Astronomers speed up search for habitable worlds

Astronomers have discovered thousands of planets, including dozens that could potentially support life. To weed them out, their atmospheres need to be studied. Artificial intelligence and machine learning can help.

Modern astronomy would be hard-pressed to survive without AI and machine learning (ML), which have become indispensable tools. Only they can manage and work with the vast amounts of data that modern telescopes generate. ML can sift through large amounts of data, looking for patterns that would take humans much longer to find.

The search for biosignatures on Earth-like exoplanets is a vital part of modern astronomy, and ML can play a major role in it.

Because exoplanets are so distant, astronomers pay close attention to those that can perform transmission spectroscopy. When starlight passes through a planet’s atmosphere, spectroscopy separates the light into different wavelengths. Astronomers then examine the light for signs of certain molecules. However, chemical biosignatures in exoplanet atmospheres are very complex, since natural, non-biological processes can generate some of the same signatures.

This is a model of the Webb transmission spectrum for a planet like Earth. It shows the wavelengths of sunlight that molecules like ozone (O3), water (H2O), carbon dioxide (CO2), and methane (CH4) absorb. The y-axis shows the amount of light blocked by Earth's atmosphere, not the brightness of the sunlight passing through the atmosphere. The brightness decreases from bottom to top. Understanding Earth's spectrum helps scientists interpret the spectra of exoplanets.

Although this method is powerful, it does have some challenges. Stellar activity such as meteor showers and flares can contaminate the signal, and the light from the atmosphere can be very weak compared to the light from the star. If the exoplanet has clouds or haze in its atmosphere, this can make it difficult to detect molecular absorption lines in the spectroscopic data. Rayleigh scattering adds complexity, and there can be several different interpretations of the same spectroscopic signal. The more of these types of “noise” in the signal, the worse the signal-to-noise ratio. Noisy data—data with a low signal-to-noise ratio—is a distinct problem.

We are still discovering different types of exoplanets and planetary atmospheres, and our models and analysis methods are not complete. Combined with the problem of low signal-to-noise ratio, this pair represents a serious obstacle.

But machine learning can help, according to a new study. The work “Machine classification of potential biosignatures on Earth-like exoplanets using low signal-to-noise transmission spectra” is under review in the journal Monthly Notices of the Royal Astronomical Society. The lead author is David S. Duque-Castaño of the Computational Physics and Astrophysics Group at the University of Antioquia in Medellin, Colombia.

Webb is our most powerful transmission spectroscopy instrument, and it produces impressive results. But there’s a problem: observing time. Some observations take a huge amount of time. Detecting things like ozone can require an inordinate number of transits. If we had unlimited observing time, this wouldn’t matter so much.

One study found that, in the case of TRAPPIST-1e, up to 200 transits may be needed to obtain statistically significant detections. The number of transits becomes more reasonable if the search is limited to methane and water vapor. “The studies showed that with a reasonable number of transits, the presence of these atmospheric species, which are commonly associated with the global biosphere, can be detected,” the authors write. Unfortunately, methane is not as reliable a biosignature as ozone.

Given the time it would take to detect some of these potential biomarkers, the researchers say it might be better to use Webb to conduct signal-to-noise studies. “While this may not yield statistically significant results, it would at least allow planning follow-up observations of interesting targets with current and future more powerful telescopes (e.g., ELT, LUVOIR, HabEx, Roman, ARIEL),” the authors write, citing the names of telescopes that are under construction or planned.

The researchers have developed a machine learning tool to help solve this problem, which they say could speed up the search for habitable worlds by harnessing the power of artificial intelligence. “In this work, we developed and tested a general machine learning methodology to classify low-signal-to-noise emission spectra according to their potential to contain biosignatures,” they write.

Because most exoplanet atmospheric spectroscopy data is noise, the MO tool is designed to process it, determine the noise level, and classify atmospheres that may contain methane, ozone, and/or water or be interesting enough for follow-up observations.”

The team created a million synthetic atmospheric spectra based on the known planet TRAPPIST-1 e, and then trained their MO models on them. TRAPPIST-1e is similar in size to Earth and is a rocky planet in the habitable zone of its star. “The TRAPPIST-1 system has attracted considerable scientific attention in recent years, especially in the fields of planetary science and astrobiology, due to its exceptional characteristics,” the paper says.

An image of the rocky exoplanet TRAPPIST-1e, similar in size to Earth.

The star TRAPPIST-1 is known for hosting the largest number of rocky planets of all the systems we have discovered. According to the researchers, it is an ideal candidate for training and testing MO models, since astronomers can obtain favorable signal-to-noise readings in a reasonable time. The planet TRAPPIST-1e likely has a compact atmosphere similar to Earth. The resulting models were successful and correctly determined the transmission spectra with suitable SNR levels.

The researchers also tested their models on realistic synthetic atmospheric spectra of modern Earth. Their system successfully identified synthetic atmospheres containing methane and/or ozone in ratios similar to Proterozoic Earth. During the Proterozoic, the atmosphere underwent fundamental changes due to oxygen catastrophe.

It changed everything. It allowed the ozone layer to form, created conditions for complex life to flourish, and even led to the formation of the vast iron ore deposits we mine today. If photosynthetic life evolved on other exoplanets, their atmospheres should be similar to Earth’s Proterozoic atmosphere, making it a relevant marker for biological life. (The recent discovery of dark oxygen has major implications for our understanding of oxygen as a biomarker in exoplanet atmospheres.)

In their paper, the authors describe the detection of oxygen or ozone as the “jewel” of exoplanet spectroscopic signatures. But abiotic sources exist, too, and whether oxygen or ozone is biotic may depend on what else is present in the signature. “Specific spectral fingerprints can be sought to distinguish biotic from abiotic O2,” the authors write.

To evaluate the effectiveness of their model, they need to know which exoplanet atmospheres are correctly identified and which are false.

The results also had to be classified as true positives or true negatives, in relation to the accuracy of the measurements, or false positives or false negatives, in other words, errors. To organize the data, they created a classification system they called a confusion matrix.

“We have introduced the category 'interesting' in the diagram to highlight planets that are worth further observation or in-depth analysis,” the authors explain. “It should be reiterated that the focus of this work is not on detecting biosignatures using ML, but on labeling planets that are interesting or not.”

The confusion matrix has four classifications – true positive, false negative, false positive, true negative

One model successfully identified probable biosignatures in the spectra of Proterozoic Earth after a single transit. Based on their testing, they explain that Webb will successfully detect most of the “habitable terrestrial planets observed by JWST/NIRSpec PRISM around M-dwarfs located at distances close to or smaller than TRAPPIST-1 e.” If, that is, they exist.

These results could help guide Webb's future work. The researchers write that “machine-assisted strategies like those presented here could significantly optimize the use of Webb's resources to search for biosignatures.” They could streamline the process and increase the chances that follow-up observations will find promising candidates. The telescope has been operating for two years and seven months of its planned five-and-a-half-year primary mission. (The telescope could operate for up to 20 years overall.) Anything that can optimize the space telescope's precious observing time is a win.

Overall, the study presents a machine learning model that can save time and resources. It quickly scans the atmospheric spectra of potentially habitable exoplanets. While it doesn’t determine which ones contain biomarkers, it can identify the best candidates for follow-up after just 1-5 transits, depending on the type of atmosphere. Some types will require more transits, but the model still saves time.

“Identifying a planet as 'interesting' will make the allocation of observing time from valuable resources like Webb more efficient, an important goal in modern astronomy,” they write.