Portuguese 
English 
Spanish 
4 min of reading 
0 comments 
Transfer learning technique reduced the need for expensive cosmological simulations by more than ten times, but study shows that the same prior knowledge capable of accelerating the search for new laws of physics can hinder the identification of truly novel signals in the universe.
Artificial intelligence can accelerate the search for new laws of physics by reducing the use of expensive cosmological simulations by more than ten times, but the gain comes with a risk: relying too much on already learned patterns.
AI uses prior knowledge to search for laws of physics
The research examined how transfer learning can help cosmologists investigate theories beyond the standard cosmological model, known as ΛCDM. The study was published in the Journal of Cosmology and Astroparticle Physics, with the article available in JSTAT.
The ΛCDM model explains large-scale features of the universe, such as its expansion and the distribution of galaxies. Even so, scientists assess that it does not represent a definitive answer for all observed phenomena.
-
China has deployed snake-shaped robots to slide along power lines, where they have already inspected over 130 kilometers of cables using cameras and sensors to detect damaged wires, worn-out parts, and overheating.
-
Students create a Lego boat to collect plastic from beaches, focus on clean energy, and aim to combat marine pollution.
-
Scientists create an artificial photosynthesis system that “mimics” plants, transforms water and CO2 into solar fuel, and even works without relying on expensive batteries to control the energy.
-
Industry rumors suggest that Apple might equip the iPhone 21 with a 200-megapixel ultrawide camera around 2028, a leap that would bring the brand closer to Android rivals, although the sources themselves admit that the change is still uncertain.
Recent observations have raised questions associated with massive neutrinos, modified gravity, and evolving dark energy. Each possibility requires detailed simulations, in which virtual universes are constructed with different physical assumptions.
This process is computationally expensive. To test many hypotheses, researchers need to generate a large number of complex simulations, which demands substantial processing power and increases the cost of analyses.
How transfer learning reduces simulations
Transfer learning allows a neural network to leverage knowledge acquired in one task and apply it to another related task. The team used this principle to first train the AI on simpler simulations.
In this initial stage, called pre-training, the neural network learned patterns from simulations based on ΛCDM. Then, it received additional training with more sophisticated models, which include possible physical phenomena not yet incorporated into the standard model.
Adrian Bayer, a cosmologist at the Flatiron Institute and Princeton University, described the strategy as a shortcut. Instead of training the AI directly on the most costly simulations, the system starts with simple models and progresses to the complex ones.
Bayer compared the method to the use of textbooks. First, the student reads a basic book to form a general notion of knowledge; then, they move on to more complicated and specialized material.
For Veena Krishnaraj, the first author of the study and an undergraduate student at Princeton University, this strategy prevents the AI from having to assimilate everything at once. The result was a significant reduction in the most expensive simulations.
In some cases, transfer learning reduced by more than ten times the number of costly simulations needed to investigate parameters related to new physical possibilities in the universe.
When the Shortcut Hinders Discovery
The study also identified a problem called negative transfer. It occurs when the AI’s prior knowledge, instead of helping, leads the system to interpret new signals as if they were variations of already known patterns.
The difficulty appeared in simulations with massive neutrinos. Some observational signatures associated with neutrino mass resemble changes linked to the ΛCDM σ8 parameter, which measures the intensity of matter clustering in the universe.
Due to this similarity, the pre-trained neural network initially struggled to separate the two effects. The AI analyzed unknown information based on familiar references, which could mask evidence of new laws of physics.
Krishnaraj stated that negative transfer is not random, but driven by underlying physical degeneracies in the model. This means that different processes can produce very similar observable signatures.
Promise for Future Astronomical Surveys
The results highlight the benefits and limits of using concepts inspired by fundamental models in physics. Pre-training can accelerate inference but also hinder the learning of new physical concepts.
So far, the approach has been tested in simulations. The next step will be to apply it to real astronomical observations, in the context of cosmological surveys that are expected to gather unprecedented volumes of high-precision data.
The team sees transfer learning as an important tool for future cosmology, provided its risks are considered. Comment on what you think of this balance between speed and caution: should AI gain more ground in the search for new laws of physics, or does this type of limitation show that human oversight will remain indispensable?
Be the first to react!