Development of Innovative Statistical Tools to Address Data-Analytic Challenges in Physics and Astronomy

点击翻译按钮获取中文摘要

Astronomy is fertile ground for high-impact statistical challenges. The overarching goal of my new research program is to advance astronomy by developing statistical methods that incorporate physics-based computer simulators, are suited to the particular scientific and data-analytic challenges at hand, and provide uncertainty quantification. This effort involves four objectives in two application areas. To be clear, the proposed program aims to develop statistical methods to address astronomy challenges but is not a physics proposal. Hunting for Earth-like Exoplanets in the Presence of Stellar Activity. A prized goal in astronomy is the discovery of Earth-like exoplanets. A complication is that most stars exhibit activity (e.g., star spots) that can mimic a planetary signal and lead to false detections. Objective 1 is to develop a framework for detecting Earth-like exoplanets orbiting Sun-like stars. This involves developing a model selection procedure to identify stellar activity models with the highest exoplanet detection power, while also using machine learning to derive data-driven stellar activity proxies and proposing flexible statistical models to capture them. Objective 2 adapts the framework to apply to host stars that differ substantially from the Sun, using hierarchical models to pool information across similar stars and thereby learn population-level stellar activity distributions. Computer Model Emulation and Calibration with Chemical Spectra. The ChemCam instrument on the Curiosity Rover obtains chemical spectra to learn about the composition of rocks and soils on Mars. Disaggregation, i.e., determining the composition of a target, is complicated by matrix effects-interactions between chemical compounds that amplify or suppress peaks in the observed spectrum. Objective 3 is to combine computer simulators with ChemCam's spectral data and Bayesian variable selection techniques to directly solve the disaggregation problem. A challenge is that a multi-compound simulator run takes hours on modern parallel computing platforms. Objective 4 aims to overcome this limitation by first constructing fast emulators of many single-compound simulators, then using a hierarchical model to combine the fitted emulators with a few runs of the multi-compound simulator. Billions of dollars are spent developing instruments such as exoplanet-hunting telescopes and Mars rovers; comprehending and analyzing the complex datasets generated by these instruments requires sophisticated statistical methodology that is lacking. This research program will provide astronomers the tools they need to accomplish high-impact goals such as confidently detecting Earth-like exoplanets or probing the history of Mars via the composition of its terrain. The program will support the training of highly qualified personnel (HQP), providing them with in-demand skills in statistics and machine learning, and will generally grow the presence of astrostatistics research in Canada.