MS1-10 - Scientific machine learning for non-intrusive reduced-order models

C. Pain (Imperial College London) , C. Heaney (Imperial College London) *

Machine learning has had a significant impact on so-called non-intrusive reduced-order models (NIROMs) over the last 10 years [1]. Coming under the umbrella of data-driven ROMs, NIROMs describe the evolution of a system using interpolation techniques such as splines, radial basis functions [2] and machine learning algorithms [3]. Developing efficient and reliable NIROMs is vital in order for these models to contribute to the use of numerical simulation in solving fluid dynamics problems. For example, NIROMs could play an important role in uncertainty quantification, data assimilation and digital twins, which are all computationally expensive procedures. Although these methods show great promise, challenges remain relating to how well these models can make reliable predictions for unseen scenarios. In this mini-symposium, we invite talks which propose strategies incorporating machine learning techniques within non-intrusive reduced-order models, with an emphasis on building reliable models. Some current steps towards this include physics-informed networks [4,5], adversarial networks [6]. It would be especially interesting to discuss methods to increase reliability of the neural networks and how authors assess the quality of the predictions.