What can AI learn about the universe?

Artificial intelligence and machine learning are widespread, with applications ranging from data analysis, cybersecurity, drug development, music composition, and artistic rendering.

In recent years, large-scale language models (LLMs) have also emerged, adding human interaction and writing to the long list of applications. This includes ChatGPT, an LLM that has had a huge impact since its introduction less than two years ago. This application has generated considerable discussion (and controversy) about the potential uses and impacts of AI.

Astronomy will also benefit greatly, using machine learning to classify large amounts of data, look for signs of planetary transits, correct for atmospheric interference, and find patterns in noise. According to an international team of astrophysicists, this could be just the beginning of what AI can do in astronomy.

In a recent study, the team used astronomical observations to fine-tune a generative pre-trained transformer (GPT) model. In the process, they succeeded in demonstrating that his GPT model can effectively support scientific research.

The research was carried out by the International Network for Relativistic Astrophysics (ICRANet ) was carried out by. China, Chinese Academy of Sciences Institute of High Energy Physics (CAS-IHEP), University of Padova, Isfahan University of Technology, Ferrera University.

Their paper “Can AI understand our universe? Fine-tuning tests of GPT with astrophysical data” was recently published. arXiv Preprint server.

As mentioned earlier, astronomers rely heavily on machine learning algorithms to classify the large amounts of data acquired by modern telescopes and instruments. This practice began about a decade ago and has since grown exponentially to the point where AI is integrated into the entire research process. Yu Wang, ICRA president and lead author of the study, told Universe Today via email:

“Astronomy has always been driven by data, and astronomers were among the first scientists to introduce machine learning. Machine learning is now being integrated into the entire astronomical research process, from manufacturing and control on the ground and in space. from telescopes (e.g. optimizing the performance of adaptive optics systems, improving the initiation of certain movements (triggering) of satellites under certain conditions, etc.) to data analysis (e.g. noise reduction, data completion, classification). , simulations, etc.), and the establishment and validation of theoretical models (e.g., tests of modified gravity, constraints on the equation of state for neutron stars, etc.).

Data analysis remains the most common of these applications, as it is the easiest area in which machine learning can be integrated. Traditionally, dozens of researchers and hundreds of citizen scientists analyzed large amounts of data generated by observation campaigns.

However, this is not realistic in an age where modern telescopes collect terabytes of data every day. This includes all-sky surveys such as VLASS (Very Large Array Sky Survey) and many stages carried out by the Sloan Digital Sky Survey (SDSS).

Since LLM is a relatively recent creation, it has so far been applied only sporadically in astronomical research. But supporters like Wang say it will have huge social implications, with the potential to be on the lower end of the “industrial revolution.”

As for the upper limit, Wang predicts there is a wide range, perhaps leading to humanity's “enlightenment or destruction.” However, unlike the industrial revolution, the pace of change and integration of AI is much faster, raising questions about how far its adoption will go.

To determine its potential in the field of astronomy, he and his colleagues took a pre-trained GPT model and fine-tuned it to identify astronomical phenomena, Wang said. Masu.

“OpenAI provides a pre-trained model, and what we did was fine-tune it, change a few parameters based on the original model to recognize astronomical data and This is similar to what OpenAI gives us. We then trained them to become astronomy graduate students.

“We provided limited data at a moderate resolution and reduced the number of training times for GPT compared to regular models.Nevertheless, the results are impressive, achieving around 90% accuracy. This high level of accuracy is due to GPT's robust foundation, which already understands data processing and has logical reasoning and communication skills.

To fine-tune the model, the team introduced observations of different astronomical phenomena from different catalogs. This includes his 2,000 samples (500 each) of quasars, galaxies, stars, and broad absorption line (BAL) quasars from SDSS. We also integrated observations of short and long gamma-ray bursts (GRBs) and simulations of galaxies, stars, and black holes. When tested, their model was able to classify different phenomena, distinguish between types of quasars, infer distances based on redshifts, and measure the rotation and tilt of black holes.

“This study shows that at least LLM can process astronomical data,” Wang said. “Additionally, the model's ability to process different types of astronomical data is a capability that other specialized models do not have. We hope that we can identify common underlying principles that will help us. Of course, this is a difficult task and cannot be achieved by astronomers alone.”

Of course, the research team acknowledges that the dataset they experimented with is very small compared to the data output of modern observatories. This is especially true for next-generation facilities like the Vera C. Rubin Observatory, which recently introduced his LSST camera, the world's largest digital camera.

Once operational, Rubin will conduct a 10-year Space-Time Heritage Study (LSST) and is expected to yield 15 terabytes of data per night. Meeting the demands of future campaigns will require improvements and collaboration between the observatory and his specialized AI company, Wang says.

Nevertheless, it is a foregone conclusion that there will be more LLM applications in astronomy in the near future. This is not only a possible development, but also an inevitable one, given the vast amounts of data being generated by today's astronomical research. And this is likely to increase exponentially in the near future, making AI indispensable in academic fields.