Adjoint-accelerated Inference and Optimization Methods

The proposed research will demonstrate, with industrial examples, how scientists and engineers can efficiently combine data with prior physical knowledge. It will leverage Bayesian inference for rigour, adjoint methods for speed, and fluid dynamics for impact. Bayesian inference has been applied to fluid dynamics before but rarely using adjoint-accelerated inference and optimization methods (AXIOMs). These methods are crucial for practical applications because they dramatically accelerate data assimilation, especially when models contain thousands of parameters. This produces quantitatively-accurate physically-interpretable models that extrapolate successfully in directions in which the physics holds. Furthermore, they quantify the information content of data and rank physics-based, physics-agnostic, and combined models by calculating their relative likelihoods given the data. The proposed research will exploit AXIOMs to achieve a 10 times reduction in scan time of Flow-MRI (Magnetic Resonance Imaging) compared with state-of-the-art compressed sensing, transforming the accessibility of clinical Flow-MRI. This research will achieve a similar increase in the extractable information from experimental campaigns on gas turbine rigs, increasing reliability and reducing cost in a crucial European industry. It will also infer the rheometry of opaque fluids from a single Flow-MRI image, rigorously select the most appropriate turbulence model from data, and improve the robustness of widely-used model discovery algorithms. This project will encourage researchers to consider data in terms of information content rather than file size, enable the use of physics-based models alongside physics-agnostic models, and contribute to other areas in physical science through engagement with the UK’s Alan Turing Institute.