How is a physics model usually built? A physicist proposes a hypothesis based on intuition, analogy, or theoretical reasoning, writes down an equation, and tests it against data. A team of Brazilian and international researchers has proposed a different approach: this cycle is now performed by a large language model (LLM).
The framework works iteratively. The AI generates dark energy equations of state along with physical justifications, drawing on scientific literature. Each candidate equation is embedded into a cosmological model and optimized against real observations: supernovae (Pantheon+), baryon acoustic oscillations (DESI DR2), and Planck 2018 data. Then an independent AI critic evaluates each proposal for physical motivation, novelty, stability, and correctness — and the next generation of equations incorporates this feedback.
The result: the system found two dark energy parameterizations that, to the best of the authors' knowledge, had never been explored before. And they're not just new — the best one showed higher Bayesian evidence than traditional parameterizations by more than one unit.
Crucially, the AI doesn't replace the physicist. It generates interpretable equations with physical meaning, not a "black box." It's a tool that expands the hypothesis space — suggesting options a human might not reach due to the habit of thinking within familiar models.
Dark energy is one of cosmology's biggest puzzles. Recent DESI data hints it may evolve over time. And now another player has joined the search for answers — artificial intelligence.