Development and assessment of the uncertainty-aware data-informed continual machine learning approach

For prediction of complex heat transfer phenomena, such as critical heat flux (CHF), machine learning models gain more attraction in the research community. Throughout the institutions, a large amount of experimental
data exists, but the data is often confidential, so it is not possible to create a comprehensive database. Continual training of a machine learning model, shared from institution to institution, could tackle/mitigate this problem.
Recently, a data-informed continual machine learning (DI-CML) approach has been developed. In the present work, this method is enhanced by considering the uncertainty of the machine learning model to filter the soft data, which is generated in that approach. The Monte Carlo Dropout method, deep ensembles and the stochastic weight averaging Gaussian (SWAG) are used to approximate the Bayesian uncertainty. Assuming that machine learning models provide low uncertainty if applied to data overlapping with the data used in its training, the generated soft data is first filtered for low uncertainty and then used to continue the training of the machine learning model. Doing so, the statistical validity domain of the trained model is identified without reconstructing or disclosing the original experimental distribution. ... mehrBesides the parameter ranges, no additional information about the training data needs to be shared. Furthermore, the proposed method maintains predictive performance while reducing the amount of shared information compared to the DI-CML method.