Assembling Executable Scientific Models from Source Code and Free Text

Abstract

Modern science is increasingly driven by the use of computational models of complex systems implemented in software. This provides enormous opportunities, including increased precision in predictions, the possibility of better control over reproducibility of results, and the ability to aggregate models from different domains for larger-scale modeling. However, realizing the full potential of computational scientific modeling requires addressing a number of new challenges. A central challenge is that scientific model software is generally implemented in code bases that do not follow uniform conventions, so that model interpretation, selection and integration still requires very labor-intensive manual curation. In this thesis we introduce the Unified Model Assembly Framework (UMAF), which aims to assist domain scientists with the process of extracting a representation of models from source code that permits grounding variables to domain concepts and performing model analysis. To accomplish this task we first provide a formal structure for expressing scientific models that can be extracted from source code. Using this information we then show how an executable scientific model can be developed in the form of a dataflow program. The thesis concludes with an evaluation of the current capabilities of the UMAF framework on a real-world crop modeling code base, focusing on the kinds of source code idioms the framework can currently translate into dataflow-based scientific models and the accuracy of translation.