Physics-guided discovery of dynamical dark-energy equations of state through iterative AI reasoning

This paper presents an iterative AI framework that autonomously proposes, evaluates, and refines dynamical dark-energy equations of state, successfully identifying novel phenomenological parameterizations that outperform traditional models in Bayesian evidence when tested against cosmological observations.

The Gist

The Big Picture: Teaching AI to Be a Cosmologist

Imagine the universe is expanding, and it's speeding up. Scientists call the invisible force pushing this expansion "Dark Energy." For decades, physicists have tried to write a mathematical "recipe" (an equation) to describe how this Dark Energy behaves. Usually, humans guess the recipe based on intuition, write it down, and then check if it fits the data.

This paper introduces a new way: Let an AI be the chef.

Instead of humans guessing the recipe, the authors built a system where an Artificial Intelligence (AI) acts like a creative, iterative cooking team. It proposes new recipes, tastes them against the "ingredients" of the universe (real data), gets criticized by a "food critic" AI, and then tries again to make a better dish.

The Kitchen: How the AI Works

The researchers set up a loop with three main roles:

1. The Inventor (The Generator): This is a large language model (like a very smart chatbot). Its job is to dream up new mathematical formulas for Dark Energy. But it doesn't just pull numbers out of thin air. It reads thousands of scientific papers first to understand the rules of the game, then proposes a new formula and writes a paragraph explaining why that formula makes sense physically.
2. The Taster (The Simulator): Once the Inventor proposes a formula, the system plugs it into a simulation of the universe. It checks: "If we use this recipe, does the universe look like the one we actually observe?" It compares the simulation against real data from supernovae (exploding stars), galaxy clusters, and the Cosmic Microwave Background (the afterglow of the Big Bang).
3. The Critic (The Judge): Another AI acts as a strict food critic. It looks at the formula and the Inventor's explanation. It asks: "Is this physically plausible? Is it clear? Is it stable? Is it actually new, or just a copy of something we already know?"

The Loop: The Inventor gets the Critic's feedback. If the formula was too weird or the explanation was weak, the Inventor tries again. Over hundreds of rounds, the AI "evolves" better and better equations, learning from its mistakes and successes.

The Blind Test: Can the AI Think Outside the Box?

Before testing on real universe data, the authors wanted to make sure the AI wasn't just memorizing old recipes. They created a "fake universe" with a very strange, exotic set of rules that no human had ever written down before.

• The Result: The AI didn't just copy the old, standard recipes (which failed miserably). It successfully discovered new formulas that matched the fake universe's strange behavior.
• The Lesson: The AI isn't just a parrot repeating what it read; it can actually adapt to find solutions for things it has never seen before.

The Real Discovery: Two New Recipes

When the team let the AI loose on real data from the universe (using the latest observations from the DESI telescope, Planck satellite, and supernova surveys), the AI found two specific mathematical formulas that stood out.

Let's call them AI-1 and AI-2.

• AI-1 is like a "bouncy" recipe. It allows Dark Energy to behave in a way that crosses a specific boundary (called the "phantom divide") and then settles down. It's a bounded, smooth curve that fits the data slightly better than the standard recipes humans have used for years.
• AI-2 is like a "damped" recipe. It looks a lot like the standard human recipe, but with a special "brake" that stops the behavior from changing too wildly in the distant past. It suggests that Dark Energy might have been flexible recently but was more stable long ago.

Why are these special?
For a long time, humans have used a standard two-part recipe (called CPL) to describe Dark Energy. The AI found that these two new recipes fit the universe's data better than the human recipe, according to a statistical measure called "Bayesian evidence." This means the universe seems to prefer these new shapes, even though they are more complex.

The Catch: It's a Map, Not the Territory

The authors are very careful to say what this means and what it doesn't mean:

• What it IS: These are phenomenological models. Think of them as a very accurate map of the terrain. They describe how Dark Energy behaves mathematically, but they don't explain what Dark Energy actually is (like a specific particle or a new force of gravity).
• What it ISN'T: The AI didn't discover a new law of physics or a new particle. It didn't solve the mystery of Dark Energy's origin. It just found a better way to draw the curve that fits the data we have.

The Takeaway

This paper proves that AI can be a powerful partner in science. Instead of just crunching numbers, AI can help generate ideas.

By using a loop of "Propose -> Test -> Critique -> Refine," the AI was able to find mathematical descriptions of the universe that humans hadn't thought of, which fit the latest data slightly better than our best human guesses. It's a new way to do science: not just asking "Does this fit?" but asking "What else could fit?" and letting the machine explore the possibilities.

Technical Summary

Technical Summary: Physics-Guided Discovery of Dynamical Dark-Energy Equations of State

Problem Statement
The discovery of cosmic acceleration implies the existence of dark energy, yet the standard $\Lambda$CDM model leaves fundamental theoretical questions unresolved regarding the physical nature of this component. While dynamical dark-energy models (where the equation of state $w$ evolves with redshift) have been extensively explored, current inference relies heavily on human-designed, low-dimensional parameterizations (e.g., CPL, JBP). These traditional approaches are limited by their fixed functional forms and asymptotic behaviors, which may bias the search for new physics or obscure the true underlying phenomenology. Conversely, non-parametric reconstructions lack the compact analytic forms necessary for physical interpretation and theoretical comparison. The paper addresses the need for a framework that can explore a broader space of symbolic equation-of-state (EoS) candidates while retaining interpretability, physical motivation, and statistical rigor.

Methodology
The authors propose an iterative, AI-guided reasoning framework that treats the discovery of dark-energy EoS as a closed-loop process involving generation, evaluation, and refinement. The workflow consists of four main components:

1. Hypothesis Generation: A Large Language Model (LLM), specifically Qwen3-30B-A3B, acts as a generator. It proposes symbolic parameterizations for $w(z)$ as executable Python functions. Crucially, each proposal is accompanied by a natural-language cosmological rationale.
2. Retrieval-Augmented Generation (RAG): To ground the generation in existing physics, the system retrieves relevant passages from a corpus of $\sim$7,000 arXiv manuscripts on dark energy. This ensures proposals are contextually aware of the literature while allowing for novel combinations.
3. Cosmological Evaluation: Each candidate is embedded into a background cosmological model. Parameters are optimized against observational data (Type Ia supernovae, Baryon Acoustic Oscillations, and CMB).
   • Iterative Search Stage: Uses a reduced dataset (Pantheon+, 2D BAO, and compressed CMB shift parameters $R$ and $\ell_A$) for rapid screening.
   • Validation Stage: Uses full likelihoods (Planck 2018, DESI-DR2, Union3, DESY5) for final model comparison.
4. Independent Critic: A separate LLM (GPT-5.1 mini) evaluates the candidates based on physical plausibility, novelty, clarity of reasoning, numerical stability, and implementation validity. The critic provides scores and specific suggestions for modification.
5. Experience Buffer: Successful hypotheses (high statistical fit and high critic scores) are stored in an experience buffer. Subsequent generations are conditioned on this buffer, allowing the system to evolve both the mathematical structure and the physical reasoning over time.

The system was first validated via a "blind test" on an exotic mock cosmology with a non-standard EoS ($w_{exotic}(z) = \alpha / (1+z)^\beta$) to ensure the pipeline could recover non-standard dynamics without being biased toward familiar templates like CPL.

Key Contributions

• Framework for AI-Guided Phenomenology: The paper introduces a reproducible loop that integrates symbolic equation discovery with physical reasoning and observational constraints, moving beyond simple curve fitting to "physics-guided reasoning."
• Two Novel Parameterizations: The framework identified two new, compact, two-parameter EoS forms that are competitive with established models:
  • AI 1: A bounded quadratic parameterization with phantom-crossing phenomenology:
    $$w(z) = -1 - w_0 \left(\frac{z^2}{1+z^2}\right) + w_a \left(\frac{z^2}{1+z^2}\right)^2 - w_0 w_a$$
    This model features a smooth, bounded transition and allows for phantom-to-non-phantom trajectories.
  • AI 2: A damped intermediate-redshift parameterization:
    $$w(z) = w_0 + w_a \frac{z}{1+z} \exp\left[-\left(\frac{z}{1+z}\right)^2 (1+z)^{1/4}\right]$$
    This model preserves the low-redshift behavior of CPL but damps the high-redshift limit to $w_0$, decoupling late-time flexibility from early-time asymptotics.
• Validation of Non-Template Recovery: The blind mock test demonstrated that the system could recover the underlying cosmology of an exotic model (where CPL was strongly disfavored) without simply reusing known templates, proving the search space is not restricted to standard ansätze.

Results

• Blind Validation: In the exotic mock test, the AI-generated models (specifically AI-sim1 and AI-sim2) recovered the injected background cosmology ($\Omega_m, H_0$) within credible intervals, whereas the standard CPL model failed, yielding biased parameters despite tighter constraints. The best AI candidates achieved a log-Bayes factor ($\ln BE_i$) of 1.55 relative to the true model, compared to 6.97 for CPL (indicating strong disfavor).
• Real Data Analysis: Applied to combinations of Pantheon+, DESI-DR2, and Planck 2018 data:
  • The two AI-selected models (AI 1 and AI 2) were among the top performers, remaining within approximately one log-evidence unit of the CPL model.
  • For the DESI-DR2 BAO + Pantheon+ SNIa + Planck 2018 combination, AI 1 attained a larger Bayesian evidence than CPL (by more than one unit), favoring the dynamical model.
  • For DESI-DR2 BAO + Union3 SNIa + Planck 2018, AI 1 showed the strongest evidence among all tested models.
  • Both AI models preserve the early-universe parameters ($\omega_b, \omega_{cdm}, \theta_s$) consistent with Planck constraints, indicating their improvements arise from late-time background evolution rather than modifications to recombination physics.
  • The models predict nearly degenerate CMB temperature spectra with CPL, with differences primarily appearing at low multipoles.

Significance and Claims
The paper claims that AI-guided reasoning can complement traditional model building by proposing and evaluating interpretable phenomenological parameterizations. The authors emphasize that the framework does not assert autonomous scientific understanding but rather makes the process of phenomenological reasoning explicit, auditable, and revisable.

Key claims include:

• The ability to search the "phenomenological space" itself rather than being confined to pre-selected ansätze.
• The discovery of compact, two-parameter equations that are competitive with established forms and, in specific dataset combinations, statistically preferred over them.
• The demonstration that AI can adapt to non-standard expansion histories when data demand it, avoiding the "template bias" inherent in human-designed searches.
• The models should be interpreted as effective phenomenological descriptions rather than microphysical theories; their significance lies in showing how literature-grounded proposals and critic-based reasoning can be integrated into a reproducible search for effective models.

The authors remain modest, noting that the statistical preference for these models is selective (dependent on dataset combinations) and that the next step involves incorporating perturbation stability and growth observables directly into the search objective.