Image Courtesy: Francisco Villaescusa-Navarro
Artificial intelligence could dramatically reduce the time and computing power needed to explore some of the universe’s biggest mysteries, but new research suggests the technology may also struggle to recognize discoveries that fall outside its existing knowledge.
A study published in the Journal of Cosmology and Astroparticle Physics examined how a machine learning technique known as transfer learning can help cosmologists investigate theories that extend beyond the current standard model of the universe. The approach significantly reduced the need for expensive simulations, but researchers also found that prior training can sometimes make AI less effective at identifying genuinely new physical phenomena.
The research focuses on the standard cosmological model, known as ?CDM, which explains many large-scale features of the universe, including its expansion and the distribution of galaxies. While the model remains highly successful, scientists believe it does not fully explain reality, prompting investigations into concepts such as massive neutrinos, modified gravity, and evolving dark energy.
Testing these ideas requires creating vast numbers of detailed computer simulations, each representing a different version of the universe. Generating such simulations can consume enormous computational resources, making the process both costly and time-intensive.
To address that challenge, researchers explored transfer learning, a technique that allows AI systems to apply knowledge gained from one task to another related problem. Instead of training neural networks exclusively on complex simulations containing new physics, the team first trained the systems using simpler ?CDM-based simulations before introducing more advanced models.
The results were substantial. In some scenarios, transfer learning reduced the number of computationally expensive simulations required by more than tenfold, potentially lowering costs and accelerating future cosmological research.
However, the study uncovered an important limitation. Researchers observed a phenomenon known as negative transfer, where previously learned information interferes with the ability to identify new patterns. In certain cases, signatures of new physics closely resembled effects already associated with the standard cosmological model, causing the AI to interpret unfamiliar signals through the lens of what it had already learned.
One example involved simulations containing massive neutrinos. Some of their observable effects appeared similar to changes in an established cosmological parameter known as ?8, which measures how strongly matter clusters across the universe. Because of that overlap, the pretrained AI initially struggled to distinguish between the two explanations.
The findings highlight a broader challenge facing AI applications in science. While foundation model approaches can accelerate analysis and reduce computational demands, they may also introduce biases that limit the discovery of unexpected phenomena. For fields such as cosmology, where researchers are actively searching for evidence that existing theories are incomplete, that tradeoff could prove particularly significant.
The team plans to move beyond simulations and test the approach using real astronomical observations. With next-generation cosmological surveys expected to generate unprecedented volumes of high-precision data, transfer learning could become a valuable tool for managing the flood of information while helping scientists search for clues about the fundamental nature of the universe.
