Turbocharging your research with cloud HPC + AI

Speaker: Dr Stefano Angioni, Research Computing Manager (Acting)

Abstract: Using Cloud for high performance computing (HPC) can enable researchers to access innovation quickly, to get results faster and to pay only for the resources that have been consumed.

At the University of Bath, we have built a state-of the-art HPC system in Azure - Microsoft’s Cloud - called Nimbus. The cloud HPC system is the first (to our knowledge) campus wide HPC system within the education sector built entirely in the Cloud, not only in the UK but in the entire world. This was a team effort involving tens of people, belonging to both professional services and research staff, from many departments across the whole University and including also external contractors.

This talk will discuss what Nimbus offers to our users in terms of storage and compute, the components and architecture of the system, the challenges we faced during the implementation, and finally will also give some insight into many of the new features being considered as part of our future roadmap.

Using 3-D satellite observations, spectral analysis and HPC to produce the world's largest climatology of atmospheric waves

Speaker: Dr Neil Hindley

Abstract: Just like the surface of the ocean, the Earth's atmosphere is in constant motion. From planetary-scale flows like the jet stream to turbulence on the scale of a few centimetres, dynamic processes are crucial in shaping both short-term weather and long-term climate systems. However, the small-scale waves whose aggregated effects can drive global circulations are notoriously difficult to simulate in numerical models of the atmosphere (such as the Met Office model) and instead they must be mathematically approximated. But how do we know what values to use for these approximations? To do this, we must analyse many multi-terabytes of Earth observations to understand these waves globally throughout the atmosphere from the surface to the edge of space. This is an enormous high-performance computing (HPC) challenge.

In this talk, I will explain how we used Bath's Balena HPC system to produce the world's largest observational climatology of atmospheric waves, using 3-D spectral analysis of 3-D satellite observations in methods pioneered here at Bath. I will explain how this climatology is being used to validate numerical models at some of the world's leading weather centres, ultimately setting us on the path towards more accurate future predictions of weather and climate.

Surface-to-space atmospheric waves from Hunga Tonga-Hunga Ha’apai eruption

Speaker: Dr Corwin Wright

Abstract: The January 2022 Hunga Tonga–Hunga Haʻapai eruption was one of the most explosive volcanic events of the modern era, producing a vertical plume which peaked > 50km above the Earth. The initial explosion and subsequent plume triggered atmospheric waves which propagated around the world multiple times.

A global-scale wave response of this magnitude from a single source has not previously been observed. We used a comprehensive set of satellite and ground based observations, pre-processed for analysis using HPC, to quantify this response from surface to ionosphere. A broad spectrum of waves was triggered by the initial explosion, including Lamb waves at all heights and gravity waves in the middle atmosphere. Gravity waves have not previously been observed propagating at this speed or over the whole Earth from a single source. Latent heat release from the plume remained the most significant individual gravity wave source worldwide for >12 hours, producing circular wavefronts visible across the Pacific basin in satellite observations. A single source dominating such a large region is also unique in the observational record.

The Hunga Tonga eruption represents a key natural experiment in how the atmosphere responds to a sudden point-source-driven state change, which will be of use for improving weather and climate models.

Cluster-based HPC solution for logistic regression, random forest and XGBoost algorithms to predict memory impairment in older adults

Speaker: Dr Prasad Nishtala

Abstract: Background: Machine learning plays a crucial role in analysing large healthcare data and predicting clinical outcomes. We present here an approach aided by High-Performance Computing (HPC) to analyse one of the largest and the most comprehensive datasets of geriatric assessments, the interRAI. The world-leading interRAI dataset standardised needs assessment, and the dataset can be linked to other healthcare data, including demographic, psychosocial and other clinical variables. The interRAI dataset currently consists of 250 variables.

Objectives: This study combines mainstream machine-learning algorithms with the power of HPC to predict memory impairment in older adults 65 years and above. Logistic regression, random forest and XGBoost algorithms were considered.

Methods: We analysed the interRAI data using the HPC cluster (G4W-Isambard- Run on XCI Marvell Thunder X2 nodes) and identified 24 variables that are key discriminators of memory impairment risk to train predictors based on logistic regression, random forest and XGBoost. Each predictor was tested by 100-fold cross-validation, by parallelised implementations over a single node with 64 cores, with the aid of the map function in the R package purrr. We evaluated model performance using the area under the receiver-operating characteristic curve, F1, accuracy, sensitivity and specificity, and negative and positive predictive value.

Results: The overall statistics demonstrated the mean AUC for the logistic regression algorithm is 0.787 (se=0.00283) at lambda=0.0001. For the random forest, the mean AUC is 0.718 (se =0.00271) at mtry=4, and the mean AUC of XGBoost is 0.799 (se =0.00300). The XGBoost model achieved the highest accuracy ( 81%).

Conclusions: With the aid of HPC, we successfully ran a large model efficiently within 6 hours. The cluster-based HPC cross-validation solution indicated that XGBoost works best compared to logistic regression and Random forest. This cluster-based HPC solution for logistic regression, random forest and XGBoost algorithms will enable an analysis of large interRAI data currently greater than a million individuals and growing.

Simulations of the Realheart Total Artificial Heart Using a Novel Fluid-Structure Interaction Approach: The Influence of Heart Rate Variation

Speakers: Joe Bornoff and Dr Katharine Fraser

Abstract: Objectives: The Realheart is a novel total artificial heart (TAH) that mimics the mechanics of the native heart by vertical translation of an atrioventricular (AV) plane. It produces pulsatile flow governed by a pair of bileaflet mechanical heart valves. A fluid-driven numerical modelling strategy was developed, and in this study, that method was implemented into a previous generation Realheart TAH design and validated against a mock-circulation study for various heart rates.

Methods: A CFD model of the left side of the Realheart V11c was created using Ansys Fluent V2021R1, which included an atrial and ventricle chamber, aortic outflow valve, and mitral valve that was housed within an AV plane enclosure. Overset meshing combined these elements into a single mesh. Dynamic meshing was used to control the motion of the overset mesh zones, where the AV plane and mitral valve translated vertically, and both sets of valves were fluid-driven utilising the six degrees of freedom solver. The systole duration was 40% of the cycle duration, and the model was evaluated at 24 mm stroke length at 80, 100, and 120 bpm. A constant pressure boundary was used at the inlet, and a 2-element Windkessel model was applied at the outlet, approximating vascular resistance and compliance. Simulations were carried out using Baths HPC cloud service Janus, using HC-44 nodes with 32 cores. Each simulation took approximately 24 hours to solve two pump cycles.

Results: Average flow rates were 4.7, 5.7 and 6.8 L/min at 80, 100, and 120 bpm respectively. These are within 5% of the experimental values of 4.5, 5.5 and 6.5 L/min. Peak flow rates were 20.8, 26.0, and 31.7 L/min respectively, compared to 26.5, 30, and 35 L/min experimentally, a 10-15% difference due to a smoother numerical mass flow response.