With exascale performant simulations key to advancing fusion knowledge and especially tokamak design or “Whole Device Modelling (WDM)”, VVUQ deployment is essential for successfully creating models that can reliably guide highly expensive procurement decisions for future commercial fusion reactors. Our primary exemplar will deploy our toolkit to quantify the uncertainty characteristics of a suite of algorithms being developed to build a highly coupled, multi-physics and multi-scale exascale class simulation of the tokamak plasma “exhaust” (or “divertor” region of the tokamak) that is “actionable”.
Weather and climate forecasting
The UCL Met Office Academic Partnership led by Serge Guillas, Co-I of the SEAVEA project, is devoted to Data Sciences for weather and climate. It gathers within four working groups experts from eight departments of UCL and across the Met Office. It covers multiple domains of modeling such as the ocean, sea-ice, paleoclimate, climate change, atmospheric turbulence, space weather, air quality as well as Data Assimilation, Machine Learning for nowcasting, and UQ methods. A UCL-Met Office workshop on UQ and parametrizations held in June 2021 identified cross-cutting approaches for the deployment of UQ methods over this range of domains to solve the most pressing problems. These include the Air Quality Unified Model (AQUM), with Cambridge exploring uncertainties in atmospheric chemistry at high resolutions, and several other domains.
Turbulent fluid dynamics
The DDWG “Turbulence at the Exascale” will work with us to demonstrate how VVUQ methods can quantify the uncertainty in turbulence modelling, for configurations for which existing turbulent models have performed poorly and in the study of transitions such as arise in avalanches which require large amounts of simulation data.
The properties of advanced nanocomposite materials with desirable macroscopic properties are sensitive to interactions at the nanoscale, such as between polymers and 2D materials like graphene. We will work with the DDWG “Systems Engineering” to investigate the reliability of the scalable SCEMa multiscale application which couples the LAMMPS MD code to the DealII FE code to link the molecular interactions to engineering scale properties such as strength and toughness. We plan to apply semi-intrusive VVUQ to control how the uncertainties at one scale influence the behaviour at the other. Carefully constructed surrogates applied at the micro level, based on model-consistent training that perform well in the coupled environment, will be developed to reduce the cost and severe load imbalance caused by explicit MD simulation.
Simulations of blood flow through arteries and veins have become increasingly accurate and personalised; they are potentially useful in clinical contexts such as surgical planning. However, VVUQ procedures are essential to gain the required regulatory approvals from FDA, EMA and MHRA. We will use UKCOMES’ flagship HemeLB code to create and deploy personalised human-scale models for high resolution 3D blood flow simulation. We will use the SEAVEA toolkit to quantify how changes to both physiological and numerical modelling parameters can impact flow throughout the body. HemeLB has been scaled to over 300,000 CPU cores (on SuperMUC-NG (SNG) at LRZ) and 24,000 GPUs (on Summit at OLCF), providing a benchmark for the exascale potential of SEAVEA methodologies.
IMPECCABLE is an in silico approach to drug discovery that involves coupling molecular dynamics (MD) based ligand-protein binding free energy approaches with machine learning (ML), including docking and DeepDriveMD, each complementing the other in terms of strengths and shortcomings to accelerate the identification of drugs. The overall workflow runs on Summit and enables the rapid ranking of the best compounds for experimental study and further refinement. The next stage will endow the full workflow conjoining both ML and MD with built-in uncertainty quantification to ensure that the rankings are actionable. We have recently completed a study of parametric uncertainty in the MD system and will also use ensembles for ML; here we will undertake research on uncertainty propagation between the different models to render the output actionable.
Brunel University London (BUL) uses agent- based models, in conjunction with Save the Children, to forecast the arrival of people who are forcibly displaced by violent conflict. The core solver Flee will be run using external HPC resource allocations, and scales efficiently to 16,384 cores for large problems. It has been subjected to sensitivity analysis using VECMA software, but such analysis is often too expensive to perform routinely for new conflicts. Within SEAVEA we seek to leverage advances in the toolkit to reduce the cost of performing SA for our application to make it possible for NGO data scientists to perform sensitivity analysis themselves.
Brunel University London (BUL) has developed an agent-based simulation code which mimics the spread of COVID-19 in a local region, such as a county or city. This code, named the Flu And Coronavirus Simulator (FACS) has been used to support local hospital trusts in London, and has been trialled for modelling pandemics in four different countries as part of the STAMINA EU project (www.stamina-project.eu). Within SEAVEA we seek to perform SA for our application not only to help quantify forecast uncertainties, but also to understand to what extent the choice of geographical region affects the input parameter sensitivities of the simulation.