For most of the past decade, artificial intelligence has played the role of a tireless apprentice in cosmology — sifting through maps of the universe, measuring how matter clumps and flows, and helping physicists squeeze meaning out of data sets too vast for any human to hold in mind. A new study suggests that the apprentice may be ready for something harder: not just learning the universe we think we understand, but helping us notice the parts we do not. The twist is that to find genuinely new physics, the AI may first have to forget some of what it was taught.
What Happened
A paper titled "Transfer Learning Beyond the Standard Model," published in the Journal of Cosmology and Astroparticle Physics on June 10, 2026, asks a deceptively simple question: can an AI trained on the universe we already know be repurposed to hunt for the universe we don't? The authors — Veena Krishnaraj, an undergraduate at Princeton University and the paper's first author, together with cosmologist Adrian E. Bayer of the Flatiron Institute and Princeton, Christian Kragh Jespersen, and Peter Melchior — set out to test whether a machine-learning shortcut called transfer learning could ease one of cosmology's heaviest computational burdens.
The recipe is intuitive. The team first trained a neural network on simulations built around the standard model of cosmology, known as ΛCDM — a step called pretraining — and only afterward adapted it to richer, more demanding models that include possible new physics. "It's basically a shortcut," Bayer explained. Rather than throwing the network straight at the most expensive simulations, the researchers let it learn the cheaper, simpler universe first, then move on to the complicated one.
The payoff was striking. Drawing on the Quijote suite of simulated universes, the team found that in some cases transfer learning cut the number of expensive simulations needed by more than a factor of ten. Krishnaraj likened the strategy to studying a hard subject by reading an introductory textbook before tackling the advanced one — a way of keeping the AI from having to "digest everything at once."
Why It Matters
To appreciate why a factor of ten matters so much, it helps to understand what cosmologists are up against. ΛCDM is remarkably successful: it describes the expansion of the universe and the distribution of galaxies with a handful of parameters. Yet almost everyone in the field suspects it is incomplete. Hints from recent observations point toward phenomena it does not cleanly account for — massive neutrinos, modified gravity, or a dark energy that evolves over cosmic time.
Testing any of these alternatives means running enormous numbers of high-precision simulations, each one a virtual universe grown from a slightly different set of physical assumptions. The computational cost is brutal, and it scales badly: every new theory to be tested multiplies the demand for supercomputer time. This is the bottleneck that quietly limits how aggressively the field can chase anomalies. If a single network can be pretrained once on the cheap ΛCDM case and then nudged toward many exotic scenarios, the savings compound across every theory a team wants to explore. In a discipline where compute budgets increasingly dictate which questions get asked, that efficiency is not a convenience — it shapes what is discoverable at all.
The Catch: Negative Transfer
The more interesting result, though, is the one the researchers did not set out to find. Transfer learning, it turns out, carries a hidden hazard they call negative transfer — moments when prior knowledge actively misleads the model.
Bayer offers a medical analogy: imagine studying from an introductory textbook and then meeting a rare disease whose symptoms mimic a common one. The familiar knowledge helps most of the time, but occasionally it steers you toward exactly the wrong conclusion. The same thing happened inside the network. Some effects produced by genuinely new physics closely resemble patterns the AI had already filed under standard cosmology, so the model reached for the categories it knew instead of recognizing something unfamiliar.
The team watched this play out in simulations with massive neutrinos. The imprint of neutrino mass on cosmic structure can look a great deal like the effect of a standard ΛCDM parameter called σ8, which measures how strongly matter clusters across the universe. The pretrained network, primed on σ8, kept misreading one as the other. "The negative transfer is not random," Krishnaraj noted; it is driven by real physical degeneracies, cases where different parameters leave nearly identical fingerprints on the sky. The very prior that makes the AI fast can also make it blind.
What's Next
For now, the method has only been tested on simulations — controlled, synthetic universes where the right answer is known in advance. The natural next step is to point it at real observational data, where there is no answer key and where the stakes of a misread signal are far higher.
The timing is deliberate. A generation of sky surveys is about to deliver an unprecedented flood of high-precision data, and analyzing it with brute-force simulation alone would be punishing. The authors see transfer learning as a tool built for exactly that moment — a way to stretch limited compute across the vast parameter space these surveys will open up. The catch is that the same surveys are where negative transfer would be most dangerous, because a model that quietly forces new signals into old categories could mask the very discovery everyone is hoping for. Learning to detect and defuse that failure mode is now part of the research agenda, not an afterthought.
Closing Thoughts
There is something quietly philosophical buried in this result. The paper is, on its surface, an engineering report about saving supercomputer hours. But it lands on a deeper tension that runs through all of science: the knowledge that lets us move quickly is the same knowledge that can keep us from seeing what is genuinely new. Expertise is a form of compression, and compression discards the unexpected.
It is telling that the same "foundation model" strategy now reshaping language and image generation — pretrain broadly, then specialize — turns out to carry this double edge when aimed at fundamental physics. Pretraining can accelerate inference, the authors write, but it "may also hinder learning new physics." That sentence could almost be a note to any expert in any field. The universe, it seems, will not give up its surprises to a mind that is too sure of what it already knows — whether that mind is silicon or flesh. The real skill, for AI and for us, may be knowing when to set the textbook down.
한글 요약
프린스턴대학과 플랫아이언 연구소 연구진이 6월 10일 Journal of Cosmology and Astroparticle Physics에 발표한 연구는, AI의 '전이 학습(transfer learning)' 기법으로 우주론 시뮬레이션의 막대한 계산 비용을 줄일 수 있음을 보였습니다. 표준 우주 모형(ΛCDM)으로 먼저 신경망을 사전 훈련한 뒤, 거대 중성미자·수정 중력 같은 '표준 모형 너머'의 복잡한 시나리오에 적응시키는 방식인데, 일부 경우 값비싼 시뮬레이션 횟수를 10분의 1 이하로 줄였습니다.
그러나 연구진은 예상치 못한 위험도 발견했습니다. '부정적 전이(negative transfer)'라 불리는 현상으로, AI가 이미 학습한 익숙한 패턴에 지나치게 의존해 진짜 새로운 물리 현상을 놓치는 경우입니다. 특히 중성미자 질량의 효과가 기존 ΛCDM 변수인 σ8과 매우 비슷하게 보여, 사전 훈련된 신경망이 둘을 혼동하는 모습이 관찰됐습니다. 이는 무작위가 아니라 물리적 축퇴(degeneracy)에서 비롯된 구조적 한계입니다.
현재까지는 시뮬레이션 단계의 검증이며, 다음 목표는 실제 관측 데이터 적용입니다. 곧 쏟아질 차세대 우주 탐사 데이터를 다루는 데 유용한 도구가 될 수 있지만, 동시에 새로운 발견을 가릴 위험도 가장 큰 영역이기도 합니다. 빠르게 해주는 사전 지식이 때로는 가장 새로운 것을 보지 못하게 만든다는 점은, AI든 사람이든 모든 탐구에 적용되는 교훈일지 모릅니다.
참고 / 출처: EurekAlert! (Sissa Medialab), Phys.org, ScienceDaily, Journal of Cosmology and Astroparticle Physics