Análisis de clústeres para simulaciones de mecánica granular mediante algoritmos de aprendizaje automático

RIM, DANIELA NOEMI; MILLÁN, EMMANUEL N.; PLANES, MARÍA BELÉN; BRINGA, EDUARDO M.; MOYANO, LUIS G.

Entre ciencia e ingeniería

Universidad Catolica de Pereira

Año: 2020 vol. 14 p. 81 - 86

1909-8367

Molecular Dynamics (MD) simulations on graincollisions allow to incorporate complex properties of dustinteractions. We performed simulations of collisions of porousgrains, each with many particles, using the MD softwareLAMMPS. The simulations consisted of a projectile grainstriking a larger immobile target grain, with different impactvelocities. The disadvantage of this method is the largecomputational cost due to a large number of particles beingmodeled. Machine Learning (ML) has the power to manipulatelarge data and build predictive models that could reduce MDsimulation times. Using ML algorithms (Support VectorMachine and Random Forest), we are able to predict the outcomeof MD simulations regarding fragment formation after a numberof steps smaller than in usual MD simulations. We achieved atime reduction of at least 46%, for 90% accuracy. These resultsshow that SVM and RF can be powerful yet simple tools toreduce computational cost in collision fragmentation simulations.