Ad Astra: KCNSC employee's algorithm leads to cosmic discovery
The night sky has long been a source of inspiration for scientists, artists and poets, posing stellar mysteries that appear just out of human reach. For Patrick Aleo, the moment the cosmos first felt real to him is a distinctive memory — one that ultimately set him on a path to uncovering some of the universe’s secrets.
“I remember looking at a full moon through a telescope my parents bought my brother,” said Patrick. “Before that, the night sky seemed like something far away that just existed, but through a telescope, it was bright. It felt tangible for the first time.”
As Patrick’s interest in life beyond Earth grew, so did his understanding, helping his curious mind make sense of previously heady concepts, like galaxies and light years.
“In fifth grade, I started telling people I wanted to be an astrophysicist,” he said with a laugh.
Patrick did follow his interest in space, pursuing his doctorate in astronomy at the University of Illinois at Urbana-Champaign. Although he currently works as a Senior Data Scientist for the Kansas City National Security Campus (managed by Honeywell FM&T), his journey back to earth-side intricacies initially began with a galactic discovery.
As he uncovered new opportunities within his studies at the University of Illinois, Patrick realized he was innately drawn to celestial events that were rare or not well understood. Through his program, he began developing models to aid the detection and classification for more anomalous occurrences, including exploding stars, or “supernovae.”
Due to the sheer number of stars in the universe, Patrick realized he needed a way to sift through the noise of those that regularly brighten and dim due to internal influences — known as variable stars — and zero in on the ones nearing the end of their lifecycle. One useful clue to the onset of a supernova event is a change in brightness, with a dying star often emitting luminosity that can be comparable to that of an entire galaxy.
“When these stars explode, they get so bright that you can see them not just within our galaxy, but from far away galaxies,” Patrick explained. “But with a telescope, they just look like a region of sky that gets brighter and dimmer, so I needed a way to disentangle the normal stars that vary in brightness from these exploding stars.”
Because most stars are not explodign at a given time, only a small number of supernovae are observed annually. However, simply locating these dying stars was only part of the challenge.
“Within supernovae, there are subsets of exploding stars,” he said. “These unique populations can be more or less common. To me, the less common ones are the most interesting — there are less out there, so less of them have been studied.”
These special subsets, Patrick continued, are often the product of unusual systems that scientists are still studying and may not fully understand yet. But, once a star explodes, there’s no way to review the factors that ultimately led to its death. Therefore, successfully identifying these hyper-unique supernovae events prior to their explosion — when they demonstrate anomalous behavior that can be observed and recorded — presents distinctive opportunities for astronomers to learn more about star formation scenarios, galaxy evolution or even new physics.
Patrick could determine star behaviors that indicated an impending supernova’s uniqueness, and he also knew catching these patterns at their commencement was key. However, manually finding individual, legitimate anomalies at their onset was time consuming and chancy. In the Information Age, consumer likes regularly inform other suggested products — what made supernovas exempt from these types of automated similarity searches?
This gave him an idea: what if, by using industry code that recommends music based on user preferences, artificial intelligence could help pinpoint supernova onsets that met his requirements for a study-worthy event?
Enter the Light curve Anomaly Identification and Similarity Search (LAISS). By leveraging open-source Spotify code in combination with a random forest classifier algorithm, Patrick created code to help locate emerging supernova that exhibited the behaviors he wanted to focus on. In practice, the program scans vast volumes of data (read: the universe) to identify anomalous events based on supernovae light curve and galaxy information. Then, LAISS can use the characteristics classified as anomalous to compare and quickly zero in on other, similar events, known in data analytics as “approximate nearest neighbors.” In the same way that Spotify can help listeners locate new artists that showcase traits similar to their favorite bands, the algorithm could now help Patrick consistently identify triggered supernovae that demonstrated the unique qualities he was interested in studying.
It was through LAISS that in February 2024, Patrick was clued into intriguing star behavior occurring more than 730 million light years away.
“Sometimes, toward the end of a star’s life, it can go through periods of instability, and there can be big outbursts in the years prior to its explosion,” he said. “When the star finally explodes, not only do you get a big brightness spike, but as the explosion expands, it does so through previously ejected clouds of dust and gas. Based on how active it was before it died and if it had these pre-explosions, you can see if there’s activity and if it’s breaking away before it explodes for good.”
This was the case for SN 2023zkd, a rare type of supernova identified by Patrick’s LAISS algorithm. As astronomers pieced together the picture of the star’s demise, they realized these bursts had started as early as four years prior to its explosion and that they exhibited intrinsic brightness far greater than any observed in these types of pre-explosions before.
“The signatures of this pre-explosion timeline aligned with models that indicated this type of thing would happen with an interaction between a star and a black hole,” Patrick explained. “They were spiraling around one another and eventually, the black hole overtook the star, causing the final explosion.”
The excessive intrinsic brightness detected right before the event, he added, was due to the black hole’s effect on the star, causing it to forcibly explode.
Although supernovae are routinely discovered, LAISS was the first tool that signaled SN 2023zkd’s anomalous behavior to researchers. By flagging the activity based on Patrick’s inputs, LAISS alerted him to the event following its first brightness peak, allowing his research group to begin investing resources in further exploration. When other scientists realized its noteworthiness later in its second peak of activity, Patrick’s group had already been collecting data that helped them understand why the scenario was unique.
“It’s unlikely that if my algorithm didn’t say it was interesting that people would have known about this,” he said. “They would have found out about it very late and would have missed critical information in putting it together.”
Now, as a Senior Data Scientist, Patrick has learned to apply the skillset he developed in astronomy to other types of analytics.
“In my studies, I was setting myself up to really understand a lot of big, messy data and how to gain insights from it that people haven’t learned before,” he said. “As I was finishing my Ph.D., I wanted to take the skills and methods I learned and apply them to a new domain.”He noted that he has a passion for algorithms and data science, and enjoys developing models to not only generate predictions, but also study areas or processes that may not be fully understood yet. As a member of KCSNC’s Receiving Inspection organization, which oversaw quality inspection for more than 9.1 million purchased parts in fiscal year 2024, Patrick is invested in modernizing the data landscape to help continually improve our purchased product operations. Although the first year and a half of his tenure has been largely focused on building data sets and establishing the basis for future enhancements, he sees the opportunities to drive new efficiencies beginning to materialize.
“Looking forward, now that I’m building these data sets and getting the structure set up, I can start applying these algorithms and models to specific problems,” said Patrick. “I’m really excited about this next fiscal year — I have TechMat funding scoped out for machine learning projects within Receiving Inspection flow times, so there are a lot of opportunities to make an impactful difference in service of our mission.”
With millions of piece parts produced each year at KCNSC and infinite stars in the night sky, one thing our organization has in common with the cosmos is sheer scope. In both cases, these seemingly impossible numbers become more manageable when artificial and human intelligence work together. As forthcoming technological advances at KCNSC will require interpretation, so does Patrick’s LAISS algorithm, which still runs daily.
“LAISS is still running every night, and it picks out five to 10 new events that are interesting,” he said. “There’s a human element; AI gives suggestions, and a person uses their domain expertise and knowledge to decide how we maximize resources and scientific return on investment. We need these algorithms to tell us to look at this stuff, but we still need the human piece of it within the ecosystem of machine learning.”
