Data-driven computational model selection via recommender systems

RANS is a widely used computational tool in fluid engineering. The accuracy of a RANS calculation depends on the RANS model and the numerics of the code. As a result, experience plays an important role in selecting a RANS model for reliable results. This work aims to distill such human experience in a recommender system. A recommender system is a subclass of information filtering systems that seeks to predict the "rating" a user would give to an item. It has been well developed in commercial applications as recommending content for social media. Here, we utilize it to address the challenge of RANS model selection. We generate a simulation dataset for commonly used RANS models for 9 flows with diverse flow characteristics. Human experience is exploited in rating the performance of each model in quantities of interest (QoIs). We feed the recommender system with this rating data and train. The recommender system then gives estimates of the unknown performance of a specific model in a specific flow of interest. Detailed analysis shows that the accuracy of the prediction relies on the consistency of human experience and the richness of flow feature input.