Aerodynamic design is nowadays conducted as a simulation-based search and optimization task whose expense is driven by the computational and/or experimental cost of the analyses evaluated at each search query. The intensive-query nature of this search and optimization task determines the preference for less expensive low-fidelity models, especially at early design stages, although those can lead to the premature exclusion of high-potential design solutions [1, 2]. To mitigate this risk, computational methods have been advanced that combine aerodynamic analyses conducted with different models spanning a spectrum of physical abstractions and evaluation cost, therefore including higher fidelity models in the library. Those methods rely on multisource strategies that virtually and variably fuse information from different models to bring higher accuracy into the exploration effort while containing the overall associated expense [3, 4, 5]. Among those, data-driven learning methods are gaining popularity that consider aerodynamic models’ evaluations at each analysis/query point as observed realizations of an event and use those data to agnostically learn a model whose predictive quality is frequently challenged along the reliability and robustness dimensions. Stemming from those considerations, this work will provide an overview of multisource (MS) methods we developed based on scientific machine learning (SciML) principles. Specifically, formulations will be discussed that embed domain and fidelity awareness in the learning scheme and provide avenues to provide predictions with reliability measures. The MS-SciML methods are discussed and compared for applications to aerodynamic design and benchmark problems of different complexity. An outlook will be proposed about the role and relevance of those methods for the development of sustainable digital twinning capabilities.