AI has recently begun to emerge as one of the most effective technologies used in cosmology. The power of machine learning technology can be seen when we analyze models that predict the evolution of galaxies and the universe. However, recent research has revealed an unexpected complication. When researchers trained an AI model based on the standard cosmological framework that describes the universe, the system became highly efficient at recognizing familiar patterns, but significantly slower at detecting truly new physics. This discovery highlights the growing challenges in scientific AI. In other words, the very knowledge that makes these systems powerful can also create resistance to innovative discoveries. As astronomers prepare for unprecedented amounts of data from the next generation of observatories, understanding this limitation may prove to be as important as the technology itself.
How did AI learn the rules that govern the universe?
Researchers from Princeton University’s School of Astrophysical Sciences say modern cosmology is built around the lambda cold dark matter (ΛCDM) model. This framework explains the large-scale structure and evolution of the universe. This model successfully explains phenomena ranging from the formation of galaxies to the expansion of the universe, and is still widely used in cosmological research.To accelerate scientific investigations, researchers trained neural networks using simulations generated based on the ΛCDM assumption. This approach relies on a machine learning technique known as transfer learning, in which the AI first learns broad patterns from simple datasets before being adapted to tackle more specialized tasks. Including massive neutrinos, modified gravity, primitive non-Gaussianity, etc. allows inference with significantly fewer simulations beyond ΛCDM.However, we also show that negative transitions can occur if a strong physical degeneracy exists between the ΛCDM parameter and the parameters beyond ΛCDM. We considered various forwarding architectures and found that including a bottleneck structure yielded the best performance. Our findings demonstrate the opportunities and pitfalls of fundamental model approaches in physics. Although pre-training can speed up inference, it can also hinder learning new physics.The results were impressive. Scientists have discovered that transfer learning dramatically reduces the number of computationally expensive simulations needed to analyze alternative cosmological models. In some cases, this approach reduced computing demands by more than an order of magnitude, potentially saving years of processing time and significant research costs.The study’s principal investigators demonstrated that the AI system can quickly identify subtle relationships within vast cosmological datasets, making it invaluable for future projects generating petabytes of observational information.
An unexpected problem that physicists did not foresee
The same prior knowledge that made AI efficient also created significant weaknesses. The researchers found that once neural networks become familiar with ΛCDM-based patterns, they may struggle to recognize signals that deviate from those expectations. Essentially, this system created a kind of scientific bias. They tended to interpret new information based on what they had learned, rather than embracing the possibility of change and innovation.This poses a big problem for cosmologists looking for evidence of things that don’t conform to the Standard Model, such as gravitational modifications, changes in dark energy, and the effects of giant neutrinos.The researchers say that because transfer learning becomes so effective at recognizing well-known structures, it may inadvertently suppress the very anomalies scientists are trying to discover. This challenge reflects a long-standing problem in the human sciences. Scientists may begin data analysis with biases based on existing theories. Research shows that if AI models are trained based on existing paradigms, they can be subject to the same biases.
Why this discovery could shape the future of cosmology and artificial intelligence
This discovery is important for the future of astronomy because new telescopes and surveys will generate vast amounts of data in the field. Nevertheless, this study highlights that future AI models will need to be trained to keep an open mind.Therefore, instead of relying on existing theories, scientists may need to create approaches that make AI sensitive to anomalies.This challenge extends beyond cosmology. Across physics, scientists are increasingly studying AI not just as a data processing tool, but as a mechanism for uncovering entirely new laws of nature. Recent research has shown that machine learning systems can identify previously hidden physical relationships within complex plasma systems while maintaining interpretability, demonstrating the potential of AI as a true discovery engine.Physicist Justin Burton, Emory Professor of Experimental Physics and senior co-author of the paper, told Mirage about AI-driven discoveries in plasma physics:“We have shown that AI can be used to discover new physics. Our AI methods are not black boxes. We understand how and why AI works. The framework it provides is also universal. It has the potential to be applied to other many-body systems, paving the way to new discoveries.” “Physics-aligned machine learning reveals unexpected physics in dusty plasmas,” he and his co-authors wrote in their study.New cosmological research adds an important caveat to that optimism. AI can accelerate scientific discovery, but only if researchers ensure that they can maintain a system that allows them to question the assumptions they have been trained to understand.
