Machine learning models have become powerful tools in cosmology, sifting through vast simulation datasets to infer the parameters that govern the universe. But a new study reveals an ironic limitation: the very thing that makes these models efficient, learning from known physics, can blind them to genuinely new phenomena.
Researchers at Princeton University and the Flatiron Institute trained neural networks on standard ΛCDM (Lambda Cold Dark Matter) cosmological simulations, then fine-tuned them on simulations containing new physics, massive neutrinos, modified gravity, and primordial non-Gaussianities. The goal was to see whether transfer learning could reduce the enormous computational cost of running bespoke simulations for every new physics scenario.
It worked, but only up to a point.
The efficiency problem
Cosmological N-body simulations are expensive. Running a single high-resolution simulation with new physics parameters can take days on supercomputers. Training a neural network from scratch on each new simulation suite multiplies that cost.
The team, led by undergraduate first author Veena Krishnaraj and Princeton professor Peter Melchior, used a standard transfer learning approach: pre-train a fully connected neural network on a large set of ΛCDM simulations from the Quijote suite, then fine-tune it on much smaller sets of beyond-ΛCDM simulations. The models took three types of input, the nonlinear matter power spectrum, the halo mass function, and the void size function, and learned to map them to cosmological parameters.
The efficiency gain was real. Transfer learning reduced the number of required beyond-ΛCDM simulations by roughly tenfold.
The problem: negative transfer
But the shortcut came with a cost. When the team tested their models on scenarios with massive neutrinos, where the sum of neutrino masses (Mν) is a free parameter, they found that the transfer-learned models systematically misestimated both Mν and σ₈, the amplitude of matter fluctuations.
The reason is instructive. Increasing neutrino mass suppresses small-scale structure formation in a way that looks similar to decreasing σ₈. A human physicist understands the underlying physics and can distinguish the two effects. But the AI, having learned the ΛCDM pattern during pre-training, mapped the new neutrino signal onto a shift in σ₈, a familiar parameter, rather than recognizing it as a genuinely new physical effect.
This is a textbook case of negative transfer: the representations learned during pre-training became a liability because they forced the model to interpret new physics through the lens of the old.
What this reveals about machine understanding
The paper, published June 10 in the Journal of Cosmology and Astroparticle Physics (and available as an October 2025 arXiv preprint), touches on a deeper question in AI science: what does it mean for a model to “understand” physics?
A neural network trained on ΛCDM simulations learns correlations, not causal laws. When the physical assumptions change, the model cannot recognize that new causal mechanisms are at work, it only sees that the statistical relationships have shifted, and it tries to explain the shift using the concepts it already has.
The authors describe this as the model’s internal representations diverging from human-interpretable physics. The network does not “know” that neutrinos exist or that they suppress structure formation. It knows that certain input patterns map to certain output numbers, and that mapping was shaped by pre-training in a way that resists revision.
What this means for AI in science
The finding does not invalidate machine learning approaches to cosmology. Transfer learning remains useful, the authors stress that it works well for many parameters and scenarios, and the tenfold reduction in simulation cost is a meaningful gain. But the negative transfer effect is a caution: AI models used for discovery must be tested specifically where their training history could bias their inferences.
For cosmology, this is especially relevant. Upcoming surveys from the Rubin Observatory, Euclid, and the Nancy Grace Roman Space Telescope will generate unprecedented datasets, and machine learning will be essential for extracting cosmological parameters from them. The question of how to train those models, and how to ensure they remain sensitive to signals that do not fit the standard model, is not a technical detail. It is central to whether AI helps cosmologists find new physics or simply confirms what they already expect.
The study suggests a practical path forward. Bottleneck architectures, neural networks with a deliberately narrow middle layer that forces compressed representations, showed better resistance to negative transfer, possibly because the compressed encoding preserves less room for pre-training bias. And the transfer learning approach itself can be adapted: instead of fine-tuning on new data, the pre-trained model can be used as a base for more targeted training strategies that explicitly allow for new parameter dimensions.
But the deeper lesson, the authors note, is that an AI model trained within a theoretical framework should not be trusted to find phenomena outside it. The same caution applies in drug discovery, materials science, and climate modeling, any domain where machine learning is trained on existing theory and asked to find something new.
Source: Krishnaraj, V., Bayer, A.E., Jespersen, C.K., & Melchior, P. (2026). Transfer Learning Beyond the Standard Model. Journal of Cosmology and Astroparticle Physics, 2026(06), 026. DOI: 10.1088/1475-7516/2026/06/026. arXiv: 2510.19168 [astro-ph.CO]
The work was presented at the NeurIPS 2025 Workshop on Machine Learning and the Physical Sciences. The researchers are affiliated with Princeton University and the Center for Computational Astrophysics, Flatiron Institute.