Research Fellows
Pavanakumar Mohanamuraly
Early Stage Researcher 3 at Queen Mary University of London
FastUQ: A New multi-level multi-fidelity Monte Carlo Method for Uncertain Quantification (UQ)
A new adjoint-assisted multi-level multi-fidelity method was developed for fast and accurate UQ. The method was demonstrated on aerodynamic UQ of manufacturing imperfections on Turbine blades (cascade). The method was shown to yield results as accurate as the original high fidelity method but with a fraction of the computational cost.
Figure 1: Comparision of low fidelity and FastUQ results against high fidelity simulation
Figure 2: Computational cost comparison between low fidelity and fastUQ against high fidelity simulation
[1] P. Mohanamuraly, "Fast Adjoint-assisted Multilevel Multifidelity Method for Uncertainty Quantification of the Aleatoric Kind", PhD thesis, Queen Mary University of London, 2019.
Surrogate model for uncertainty quantification (Inexpensive Monte-Carlo)
Geometrical/shape uncertainties occur naturally in all practical realms of industrial application. For example, ice formation on aircraft wing leading edge, manufacturing/machining defects, corrosion and wear of moving parts, and the list goes on. Industrial design should thus be robust enough to deliver optimal performance even under such uncertainties. My task is to quantify and estimate the effect of these uncertainties on a given cost or objective function at a practical computational cost.
Publications
[1] P Mohanamuraly, J Mueller, "Adjoint-assisted multi-level approach for quantifying geometry-induced uncertainties in robust aerodynamic shape optimisation", UNCECOMP 2017.
MPI Adjoints for hand-assembled fixed-point iteration solver
Parallel CFD and adjoint solvers are highly essential to obtain results in turn-around time, which is practical for industrial application. Discrete adjoint solvers employing hand-assembled adjoint fixed point iteration present some key advantages over brute force differentiation. But making them work in MPI mode can be quite a challenge because the hand-assembly can lead to non-intuitive adjoint MPI code. We illustrate this using an example. Consider the drag adjoint cost function, where each processor calculates its local value of drag and then accumulates overall MPI ranks using MPI_Alleduce (MPI_SUM) to obtain the final result. For a zero-halo partitioned mesh, we find that in the adjoint (reverse mode) one has to accumulate the adjoint values at the shared nodes to get correct results. This is totally non-intuitive and requires careful examination to get the adjoint right. This is an important contribution of this research to the adjoint research community.
Publications
[1] P. Mohanamuraly, L. Hascote, J. D. Mueller, "Seeding and adjoining zero-halo partitioned parallel scientific codes", OMS Journal, (accepted), 2019.
[2] P. Mohanamuraly, J. Huckelhiem, and J. D. Mueller, "Hybrid parallelisation of an algorithmically differentiated adjoint solver", ECCOMAS Congress 2016, 2016.
[3] P Mohanamuraly, J Mueller, "Data partitioning and MPI adjoints", 20th European Workshop on Automatic Differentiation, INRIA Sophia-Antipolis, 2017.
[4] P Mohanamuraly, J Huckelhiem, J Mueller, "STAMPS: an Efficient Hybrid-Parallel Discrete-Adjoint CFD Solver for Aerodynamic Design", Eurogen 2017.
Meshless optimised mesh smoothing framework
In a typical aerodynamic shape optimisation loop, one perturbs a design surface based on a predicted value and obtains the solution to the perturbed configuration. In a CFD workflow, the design surface perturbation must be translated into the volumetric mesh node perturbation using a suitable model like spring analogy, or linear elasticity equations. Often it is possible to obtain volumetric meshes free from non-zero volumes but the quality of the elements is much worse than the unperturbed mesh. Mesh smoothing techniques help alleviate this problem by moving mesh nodes optimised for a specific mesh quality metric.
Many algorithms are available in the literature to tackle this problem but most of them suffer from one conceptual problem: mesh nodes are perturbed based on an objective function (quality metric), which is defined at mesh elements. We alleviate this issue by defining a quality metric, which is not based on elements (meshless). But the metric does contain the overall mesh quality information and is quite suitable for highly stretched boundary layer types meshes. This mesh quality metric has an exact analytical form for the forward and reverse differential and can be constructed using fast matrix-matrix, or matrix-vector products.
Publications
[1] P. Mohanamuraly, and J. D. Mueller, "A meshless optimised mesh-smoothing framework", ASMO UK/ISSMO/NOED2016, 2016
CAD-based optimisation with geometric/physical constraints
Publication
[1] O Mykhaskiv, P Mohanamuraly, JD Müller, S Xu, S Timme, "CAD-based shape optimisation of the NASA CRM wing-body intersection using differentiated CAD-kernel", 35th AIAA Applied Aerodynamics Conference, 2017.
[2] O Mykhaskiv, M Banovic, S Auriemma, P Mohanamuraly, J Mueller, A Walther, H Legrand, "NURBS-based and Parametric-based Shape Optimisation with differentiated CAD Kernel", Computer-Aided Design and Applications, 2017.