HAIEM is rooted in multiple interactions that took place in the framework of the recently finished Flanders Make ICON project MODA. Flanders Make DecisionS, FlandersMake@UGent-EEDT-DC, Bekaert and Atlas Copco were partners in MODA.  These interactions as well as those with many other companies confirmed that Flemish industry heavily relies on models of physical dynamic systems (e.g., drivetrain models, thermal models, actuator models) to realize applications such as monitoring (e.g., anomaly detectors, condition estimators, lifetime predictors), selection of (control) settings (e.g., model-based controllers) and controller designs (e.g., data driven controllers). However, physical modelling of certain quantities like friction, flexibilities, wear out to the desired level of accuracy is very time consuming. The growing availability of industrial data is an opportunity that can be exploited to overcome these shortcomings by combining physical models with AI technology. Hybrid models (HMs), for which a basic framework has been developed within the MODA project, allow physical models and AI driven models working in tandem. The key innovation is that instead of merely matching unknown parameters in a physical model with data, unknown dynamic behavior is modelled in a data driven manner.  Hybrid models have increased extrapolatability and/or transferability (hereafter referred to “transpolatability”) compared to pure AI models. They are also more interpretable due to the inclusion of expert knowledge. The limited number of degrees of freedom in the hybrid models –  compared to pure AI models – naturally leads to higher robustness, which combined with interpretability significantly increases their trustworthiness. HAIEM will further develop the MODA framework towards transpolatability. To this end, there will be a specific focus on methods to quantify extrapolatability to other operating conditions and on methods to ensure better transferability of models towards other machines. The latter will be achieved by investigating how transfer learning methods for pure AI models can be tailored to apply on hybrid models without the strong need for rich datasets. To this end, we will investigate the relation between the size and quality of the training dataset to the hybrid model performance in order to steer the collection of experiments to improve performance.