Development and assessment of the data-informed continuous machine learning approach based on CHF prediction

Abstract

Machine learning (ML) method has attracted more and more interests in engineering applications. Despite extensive efforts in the last decades in the application of the ML-method to thermal- and fluid mechanics, there exist in general some obvious shortcomings. Generally, neither sufficient information about the data base used by the previous researchers nor information about the uncertainty (or error) is included in the ML-model and is not available for the next researchers. This makes the continuous learning process difficult or even impossible.

This paper proposes the new data-informed continuous machine learning (DI-CML) approach, to overcome the above shortcomings. The main feature of the DI-CML approach is to generate a machine learning package, which, in addition to the ML-model, contains the distribution functions of the input variables and the distribution function of the uncertainty (error). With this ML-package, an artificial data base can be produced, which should be as similar as possible to the original data base used for the development of the ML-model. This would make the continuous learning process possible and efficient.

The main idea and the procedure of the DI-CML approach is described. The feasibility of the DI-CML approach is assessed by means of CHF prediction. The large CHF data base provided by the OECD-NEA benchmark working group is used. The accuracy of the CHF prediction by the DI-CML approach is analysed by using different features of data base sets, different methods to derive the distribution functions of the input variables as well as different methods for the generation of the artificial data base. The results confirm the good feasibility of the proposed DI-CML approach. Furthermore, challenges and future research needs are also identified.