|
|
|
Development of Active Learning Capability for Numerical Simulations
|
|
|
| We propose to develop efficient machine learning techniques that will enable vast improvements in large-scale numerical simulations in the space sciences. Simulations play a fundamental role in studies by scientists and engineers across NASA, DoE, NOAA, FAA, industry, and academia. In many cases, simulations provide the means to examine processes that could otherwise be infeasible or impossible to study. Here, we team domain experts (scientists expert in particular simulations) with computer scientists who bring to bear cutting-edge research in computing techniques and technology. We will employ novel active-learning techniques, for which we have already shown proof-of-concept. In a just-completed NASA Intelligent Systems grant, we demonstrated that active learning can make running suites of numerical simulations dramatically more efficient. Here, we seek to mature that technology to the extent that it can be presented for funding to a NASA OSS research program. To do this, we choose as our primary application the simulation of asteroid collisions, resulting in creation of asteroid satellites. This application will demonstrate the depth of our methodology, and when complete, will allow a substantial increase in the speed of learning from simulators. We will also demonstrate the wide-ranging potential applicability of our active learning research by development under two additional simulation areas --- understanding of data on Earth's magnetosphere, and understanding of the structure of the atmosphere of Titan, a moon of Saturn. Our three applications are very different from one another, and thus will demonstrate the power of the technique, in terms of flexibility and enhancement of science understanding. |
|
|
