AI and why astronomers no longer look through a telescope

As a child, I had the opportunity to use the two-meter Faulkes telescope in Siding Spring, Australia. I remember it was a very exciting experience to see the telescope move through a live webcam. I entered the coordinates of the sky and, within a minute, the telescope was slowly tilting towards the object. I looked forward to becoming an astronomer and spending countless nights at the sites of real telescopes – Hawaii, Chile, Spain, these are all places with powerful telescopes that I wanted to use one day.

But I did not know that the era of manual observations was coming to an end.

My first professional research experience occurred when I was hired as an intern at the Las Cumbres Observatory, headquartered in Santa Barbara, California, my hometown. Although it is called an observatory, no telescope is actually located there. Las Cumbres includes more than 30 instruments around the world, with the mission of always having a telescope in the dark, ready to make observations at all times.

Their telescope system is very simplified and is an excellent example of how astronomical observations work today. The telescopes are all robot-controlled, requiring almost no human intervention except for some engineers responsible for maintaining the instruments. Scheduling is also robotically controlled, and while astronomers may request observations during certain nights, observations are ultimately booked and controlled by an automated system that is programmed to select the best targets, based on instrument availability and weather conditions at all sites; an analysis that takes a few seconds.

It is indeed becoming increasingly rare for astronomers to visit sites and make actual observations, as manual control of telescopes, when necessary, is often done remotely from the comfort of one’s home; and all collaboration is done simply via zoom. Gone are the days of Edwin Hubble, who actually lived for long periods on the mountain, observing night after night, collecting data plates one at a time.

While I think it’s a little sad that such days are behind us, there’s a silver lining to it all. With technological advancements in telescope instrumentation and software capabilities, it has become necessary to review and streamline the observing process. We had no other choice, because man is prone to error, yes, but he also needs to sleep, eat and socialize, whereas a machine can run non-stop for decades as long as that it is supplied with electricity. As an astronomer, I value science above all else and would gladly choose an increase in quality data over the ability to live on a mountain and do manual observations, however mystical and dreamy as it could have been.

Machine learning has been used as a tool to optimize manual procedures since the 20th century, in fact my colleagues always joked that the USPS mail system has been using machine learning since the 1960s to help sort mail, so that astronomers have just recently come to appreciate its application in our field due to the Big Data transition in which we find ourselves. The James Webb Telescope, for example, will produce more data than it will ever be possible to inspect visually. There are images today that no one in this world will ever see, and with new telescopes coming into service in the coming years, more images will fall into the category of data inspected only by emotionless machines.

Astronomers using NASA's James Webb Space Telescope have combined the capabilities of the telescope's two cameras to create a never-before-seen view of a star-forming region in the Carina Nebula.  Captured in infrared light by the Near Infrared Camera (NIRCam) and Mid Infrared Instrument (MIRI), this combined image reveals previously invisible star birth zones.  What looks a lot like craggy mountains on a moonlit evening is actually the edge of a nearby young star-forming region known as NGC 3324. Called the Cosmic Cliffs, this edge of he gigantic gas cavity is about 7,600 light-years away.

In short, machine learning is the use of differential calculus to identify optimal patterns in high-dimensional data. A 50 x 50 pixel image, for example, can be represented in a space of 2500 dimensions. But what are these “optimal patterns” that the machine identifies? Unfortunately, there is no answer to this. Really, machine learning is often considered a dark art. Even if we could visualize the connections made by my machine learning engine, it would have been futile, because ultimately we wouldn’t have figured this out. The correlations the machine has found through many iterations are just too complex for our minds to comprehend. It’s really a black box – data is coming in, we don’t know what connections the machine learned during training, but good results are coming out and we’re happy.

As astronomers, we use machine learning for object recognition, signal predictions, and even as a tool to manage our instruments. It would take me my whole life to inspect two million astronomical objects, but I have a machine learning algorithm that did it for me in less than 30 minutes. These developments have led to the creation of brokerage systems that take data from the telescope, apply machine learning to distinguish certain objects, and then relay the information to science teams interested in the particular phenomenon.

Much of the data of the next century will go unnoticed, even with the help of our machine learning programs; but I guess that’s a nice thing – anyone can do astronomy just by downloading public data from their computer. We need all the help we can get, because while machine learning is hugely useful, in its current state it still can’t be compared to the eyes and brains we have, which it doesn’t. there’s just not enough.

Daniel Godines is a PhD student in astronomy at New Mexico State University. He can be reached at [email protected].

Previous One day in 1950, the work of a North Adams-born Secret Service agent became much less secret | Columnists
Next Gabrielle Zevin believes games show people who they really are