RaPiD models: Reciprocal Physics and Data models

Physics-based models are often expensive and slow, and the amount of information available to inform critical real-time decisions may be limited. Data-driven ML models are lightning-fast but are limited by the experience (data) used to train them. This does not limit the physics-based models which can be run for any conceivable (unobserved) scenario where the implemented physics is governing the system. The objective of this project is to establish a robust methodology of how to combine physics-based and data-driven models, alleviating their deficiencies by capturing their complementary advantages. The underlying idea is to combine well-established and robust physics-based models, made effective by advanced reduced-order modeling and the use of probabilistic data-driven models. This reduces uncertainty and focuses simulation efforts where the information gained produces the most value for the relevant decision context. This is a valuable contribution towards more specific, accurate, and timely decision support in the operation of safety-critical systems. In its first year, the project has run one case study where data-driven and physics-based methods have been combined to increase confidence in a critical operational decision in a physical system. The project has also initiated several open-source packages to increase the efficiency of finite element models, and how to combine physics-based and data-driven models. These are under active development and are published on https://github.com/rapid-models. Follow the project on https://rapid-models.dnvgl.com/

This project aims to provide more specific, accurate and timely decision support in operation of safety-critical systems, by combining physics-based modelling with data-driven machine learning and probabilistic uncertainty assessment. The underlying idea is to combine well-established and robust physics-based full order models (FOM), that are made effective by reduced order modelling (ROM), and use of probabilistic data-driven models to both increase the accuracy as well as focus simulation efforts where the information gained produce the most value with respect to the relevant decision context. Creating hi-fidelity insights for better real-time decision-support in complex and high-risk systems requires innovative and new ways of combining data-driven and physics-based modelling, beyond current approaches (e.g. hybrid modelling, multi-fidelity modelling, etc). Integrating these model domains in a way that ensures both that the computational speed and the overall uncertainty is acceptable is the biggest research challenge of the project. This will enable the asset owners to make faster and better safety-critical decisions and reduce (overly conservative) restrictions on operating windows, maintenance and inspection intervals, etc. Increasing uptime, more optimal production, and avoiding safety critical incidents, if only by a few percent, has a huge value for the asset owners. “Conventional finite-elements simulations appear to be conservative compared with what we observe from sensor data. However, models based on sensor data do not capture safety-critical behaviour which we have not yet experienced. In effect, neither model is able to give proper decision support and we are left with conservatism to offset the uncertainty”. Paraphrased statement from asset owner operating on the NCS