Lab experiment optimisation using coupled Finite Element Analysis and Machine Learning
conference contributionposted on 14.05.2021, 08:16 by R. Lewis, L.l.M. Evans, R. Otin, K. Leng, A. Davis, A.D.L. Hancock, J. Thiyagalingam, P. Nithiarasu
The inside of a fusion energy device is one of the most challenging environments known about, with temperatures exceeding 100,000,000 °C in the centre of the plasma. The Heating by Induction to Verify Extremes (HIVE) testing facility at the UK Atomic Energy Authority (UKAEA) uses high heat flux experiments to research and develop plasma-facing components which experience the most severe thermal loads in such a machine. Induction heating is used to deliver up to 20 MW/m2 of surface power to a component while high-pressure coolant passes through it. Performing these experiments is expensive due to their high-power usage and the setup time required by technicians. It is imperative therefore that each experiment provides as much insight as possible about a component’s performance. Experimental parameters, such as induction coil design, frequency or power, are decided based on the memories, experience and gut-feeling of an experienced operator. In the era of High-Performance Computing (HPC), Big Data and Machine Learning (ML) the optimal parameters can be deduced utilising these methods. Moreover, ML algorithms have proved to be excellent at gaining insight from high-dimensional data, which is a common struggle for the human mind. A fully automated Virtual Twin (VT) of HIVE has been developed using VirtualLab, an open-source package created by the authors used to perform virtual experiments. Simulation results provide data points across the parameter space which VirtualLab’s ML routine can use to gain valuable insight, including the optimal experimental parameters. Utilising VirtualLab’s multi-node and High Throughput Computing (HTC) capabilities, a large number of simulations are performed for a variety of experimental parameters in a very short space of time, allowing rapid iterations of the engineering design cycle. This is an example of human-machine collaboration guiding, but not making, decisions which is becoming an increasingly important and useful tool in a wide variety of fields.
- Mathematical Sciences