Publications

@conference{servintowards,
title = {Towards a graph neural network solver for granular dynamics},
author = {Martin Servin and Holger Götz and Tomas Berglund and Erik Wallin},
url = {https://www.researchgate.net/profile/Martin-Servin/publication/350089857_Towards_a_graph_neural_network_solver_for_granular_dynamics/links/6050749892851cd8ce445ca6/Towards-a-graph-neural-network-solver-for-granular-dynamics.pdf},
year = {2021},
date = {2021-03-01},
booktitle = {VII International Conference on Particle-Based Methods (PARTICLES 2021)},
abstract = {The discrete element method (DEM) is a versatile but computationally intensive method for granular dynamics simulation. We investigate the possibility of accelerating DEM simulations using graph neural networks (GNN), which automatically support variable connectivity between particles. This approach was recently found promising for particle-based simulation of complex fluids [1]. We start from a time-implicit, or nonsmooth, DEM [2], where the computational bottleneck is the process of solving a mixed linear complementarity problem (MLCP) to obtain the contact forces and particle velocity update. This solve step is substituted by a GNN, trained to predict the MLCP solution. Following [1], we employ an encoder-process-decoder structure for the GNN. The particle and connectivity data is encoded in an input graph with particle mass, external force, and previous velocity as node attributes, and contact overlap, normal, and tangent vectors as edge attributes. The sought solution is represented in the output graph with the updated particle velocities as node attributes and the contact forces as edge attributes. In the intermediate processing step, the input graph is converted to a latent graph, which is then advanced with a fixed number of message passing steps involving a multilayer perceptron neural network for updating the edge and node values. The output graph, with the approximate solution to the MLCP, is finally computed by decoding the last processed latent graph. Both a supervised and unsupervised method are tested for training the network on granular simulation of particles in a rotating or static drum. AGX Dynamics [3] is used for running the simulations, and Pytorch [4] in combination with the Deep Graph Library [5] for the learning. The supervised model learns from ground truth MLCP solutions, computed using a projected Gauss-Seidel (PGS) solver, sampled from 1200 simulations involving 50-150 particles. The unsupervised model learns to minimize a loss function derived from the MLCP residual function using particle configurations extracted from the same simulations but ignoring the approximate solution from the PGS solver. The simulation samples are split into training data (80%), validation data (10%), and test data (10%). Network hyperparameter optimization is performed. The supervised GNN solver reaches an error level of 1% for the contact forces and 0.01% on the particle velocities for a static drum. For a rotating drum, the respective errors are 10% and 1%. The unsupervised GNN solver reaches 1% velocity errors, 5% normal forces errors, but it has significant problems with predicting the friction forces. The latter is presumably because of the discontinuous loss function that follows from the Coulomb friction law and therefore we explore regularization of it. Finally, we discuss the potential scalability and performance for large particle systems.},
keywords = {Algoryx},
pubstate = {published},
tppubtype = {conference}
}

@conference{laezzashape2020,
title = {Shape Control of Elastoplastic Deformable Linear Objects through Reinforcement Learning},
author = {Rita Laezza and Yiannis Karayiannidis},
url = {https://ras.papercept.net/proceedings/IROS20/3496.pdf},
year = {2020},
date = {2020-01-01},
booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Las Vegas (Virtual), USA, 2020-10-25 - 2020-10-29},
abstract = {Deformable object manipulation tasks have longbeen regarded as challenging robotic problems. However, untilrecently, very little work had been done on the subject, withmost robotic manipulation methods being developed for rigidobjects. As machine learning methods are becoming morepowerful, there are new model-free strategies to explore forthese objects, which are notoriously hard to model. This paperfocuses on shape control problems for Deformable Linear Objects (DLOs). Despite being one of the most researched classesof DLOs in terms of geometry, no other paper has focusedon materials with elastoplastic properties. Therefore, a novelshape control task, requiring permanent plastic deformationis implemented in a simulation environment. ReinforcementLearning methods are used to learn a continuous controlpolicy. To that end, a discrete curvature measure is usedas a low-dimensional state representation and as part of anintuitive reward function. Finally, three state-of-the-art actor-critic algorithms are compared on the proposed environmentand successfully achieve the goal shape.},
keywords = {External},
pubstate = {published},
tppubtype = {conference}
}

@conference{berglund2018virtual,
title = {Virtual commissioning of a mobile ore chute},
author = {Tomas Berglund and Kjell-Ove Mickelsson and Martin Servin},
url = {http://umit.cs.umu.se/modsimcomplmech/docs/papers/vcmoc_chops.pdf
http://umit.cs.umu.se/chute/
https://www.linkedin.com/pulse/virtual-commissioning-mobile-ore-chute-martin-servin/
},
year = {2018},
date = {2018-01-01},
booktitle = {The 9nth International Conference on Conveying and Handling of Particulate Solids (CHoPS), London, UK (2018)},
journal = {simulation},
volume = {10},
pages = {14th},
abstract = {This paper describes the virtual commissioning of a mobile ore chute for sequential loading of trucks from a conveyor system with a continuous material flow. The design and control were tested in simulation environment and improved prior to its installation in an underground mine in full production. The altered design met the performance goal and the amount of rock spill and wear on surrounding equipment could be reduced significantly. The simulations were based on a novel combination of discrete element and multibody simulation using a nonsmooth dynamics formulation, integrated in a 3D modeling software. This enable both fast simulation, based on original CAD drawings, and high flexibility in modifying the design and control.},
keywords = {Algoryx},
pubstate = {published},
tppubtype = {conference}
}

@conference{servinimpact,
title = {Impact force analysis with the nonsmooth discrete element method},
author = {Martin Servin and Tomas Berglund and Kjell-Ove Mickelsson},
url = {http://umit.cs.umu.se/modsimcomplmech/docs/papers/Impact_Analysis_Abstract.pdf},
year = {2017},
date = {2017-01-01},
booktitle = {Particles 2017},
abstract = {Analysis of impact forces is applicable for predicting particle breakage and equipment wear in systems for transportation and storage of granular materials, such as iron ore pellets [1,2]. The nonsmooth discrete element method (NDEM) [3,4] can be understood as a time-implicit version of the conventional, or smooth, discrete element method (DEM). The NDEM enable large time-step integration by allowing the velocities to change discontinuously in accordance with an imposed contact and impact law, expressed in terms of inequality and complementarity conditions, in addition to the rigid body equations of motion. In some applications this shorten the simulation time by several orders in magnitude [4] and thus enable quick design experiments and systematic exploration of the design and parameter space. An apparent deficiency with the NDEM is that the force that occur due to highvelocity impacts is not explicitly computed. Instead, the NDEM solver compute the net impulse that is necessary for fulfilling the imposed contact and impact law. The conditions for particle breakage and surface wear are directly dependent on the contact force and not the impulse.

A method is presented for a posteriori computation of the impact contact forces from NDEM impulses. For pair-wise contacts, the equivalent contact force can be deduced from the impulse by assuming the Hertz contact law [5]. This is extended to more complex contact networks, with simultaneous impacts, and validated using smooth DEM simulation. The method is demonstrated by studying systems are representative for transportation and storage of granular materials and analysing the statistical distribution of impact forces and energy dissipation, thus indicating breakage and wear.
},
keywords = {Algoryx},
pubstate = {published},
tppubtype = {conference}
}

@conference{servincomputational,
title = {Computational modeling of flow and size segregation in a stockpile with multiple outlets},
author = {Martin Servin and Tomas Berglund and J Boström},
url = {http://umit.cs.umu.se/modsimcomplmech/docs/papers/Stock_pile_flow_Abstract.pdf},
year = {2017},
date = {2017-01-01},
booktitle = {Particles 2017},
abstract = {Gravity reclaim stockpiles are widely used for storing large amounts of granular materials. Material is added to the surface via conveyors and is drawn from one or multiple outlets at the base of the pile, at a speed controlled with belt or apron feeders. In silos, the discharge typically occur as mass flow, which means that all the material move downwards at approximately the same speed and with little size segregation. The flow pattern in stockpiles, on the other hand, is that of funnel flow, where the material flow through vertical channels formed over each outlet [1]. These channels of active flow are surrounded by material that is stagnant, except near the surface where material flowing when the angle of repose is exceeded. Stockpiles are subject to significant size segregation due to several mechanisms [1,2]. The material is normally segregated already on the incoming conveyor and this is enhanced by trajectory segregation. Rolling and sifting segregation occur when the material spread on the pile surface. Segregation by percolation takes place in the shear zones between the funnel flow and stagnant zones.

We estimate the flow pattern and size segregation transport coefficients in a stockpile model with multiple outlet using the nonsmooth discrete element method [3,4]. Based on the analysis, similar to [5], a kinematic hybrid particle-cellular automata stockpile model is developed. Finally, we examine the possibility of realtime monitoring of the transport and the size segregation in a stockpile and feasibility of maintaining an net outflow with stable size distribution by controlling flow rate of the individual outlets.
},
keywords = {Algoryx},
pubstate = {published},
tppubtype = {conference}
}