Combining astrophysical knowledge and machine learning techniques can address issues that can't always be solved with domain expertise or computational methods alone, according to Eric Ford, Penn State Institute for Computational and Data Sciences (ICDS) co-hire and distinguished professor of astronomy and astrophysics.
Ford, who was joined Penn State in 2013, balances statistical, computational and machine learning methods to research in detecting low-mass, potentially Earth-like planets around stars beyond our solar system.
"I was attracted both to ICDS and the [astronomy and astrophysics] department," Ford said. "Penn State faculty played a big role in some of the first exoplanet discoveries and built up one of the top programs for exoplanets in the world. Penn State's combination of ICDS, the Center for Exoplanets and Habitable Worlds and the Center for Astrostatistics was very well aligned with my research interests."
From early on, Ford said he was interested in the exploration of planets in our solar system, but it was once he was in college that everything fell into place.
"I was in college when the first exoplanet around a sun-like star was discovered," Ford said. "I really liked that I could be almost as much of an expert as the professors because the field was so young. I really liked the vibrancy of the field... it started me on the trajectory to research exoplanets."
Ford's ongoing research project with students and post-docs is using multiple approaches to detect Earth-like planets.
"We use spectrographs to observe the spectrum of stars to deduce the velocity of the target star due to any orbiting planets," Ford said. "We're working to improve on the traditional approach so it detects the potentially Earth-like planets."
According to Form, to interpret future observations of exoplanet atmospheres and characterize their habitability, it is essential to measure the masses of these potential Earth-like planets.
"For many years, astronomers were limited by inevitable fluctuations in the star spectra due to how much starlight was collected or the variations due to the instruments used to collect data," Ford added. "With the new generation of highly stabilized spectrographs such as HPF and NEID spectrographs built at Penn State, now we're sensitive to even smaller variations due to the stars themselves."
Ford noted that stars have spots, pulsations and convective patterns that also cause fluctuations in the measured stellar spectrum.
"Machine learning is used to distinguish which variations are due to planets orbiting the star and which are due to stellar processes," Ford said. "A lot of my work involves adapting modern statistical techniques for astronomy. My approach is first to ask, 'What would we like to do in the ideal world?' and then find ways we can jump over the computational hurdles so we can answer questions using a solid statistical foundation. Sometimes that involves combining high- and low-fidelity models. Machine learning can increase the efficiency of some calculations, allowing us to devote more resources to other aspect of the problem for which machine learning isn't well-suited."
Due to the research's complexity, Ford has used multiple approaches and said he worked to stay true to the interdisciplinary research mission of ICDS and Penn State.
"One cool thing about interdisciplinary research is that researchers in fields like computer science and artificial intelligence are trying so many new things," Ford said. "They have their own goals which often aren't directly applicable to the physical sciences. But there are often nuggets of valuable insight in their work that we could adapt to benefit astronomy. The art of interdisciplinary research is that you can't do all the calculations required to try out for every idea. You have to learn a little about a lot of potential methods and pick those to spend more time on to mature. It's both frustrating and fun."
The outcome of the research and the training of future generations holds equal importance, Ford said.
"Some of the students we train become astronomers and others go on to apply their data science and computational skills to other fields such as energy and environmental monitoring," Ford said. "There's room to make advances... discovering a wide variety of exoplanets, improving our understanding of they form and learning whether our solar system is one in a million or if others like it are common. As we learn more about other planetary systems and how they formed, those insights can trickle back to inform how our own solar system formed and what about our solar system may be special."
