BIO
Dr J.P. Morrissey is a research fellow specialising in particulate mechanics and the Discrete Element Method (DEM) at the University of Edinburgh, Scotland, UK where he is a member of Granular & Geomechanical Processes Group within the School of Engineering.
He graduated with an M.Eng in Civil Engineering from the University of Edinburgh where he also completed his Ph.D, studying the behaviour of complex cohesive granular solids. His research covers both experimental methods and numerical simulations with a strong focus on DEM and its exploitation for industrial use cases. He has been involved in numerous European ITN’s and collaborated with and consulted for companies such as Altair, AstraZeneca, LKAB, Johnson Matthey, P&G, PepsiCo, Pfizer, Siemens and Tensar.
His research interests include:
- AI/ML
- Mechanics of particulate solids
- Experimental methods
- Characterisation and calibration
- Computational methods
- Numerical simulation
- Data analytics
PROPOSAL
Particulate solids are commonplace across numerous industries and in nature, but these complex materials are often poorly understood with difficult to predict behaviour. In recent decades, computationally expensive numerical simulations have increasingly been applied to develop understanding and better predict behaviour. The last decade has seen an explosion in the adoption of AI/ML methods in various scientific fields, with significant development focused on tools such as Physics-Informed Neural Networks (PINNs). While these tools have shown significant promise in other fields (fluid dynamics, weather prediction), their applicability to granular systems is limited by the lack of governing equations for such complex systems.
Graph-based Neural Networks (GNNs) offer a logical starting point for modelling dynamics granular systems: particles are naturally represented by nodes on the graph while edges represent the particle interactions. This enables GNNs to learn underlying interaction laws that govern the particle dynamics. GNN-based tools are often able to make reasonable predictions of particle dynamics but currently do not predict forces or stresses in a system, limiting their usefulness for making engineering decisions about complex industrial systems, e.g., predicting the pressures within a silo.
Current state-of-the-art GNN-based simulators [1–4] can handle simple geometry primitives but not complex boundaries and dynamic systems. A further limitation is that GNNs have high memory demands. A typical desktop GPU with 16GB of memory may be limited to just 20,000 particles: far fewer than needed to model most industrial systems. However, next-gen systems such as Isambard-AI create the opportunity to apply such tools on a much larger scale, o ering faster training and ability to solve larger problems.
Building upon existing tools, this work will develop the means to capture the dynamics of complex geometries and non-spherical particles in a GNN-based simulator. New encoder and decoder architectures will be added to the chosen tool and additional training datasets created. Currently available training datasets are low fidelity, often produced with limited or incorrect physics in MPM or SPH solvers, and lack the complex and important shear response of frictional particles.
With a fully trained model, the work will explore how such simulators can be integrated into existing workflows. A possible hybrid approach of using GNNs to predict particle positions and dynamics, coupled with an accurate force prediction model between timesteps, could be significantly faster than existing DEM solvers. By effectively replacing the expensive contact search step with a GNN, it may be possible to create an accurate and performant surrogate for a DEM solver.
Since the applicant is within the Explore AI stream, targeted training to fill skills gaps will be an essential precursor to this technical development. The other pillar of this project is knowledge transfer to the broader CCC-ParaSolS community through sustainable software practices and a series of webinars. These are essential for dissemination of knowledge beyond the applicant to the diverse community we have endeavoured to create in CCC-ParaSolS.
[1] Sanchez-Gonzalez et al. (2020): https://doi.org/10.48550/arXiv.2002.09405.
[2] Choi & Kumar (2024). https://doi.org/10.1016/j.compgeo.2024.106374.
[3] Kumar & Vantassel (2022): https://doi.org/10.48550/arXiv.2211.10228.
[4] Mayr et al. (2023): http://arxiv.org/abs/2106.11299.