Skip to main content
European Commission logo print header

EXAscale Quantification of Uncertainties for Technology and Science Simulation

Periodic Reporting for period 2 - ExaQUte (EXAscale Quantification of Uncertainties for Technology and Science Simulation)

Reporting period: 2019-12-01 to 2021-11-30

Recent advancements in high performance computing (HPC) will soon allow for the use of exascale systems in industrial practice, bringing the immense computational power of today’s machines to real engineering applications. The ExaQUte project aims at exploiting such systems, pushing our current physics simulation capabilities by performing Uncertainty Quantification (UQ) and Optimization Under Uncertainties (UQQ).

If we analyze the performance of current simulation tools, nowadays we can use numerical methods to accurately simulate and predict the behavior of a real phenomenon (for instance the interaction of a building and the wind), provided that the values of relevant parameters at the initial state from external contributions to the system are known beforehand. This type of analysis can be costly for large and accurate cases such as high-rise building.

Uncertainty comes into play when the input parameters of our problem are not exact, meaning that the solution of the problem we are simulating will not be the exact one. What if we had to make decisions worth millions of euros based on something that we do not know at its fullest?

ExaQUte overcomes this situation performing a large number of repeated simulations with similar but different input and boundary parameters, as opposed to running only one huge simulation. This approach, which can efficiently leverage supercomputers, allows evaluating how different conditions impact the solution and thus to increase our insight into the problem of interest.

Next, once we are able to model an uncertain phenomenon such as wind, how can we use this knowledge in engineering? How can we exploit the uncertain solution? Is it possible to use this result to optimize the shape of buildings accordingly? ExaQUte will answer this by combining the concept of uncertainties with optimization. Typically, optimization is used to find the optimal design that yields the best performance. This, however, may result in fragile designs that are only adequate for the set of conditions they were conceived for. In this regard, the information given by uncertainty quantification allows us to converge to a final robust design which takes into account a wide range of different and uncertain conditions.
We defined a common API for the PyCOMPSs programming model and HyperLoom scheduler provided by BSC and IT4I. To optimize the applications on top of the programming models, we performed an analysis of the current status of the partners tools and testings and benchmarking experiments A benchmark using the MLMC algorithm was run on the two systems. Lastly we defind an interface to fast local storage.

The capability of employing NURBS In the calculations, and of mapping to and from exact (NURBS-based) geometries to embedded geometries was developed and introduced in the computational pipeline. This isi completed by the development of different strategies to leverage adaptive refinement capabilities, either based on solution-based error estimation or on the use of the geometrical distance to the object.

The framework for octree-based mesh generation & adaptation is now fully functional in FEMPAR (https://github.com/fempar) and its p4est interface. A conformal mesh adaptation library constructed on top of the MMG software is developed. The second release implemented an “in-house” migration by an advancing front method available at (https://github.com/MmgTools/ParMmg).

The new XMC library released has been interfaced with the adaptive meshing framework of the Kratos MultiPhysics engine. Users are able to run MLMC simulations where the meshes to compute finer samples are adapted successively on the results of the coarser levels. First results have been obtained with a simpleMC algorithm by using metric-based adaptivity for a time-dependent CFD problem with high Reynold’s number.

A hierarchy of benchmark optimization problems for the design of tall buildings has been developed accounting various uncertainties in wind engineering. Aldjoint-based stochastic optimization has been performed for the task of airfoil design and the design of thin shells and design of thin shells subject to uncertain loads.
The chosen strategy for Uncertainty Quantification (UQ) is to exploit MC algorithms. The parallelism can be easily exploited when running in supercomputers, and the most important point is to properly handle different execution sizes. We have already developed a framework capable of dynamically allocating resources and computing more MC iterations, which is a fundamental feature for hierarchy-adaptive MC algorithms.

To ensure full usage of the machine, the asynchronous analogous of the algorithms has been developed, and proved how this strategy increases the usage efficiency. MLMC estimators, capable of handling goal/solution-oriented adaptively refined hierarchies, were also added.

Concerning Optimization Under Uncertainties (OUU), we aim to provide different risk measures and to update those on-the-fly to keep a maximum parallel efficiency: expected value, variance, higher moments, quantiles, buffered failure probability and Conditional Value at Risk. Another open question is the application to complex engineering problems. A gradient-based techniques based on adjoint calculus was developed for the full-potential equation and applied for the optimization of a wing profile. Parametric optimization was developed in application to a wind engineering problem.

A novel approach, based on the combination of time average and ensamble averaging is proposed to reduce the computational effort needed for wind predictions. The new approach is particularly convenient when running in distributed environments and sufficiently computational resources are available. We found that computational times decreases with respect to the standard method of running a unique simulation with a large time window.

IMPACT

Aside for the possibility of scaling up to any system size, ExaQUte took advantage of persistent local storage. This allowed sidestepping IO bottlenecks, which represent one of the practical limitations of future HPC hardware.

The results from ExaQUte are open source, allowing industry and academia to develop world-class products and services that will help to maintain EC leadership in HPC. The modular structure used in the software framework, will allows the project outcomes to be used beyond the demonstrator application by a larger user community, including industry (civil engineering) which has historically not been a major customer of HPC infrastructure.

The application field of ExaQUte has potencial high socialeconomic and environmental impact for predictive safety of civil constructions( buildings, bridges, harbours, …) under forces due to water or natural hazards . The output of this research is essential for enhanced analysis, risk assessment and performance-based design of constructions, to protect population and infrastructure against hazards.

ExaQUte will contribute to keeping and fostering that knowledge and know-how of scientists within Europe. The use of HPC computers requires significant expertise. ExaQUte provides room for highly skilled scientists to develop their work at the highest level of international research in European institutions,
Geometry (L) and pressure (R) distributions for the full-potential equation
solution-oriented adaptive refined unstructured anisotropic mesh of a 2-D building