Digital twins for multiphase flow systems

Explaining the science

Process engineering within the energy and manufacturing sectors relies on a thorough understanding of multiphase flow systems. Systems comprising liquids, gases and solid particles are complex to control and feature poorly understood dynamics regarding phase formation and transitions. Digital twin technologies will unlock barriers to operational efficiency and industrial process safety through a fusion of simulation and experimentally derived data streams.

A key metric of digital twin performance is the underlying simulation fidelity. In many engineering cases, physics based models outperform purely statistical models; as limited data are typically available for inference, or data may not exist due to the intrusiveness of measurement techniques. Simulation fidelity, however, comes at a cost: slow computation times limit iteration and parameter-exploration opportunities, while high-resource demands restrict computations to energy-intensive high-performance clusters. New hybrid physics/machine learning models will hence unlock real-time digital twin coupling, and allow for computation on low-intensity cloud resources. A tradeoff between simulation realism and relevancy must be navigated in order to deliver solutions: rapid, low-fidelity solutions can be of high decision-making utility, provided that uncertainty is explicitly quantified.

Advances in machine learning can be exploited to calibrate underlying simulations with physical data streams, leading to rapid prediction tools for complex multiphysics scenarios. This project's focus areas include reduced-order modelling, surrogate modelling and active learning strategies, in addition to numerical solvers for continuum mechanics partial differential equation systems. Incorporating physics based constraints and insights, such as energy/mass conservation and empirical closures, into machine learnt models is a priority.