First HPC@QMUL Annual User Meeting

When: Thursday, November 28, 2024, 9:00 AM - 5:00 PM
Where: BLOC, ArtsOne Building, Mile End

Book now

We're excited to invite our HPC users to our First HPC@QMUL Annual User Meeting.

Already an Apocrita user?
Would you like to share the greatness of Apocrita in your research?
Fancy to show to the community the High Performance Computing you’ve used?

Book your spot

We're inviting researchers who desire to give a 10-to-20-minute oral presentation to share their research and achievements using Apocrita. Please be advised that we only have 60 seats, so take this good opportunity for explain how technology and our services assist in their research, and you will be assured of a place in the event.

Our speakers will be invited to a catered break and will be given an Apocrita t-shirt. We also plan to highlight their research on our website HPC@QMUL. We will provide a separate space in the BLOC, where researchers can give a poster presentation to share their research and achievements using Apocrita.

We are looking forward to seeing you at the event, as we know each of you along with your team have so much to share.

Speakers so far: 

Qilei Li

Person re-identification (ReID) is a computer vison task that aims to identify individuals across various locations and time frames using images or videos collected from widely distributed camera networks. This task is fundamental in surveillance and security applications, where consistent tracking of individuals across diEerent environments is crucial. Accelerated by deep learning, ReID extracts reliable features that enable accurate person searches within large-scale datasets.
However, these models face the challenge of generalizing eEectively across data from diverse domains, each with unique visual characteristics and distribution patterns. Ensuring generalizability requires advanced representation learning strategies that can bridge these domain gaps, allowing ReID models to perform robustly in new environments. High-performance
computing (HPC) becomes essential here, as it provides the computational power needed to train and fine-tune these models on massive datasets. This capability enhances the scalability and adaptability of ReID, allowing models to process and
analyze data from vast, dynamic sources, ultimately ensuring more reliable and precise person identification across complex real-world scenarios. 

Jorge Eduardo Castro Cruces

The eEects of climate change are intensifying globally, driven by severe weather events that have
resulted in the decline of numerous species. It is imperative to obtain precise insights into the threats posed to our ecosystems, enabling us to eEectively manage and potentially alleviate further environmental deterioration in the future.
Ecological interconnectedness plays a crucial role in anticipating responses to climate change, as the decline of a single species can cascade through the network, impacting the populations of its predators or prey as species adapt to shifting environmental
conditions. Fortunately for us, this project endeavors to enhance our comprehension of how Global Environmental Change (GEC) disrupts ecological networks and redefines community organization. Leveraging machine learning techniques, we aim to unravel the structural and reassembly principles governing ecosystems under stress. By employing advanced graph data science, we will decode topological characteristics and forecast species adaptation
and rewiring in response to climate change.

Our model will pioneer a predictive framework to anticipate ecosystem-level responses, providing evidence-based guidance for conservation interventions. Furthermore, this research contributes to the burgeoning field of ecoinformatics by oEering novel methodologies for deciphering vast ecological datasets. Through these endeavors, we strive to empower stakeholders with tools to navigate the complexities of ecosystem resilience in the face of environmental change. In conclusion, this research outlines a comprehensive plan to predict new links  and ecosystem re-assembly using machine learning techniques. The findings from this research will contribute significantly to the field of Ecology and Machine Learning, providing valuable insights into community organization and resilience in the face ofbiodiversity loss.

Rabia Abid

The recent Covid pandemic has highlighted the need to understand and control the air flow indoors and particularly in public settings. An important sector that has been seriously affected by the pandemic was the educational sector, where many pupils were forced to go online and have missed much of the learning experience. It is only a question of time to when humanity will have another wave of airborne diseases pandemic and thus it is of importance to prepare the tools and settings for such events. Breaking the transmission chain by reducing the pathogen load in the air through ventilation is obvious. Reduced-order modelling is commonly based on the fully mixed air assumption. Although it can lead to useful guidance such as the required level of air changes per hour (ACH), it cannot explain why there are places in the room that seem to be more vulnerable for infection and cannot provide guidance where best to place the ventilation. In this context computational fluid dynamics (CFD) that includes particle dynamics can be useful. We will present a CFD implementation in OpenFOAM using the Eulerian-Lagrangian approach. The air flow is modelled using the URANS method where people’s thermal plumes are simulated using the Boussinesq approximation. The motion droplets are simulated through a Lagrangian simulation accounting for drag, buoyancy and gravity. The focus was on far field spreading and thus looking at droplet sizes of 2 and 15 microns which do not pass an evaporation phase (although such phase can be added using existing OpenFOAM tools for large droplets in the near field). A generic medium size class of two standing instructors and thirty sitting pupils is investigated, where one infector is assumed (an instructor or a pupil). A short speech of 5 seconds is simulated to study the motion of 10000 droplets, focusing on far field spreading. Two scenarios are simulated, a poorly ventilated classroom and a well venerated classroom (6 ACH).

Stark difference is observed between the two. The poorly ventilated classroom shows the dominance of the people's thermal plumes over the air flow and hence determining the droplets convection. This leaves places in the classroom which are more prone to accumulation of the droplets (and hence a higher infection risk) than others, e.g.  certain corners in the room. On the other hand, the ventilation causes more uniform spreading of the droplets in early times and better protects those sitting far from the infector at later times.

The effects of the infector’s position and the level of occupation in the classroom will be discussed in the presentation as well as an infection risk analysis adapted to CFD calculations and incomplete air mixing situations.

Mi Shuo

A new integratednumerical framework is developed for simulating he fluidstructure interaction of aquaculture systems.
The framework is based on our previous OpenFOAM formulation [1], while being coupled with a lumped massmooring model, MoorDyn [2], and a finite-element structural solver, EndoBeams [3]. The turbulent flow is taken as incompressible fluid solver and solved using a volume of fluid surface capturing method. The motion and deformation of the flexible nets are calculated using the screen and massspring methods. MoorDyn is used for simulating mooring lines while EndoBeams are used to calculate the deformation of other components of the aquaculture system, such as collars and frames. The coupling of all the components follows a loose-coupling method. The immersed boundary method is employed for the interactions between the fluid and all components of the aquaculture system. Fluid particle dynamics is also modelled using the Eulerian-Lagrangian to simulate fish disease waterborne transmission within aquaculture system area. The framework has been validated with extensive experimental data from the literature and is demonstrated as a robust tool to simulate the complex fluid-structure-particle dynamics of aquaculture systems.

Sebastien Paine

The environment in which circumstellar discs evolve plays a crucial role in their evolution, and the formation of planets. In stellar clusters, nearby large stars will irradiate gas and dust: heating the disc and entraining planet-forming materials. An important, but understudied aspect of this is what sizes of dust get carried in the disc wind. This aEects the amount and position of planet-forming solids, as well as dust shielding the disc from UV radiation. We have developed a particle solver to track the motion of dust in multidimensional simulations of photo-evaporating discs.

The code independently evolves 1000s of dust grains in parallel, using a golden-section search algorithm to find the maximum dust size entrained across the disc. We validate an existing analytic estimate in 1D by Facchini et al. (2016) but find the situation more complex for 2D axisymmetric models. There, dust entrainment varies significantly depending on where from the disc surface the dust is launched into the wind. However, we also find that the dust opacity is mostly uniform from all directions. This has implications for the structure of gas mass loss and observational characteristics of externally irradiated discs