarse-grained model would allow a typical simulation to be performed with a reduction in computation time between 160x and 1800x (simulation software dependent) compared to a full, non-coarse-grained simulation.

The algorithm itself is completely DEM engine-agnostic; it was implemented in an open-source Python library, providing an interface that is easy to use, but powerful enough to automatically parallelise arbitrary user scripts through code inspection and metaprogramming. It was used successfully from laptop-scale shared-memory machines to multi-node supercomputing clusters. Though ACCESS was built as a tool for the computational granular mechanics community, it is a fully fledged massively-parallel evolutionary optimisation framework that can work - without any modifications! - with arbitrary simulation techniques that require calibration against experimental data.

REFERENCES

[1] Rosato AD, Windows-Yule C. Segregation in Vibrated Granular Systems. Academic Press; 2020 Jun 5.

[2] Luding S. Introduction to discrete element methods: basic of contact force models and how to perform the micro-macro transition to continuum theory. European journal of environmental and civil engineering. 2008 Aug 1;12(7-8):785-826.