Stipendiatstilling 17 SINTEF (2021-2023)

Modeling and simulation of physical systems provide insights into their design, troubleshooting, and predictive maintenance. Traditionally, numerical and analytical methods have been the primary approaches, but they often require simplifying assumptions to remain computationally feasible or they become prohibitively expensive for highly complex systems. Machine learning became a promising alternative, leveraging the increasing availability of observational data. However, in many cases, data are sparse, noisy, or unrepresentative of the entire system, leading to unreliable predictions and limiting the applicability of these methods. This Ph.D. project advanced physics-informed machine learning (PIML), which bridges data-driven and physics-based approaches, to enhance predictive accuracy, interpretability, and efficiency in scientific and engineering systems. The key contributions fall into two main areas. The first focuses on physics-based feature engineering and anomaly detection, where we developed an interpretable fault detection algorithm for maritime systems, validated on real-world operational data. Additionally, we proposed a structured framework for anomaly detection that incorporates expert validation, improving reliability and interpretability in industrial monitoring applications. The second area aims to enhance neural operators, encompassing multiple developments. One aspect of this work involved extending neural operator models with recurrent networks, enabling stable long-term simulations. Another key contribution was the development of a novel training strategy that combines low- and high-resolution data, improving predictive accuracy while reducing reliance on high-resolution datasets. Additionally, we explored ways to enforce physical consistency in neural operators through hard and soft constraints, ensuring compliance with fundamental conservation laws. Each contribution demonstrated the potential of physics-informed machine learning in tackling real-world scientific and engineering challenges, improving the reliability and efficiency of data-driven models.

Prosjektet leverte både fremskritt innen fysikk-informert maskinlæring og eksempler på praktiske anvendelser. Det bidro til videreutvikling av nevrale operatorer for bruk i prediksjoner over lengre tidsserier, som ble publisert gjennom to artikler. Forskningen utdypet forståelsen av flere hybrid-KI-tilnærminger ved å sammenligne deres styrker og begrensninger på tvers av ulike anvendelser, noe som førte til bredere innsikt i hybride modelleringsstrategier. I tillegg var et direkte mål å utvikle anvendte algoritmer, noe som ble oppnådd med en feildeteksjonsmodell. Denne har vært kommersielt implementert på Brunvoll-fartøy siden 2021. Gjennom forskningsopphold ved Brown University, flere besøk og deltakelse på felles workshops, ble det etablert flere internasjonale samarbeid for SINTEF.