Skip to main content


  • Keynote 1.

Iain Styles.
School of Computer Science. University of Birmingham.

How HPC is evolving to support AI: A Case Study.

Artificial Intelligence has become the major driver for the development of new HPC systems and architectures, with both commercial companies and national governments investing huge sums to develop specialist infrastructure to support the largest AI models. AI poses particular challenges to HPC systems resulting from the nature of the computations and the size of the data and the way it is accessed. I will discuss what these challenges are and how we addressed them in the EPSRC-funded Baskerville Tier 2 system based in Birmingham which was specifically designed to support modern AI workloads. I will describe Baskerville’s hardware configuration, how that specifically supports AI workloads, and the software stack that was built to facilitate the transition of workloads away from the desktop systems beloved of so many AI researchers. I will also discuss the future needs of AI compute systems and how they may develop to meet them.


  • Keynote 2. 

David Young.
Software Sustainability Institute.
Astrophysics Research Centre, School of Mathematics and Physics. Queen's University Belfast.

Research Software Engineering: Origin Stories.

The pivotal role software now plays within most large research projects is beyond refute, yet the individuals writing the software are often overlooked. In March 2012, at Queen's College London, a small group met to discuss this lack of recognition and what could be done to rectify the situation. The historical outcome of that meeting was the birth of the Research Software Engineering (RSE) movement. In this talk, I will unpack the origin story of the RSE movement and its astounding growth over the past decade. Institutional and regional RSE groups across the UK provide valuable cross-discipline communities for like-minded developers, and many have become hubs for providing bespoke, industry-level software development to academics who previously would not have had access to such services. Host institutions benefit from the training given by members of RSE groups, teaching academics basic software and data manipulation skills and mentoring them in software best practices such as reproducibility, reusability and sustainability, helping improve productivity and the quality and reach of their research. Looking closer to home, I will also discuss how we can all benefit from investing in RSE-NI, Northern Ireland's fledgling regional RSE group.


Dr David Young is a Senior Research Software Engineer embedded in the Astrophysics Research Centre within QUB for over ten years. He has a wealth of experience working on various international, data-intensive projects, many in partnership with the European Southern Observatory and often within a technical leadership role. He mentors PGRs within his group and has a track record of advocating for best coding practices and sustainable, reusable research software. He is a Software Sustainability Institute fellow with a career history spanning research domains.


  • Talks

Syed Tauhidi.
School of Computer Science. Queen's University Belfast.

Using graphlets for accelerated subgraph isomorphism search.

The Subgraph Isomorphism (SI) problem, which finds applications in domains like bioinformatics, cheminformatics, astrophysics, graph databases, etc., requires searching for a small pattern graph within a larger data graph. Being NP-complete, no fast (polynomial-time) algorithm for the SI problem is known. Nevertheless, studies have proposed heuristic solutions that speeds up the search process using various pruning techniques. This study uses the Kelvin2 High-Performance Computing (HPC) to profile the performance of multiple heuristic algorithms on the runtime of different pattern and data graph pairs. Initially Machine Learning (ML) was used on the performance profile data to predict the most efficient heuristic algorithm for any given pattern graph. This approach achieved a speedup of up to 1.54 compared to any static heuristic algorithm. Currently internal substructures (graphlets) of the graph pairs are being used to develop a heuristic algorithm that prunes the search space to further accelerate the runtime.


Syed Tauhidi is a PhD student in Computer Science working on 'High-Performance Graph Analytics', under a collaboration between Queen's University Belfast (QUB) and Tezpur University (TU), sponsored by the Ministry of Education, Govt. of India. He is supervised by is Prof. Hans Vandierendonck (QUB), and co-supervised by Dr. Thai Son Mai (QUB) and Dr. Arindam Karmakar (TU). The research interests include machine learning, graph analytics and robotics. The author has affinities for traveling, cooking and fiddling with a guitar.


Mina Kazemi.
Hydrogen Safety. Ulster University.

Numerical Study on Characteristics of Hydrogen Under-expanded Jet Fires.

Onboard high pressure hydrogen storage is typical in automotive applications. The associated piping system may be damaged in an accident or the thermally activated pressure relief device can be activated, releasing high pressure hydrogen. The release is likely to ignite forming an under-expanded jet flame. The flame may blow out, potentially forming a flammable cloud in an enclosed space which could lead to an explosion. In this work 3D Computational Fluid Dynamics is used to model hydrogen flame stability. This requires discretization of the solution domain into extremely fine control volumes (CVs) to capture the shock structure and fuel/air mixing to delineate flow/flame interactions. Transport equations and combustion must be resolved for each CV with a high computational cost. Kelvin alleviates this issue enabling simulation of a scenario in three days using 1 node and 64 cores which is 4 times faster compared to a desktop computer with 16 cores. (150/150).


Mina Kazemi has been involved in the validation of hydrogen flame stability limits using CFD modelling since she joined the HySAFER team as a researcher at Ulster University in 2020. She presented two conference papers and submitted a journal paper on this subject.  
Her career in CFD modelling began in 2015 during her BSc program in Mechanical Engineering and continued in the field of Wildfire till she obtained her MSc in mechanical engineering-energy conversion in 2019.


Bonaventura Tagliafierro.
Civil Engineering. Universitat Politècnica de Catalunya.

High-performance GPU-accelerated software for advanced renewable energy simulations.

This work presents the outcomes of a computational challenge focused on evaluating the performance of renewable energy structures, such as wave energy converters (WECs) and floating offshore wind turbines (FOWTs). Anticipating their structural behaviour under environmental conditions exceeding typical sea and wind states has become increasingly crucial, and numerical simulations have gained greater reliability in this regard. Utilizing a Smoothed Particle Hydrodynamics (SPH)-based platform called DualSPHysics (, and with the support of the NI-HPC supercomputing centre, our research group conducted an extensive campaign to investigate wave-induced loads on a well-documented floating wind turbine platform known as DeepCwind, under extreme wave conditions. The presentation will include the results of our investigation which are part of a research paper, reviewing the computational features that we leveraged, such as the code CUDA parallelism.  In closing, we will discuss the possibility of performing multi-GPU simulations, thereby broadening the scope of the computable problems.


Bonaventura Tagliafierro is a civil engineer and architect. Former Fulbright Fellow at UW-Madison, now post-doctoral researcher at the Universitat Politècnica de Catalunya (Spain) and Visiting Scholar at QUB. Completed his doctoral program at the University of Salerno in 2022 (Italy), specializing in steel design and numerical methods for structural design. Has been collaborating with EPhysLab (University of Vigo, Spain) since 2019 as a researcher and joined the DualSPHysics code project. Computational methodologies include the Lagrangian Smoothed Particle Hydrodynamics (SPH) technique. Code applications involve wave energy converters, floating offshore wind turbine platforms, aiming at investigating performance under extreme events.


Michael Fayemiwo.
School of Computing, Engineering and Intelligent Systems. Ulster University.

Improving 2D Segmentation of Adrenal Gland using Test-time Augmentation on CT Images.

This study introduces an efficient pipeline for segmenting left and right adrenal glands from CT images, addressing the challenges posed by their small size, varied shapes and complex placement. The pipeline involves pre-processing steps like Hounsfield value adjustment, contrast enhancement, and automated slice selection. Post-processing includes test-time augmentation and removal of unconnected segments. The study utilized the Northern Ireland High-Performance Computing cluster due to high computational demands. The pipeline was implemented using 2D UNet architecture with four different backbones namely VGG16, Resnet34, Inceptionv3, and the default UNet. After post-processing, all models showed significant improvements in performance. Post-processing with Inceptionv3 increased Dice coefficient by 38% for the left gland and 11% for the right gland on AMOS dataset. Likewise, Resnet34 showed identical improvements on MICCAI dataset. Overall, the proposed pipeline outperforms existing methods with an 8% improvement in dice score, highlighting its effectiveness in adrenal gland segmentation from CT scans.


Saeideh Baghanian.
Civil Engineering. Queen's University Belfast.

Evaluating the Impact of Climate Change on marine Wave Characteristics on Kelvin-2.

High-performance computing (HPC) is vital for climate modelling, facilitating complex simulations, and processing large datasets. This study employs high-performance computing to conduct complex simulations using the third-generation wave model, SWAN. SWAN employs a full spectral representation with explicit modelling of all physical processes, highlighting the need for HPC to handle the complexities of such simulations. Utilizing wind data with high spatial and temporal resolution and bathymetric input, the simulations run in parallel, resulting in more reliable and precise insights for wave climate change. This research demonstrates the critical role of high-performance computing in scientific studies like this, allowing for advanced simulations that underpin our understanding of complex environmental systems.


Steph Merritt.
Astrophysics Research Centre. Queen's University Belfast.

Sorcha: a Survey Simulator for Solar System Science with the Legacy Survey of Space and Time.

We present Sorcha, our open-source survey simulator designed for the upcoming 10-year Legacy Survey of Space and Time at the Vera C. Rubin Observatory. Over its nominal 10 year lifespan, LSST could discover over 5 million main belt asteroids, 300,000 Jupiter Trojans, 100,000 near-Earth objects, and over 40,000 trans-Neptunian objects, providing an unprecedented opportunity for population studies of small bodies within our Solar System. However, like all surveys, these discoveries will be affected by complex observational biases, obscuring the true properties of the underlying populations. Sorcha will simulate these biases and will be the the first survey simulator capable of modelling all Solar System small-body populations, demonstrating the power and necessity of this survey simulator for accurate characterization of our Solar System. The code is presently used by the Rubin Solar System Data Products team to generate Data Preview 0.3, a simulated dataset of 15.7 million small Solar System objects.


Ryan Duddy.
Thornton Tomasetti.

Physics-Based Modelling for Machine Learning Model.

Transverse defects in rails are one of the main causes of derailment within the US. Defects can be detected by ultrasound devices called roller search units (RSU). This process is slow and requires skilled technicians to interpret the measurements (A-scans). 
Work at TT funded by the FRA proposes that machine learning algorithms can be trained to determine not only if a defect is present but the position and size of the defect, and using a fracture mechanics model, predict the remaining lifetime of the flaw. Ground truth data is sparse, so to supplement the training data, numerical modelling of the RSU is performed to artificially generate A-scans. The finite element method is used to propagate the acoustic wave equation coupled with Kirchhoff extrapolation to reduce the computational overhead.


  • Contributed posters

H. W. van der Hart, G. S. J. Armstrong and A. C. Brown.
Center of Theoretical Atomic, Molecular and Optical Physics. Queen's University Belfast.

Hybrid MPI/OMP parallel scheme for double ionization using time-dependent R-matrix theory.

The time-dependent R-matrix code, developed in Belfast, has established itself as one of the leading codes to investigate ultra-fast atomic processes in few-cycle light fields. This massively parallel code has recently been extended to describe the simultaneous emission of two electrons. The problem is divided into three regions, which use different numerical and parallelization schemes: a basis-set approach when both electrons are close to the nucleus, a finite-difference grid when both are far, and a hybrid scheme in between. Best machine utilization requires load balancing across the three different regions. We have identified the appropriate sector length required to balance the work in the hybrid one-electron-escape and the pure finite-difference two-electron-escape regions. Calculations can then be scaled up by moving from a 1-to-1 MPI communication scheme to an N-to-1 scheme.


Abdul-Akim Guseinov, Dmitry S. Karlov, Stephen Garland, Bianca Plouffe, Irina G. Tikhonova.
School of Pharmacy. Queen's University Belfast.

Opensource Virtual screening for membrane-facing and intracellular allosteric sites of NTSR1.

The search for druggable allosteric sites is an active area of drug discovery research. Molecular dynamics (MD) simulations are widely used to identify such sites. We applied our cosolvent MD-based method, called probe-confined dynamic mapping [1], to neurotensin receptor 1 (NTSR1). NTRS1 is a promising drug target for treating several neurological disorders and cancer types. 
We evaluate the functional relevance of the computationally predicted allosteric sites using in vitro experiments. For two allosteric sites, we performed structure-based virtual screening of 13 million commercially available molecules using opensource software. To increase chemical diversity, we implemented a new screening method. Promising molecules were further evaluated via MD simulations to identify compounds that form stable complexes with NTSR1. These molecules will be tested in vitro. We also screened a newly discovered NTSR1 allosteric site, recently observed using cryo-electron microscopy.

[1] Ciancetta, A.; Gill, A. K.; Ding, T.; Karlov, D. S.; Chalhoub, G.; McCormick, P. J.; Tikhonova, I. G. Probe Confined Dynamic Mapping for G Protein-Coupled Receptor Allosteric Site Prediction. ACS Central Science 2021.


Niall McElroy.
Center of Theoretical Atomic, Molecular and Optical Physics. Queen's University Belfast.

MPI and CUDA-based approaches for Electron-Impact R-Matrix Collisions.

The R-matrix approach is well known to be one of the most powerful and reliable methods for calculating phenomena such as electron-impact excitation/ionization in addition to photoionization and recombination, which are vital for accurately modelling a variety of plasmas. Recent and ongoing developments of the relativistic parallel DARC codes have enabled large advances in the accuracy of the atomic structure and subsequent collision calculations that are now feasible for lowly ionized heavy elements. MPI with the addition of CUBLAS routines has enabled an order-of-magnitude reduction in wall time in contrast to a strictly MPI-based approach. I will discuss how the development of both CPU and GPU codes has progressed on Kelvin-2 in this hybrid approach. Additionally, a selection of work undertaken on Kelvin2 that has been enabled by this will be shown.


Jiahui Zhou.
School of Chemistry and Chemical Engineering. Queen's University Belfast.

ADH Enzyme.

We investigated the asymmetric transformation of substituted cyclopentane-1,3-dione by Alcohol dehydrogenase (ADH) in the presence of the NADPH cofactor. GPU-accelerated Molecular simulations revealed the substrate's binding mode in LbADH, elucidating the origin of enzyme enantioselectivity. The opening and closing of the 191-205 loop above the substrate were found to shape the binding pocket and orientate the substrate, yielding different stereoisomeric products. Furthermore, Quantum Mechanics Molecular Mechanics (QMMM) calculations demonstrated the reaction pathway for ketone reduction by the enzyme, consistent with our experimental observations. Our research sheds light on the rational engineering of ADH for achieving stereodivergent stereoisomeric products and was supported by substantial computational resources.


Xiangwen Wang.
School of Chemistry and Chemical Engineering. Queen's University Belfast.

ALDELE: All-Purpose Deep Learning Toolkits for Predicting the Biocatalytic Activities of Enzymes.

In recent years, the rapid advancement of High-Performance Computing (HPC) and Graphics Processing Units (GPUs), along with the widespread adoption of machine learning techniques, has had a profound impact on accelerating enzyme catalysis research and the field of bioinformatics. These technological developments have provided essential tools for fast and accurate prediction of enzyme catalytic activities, facilitating the discovery of potential catalysts for industrial applications. In this study, we introduce ALDELE (an All-purpose Deep Learning-based Enzyme-catalyst prediction model), which combines structural and sequence representations of proteins with features for ligand representation, including subgraphs and overall physicochemical properties. A comprehensive evaluation of ALDELE demonstrates its capability to predict enzyme catalytic activities. Furthermore, it excels at identifying residue-based hotspots, guiding enzyme engineering efforts, and generating substrate heat maps to explore the substrate scope for a given biocatalyst. ALDELE offers an accessible and comprehensive solution by integrating various toolkits tailored for diverse purposes, all at a reasonable computational cost. Consequently, it significantly expedites the discovery of novel functional enzymes for potential industrial exploitation.


Pal Schmitt, Jeffrey Johnston.
School of Natural and Built Environment. Queen's University Belfast.

Kelvin-2: Powering Renewables.

Actuator line methods are popular tools for the assessment of wind and tidal turbines. This presentation  showcases two high impact examples which Kelvin2 enabled:
- The QUB submission to the "Tidal turbine benchmarking project", an EPSRC and ORE Supergen Hub funded project which aims to validate a large range of numerical methods against high quality experimental turbine data.
- Research on the effect of tilted wind turbines on wake development and turbine interaction, under careful consideration of atmospheric conditions.


Zohreh Moradinia.
School of Computer Sciences. Queen's University Belfast.

Machine Learning and Transprecise Computing in Multiphysics Simulation.

This research explores the application of transprecise computing in multiphysics simulations, aiming to balance accuracy and precision against performance and execution time. The study investigates how various parameters impact both execution time and simulation accuracy, with a focus on configuring simulations by adjusting key parameters to ensure reliable results in specific applications. To achieve these objectives, a classic heat transfer problem has been employed: flow over a heated plate, using the finite volume method as a numerical solution for the equations. The simulations are executed using the Kelvin2 High-Performance Computing (HPC) system. Additionally, machine learning models have been used for predicting accuracy and execution time in multiphysics simulations. This innovative approach has the potential to significantly reduce the time and resources required for such simulations, offering valuable insights for engineers and scientists in system design and optimization decisions.


Ryan Duddy.
Centre for Quantum Materials and Technologies. Queen's University Belfast.

Ab Initio Design of Plasmonic Materials.

Plasmonics is the study of plasmons, which are coherent oscillations of electrons controlled by coupling free electrons to light. Plasmons have a range of technological applications due to their ability to confine electric fields to areas below the diffraction limit of light. Novel materials that can operate in harsh environments (e.g. high temperatures) are required for plasmonics. First principal calculations have become a computationally feasible avenue to investigate these systems in the age of HPC facilities.
The current work implements real-time TDDFT on Octopus code to calculate the dielectric function of metals (Noble metals, Metal Nitrides, and Alkali metals). Two novel methods were developed to improve the description of the metal dielectric function, a real-time TDDFT approach to calculating the plasma frequency and a modified ACBN0 functional to improve the description of optical absorption in metals.


Declan French, Barry Quinn.
Queen's Business School. Queen's University Belfast.

Economic Costs of Culture Displays in Northern Ireland.

This research project will study the economic cost of cultural displays in Northern Ireland. The project will identify the prevalence of cultural displays, test for negative effects on house prices, and test for negative effects on rates of GP prescribing for anxiety and depression. The project is important because it will add to the NI Longitudinal Study, help answer policy questions about residential segregation and labor mobility, and employ state-of-the-art computer vision techniques.