Provenance-integrated adaptation of numerical approximations of differential equation models

We explore how to leverage metadata and provenance collected a priori and during the execution of a simulation of a differential equation model, with the goal of using these to adapt, improve, and predict the simulation. Such metadata include ‘low-level’ performance metrics obtained by monitoring, e.g., convergence rates, runtime, and memory consumption. By analyzing such provenance relying on data mining, knowledge discovery, and machine learning methods and combining it with mathematical and domain knowledge, further potentially more costly ‘high-level’ measures and metrics can be derived that enable predictions of the estimated difficulty of a solution or the similarity of tasks and suitability of components in different simulations.

We continue to study two fundamental problems: deciding when a (possibly expensive) provenance capture enables more informed design decisions; and adapting parameter settings to improve performance. This requires the investigation of methods and measures defining high-level provenance and algorithms leveraging it, as well as execution strategies of a. We also plan to explore provenance management and use when multiple simulations are involved. The results of this project inform multi-adaptive simulations.