ProtoScience: Deterministic System Autonomously Rediscovers Physics Laws from Raw Data
Key Takeaways
- ▸ProtoScience successfully rediscovered multiple fundamental physics laws including Kepler's Third Law, General Relativity predictions, and gravitational wave relationships using only mathematical techniques, not LLMs
- ▸The system demonstrates high accuracy across diverse real-world datasets from NASA, NOAA, and LIGO, with validation R² values exceeding 0.998 for most physics discovery tasks
- ▸The deterministic pipeline correctly identifies when no meaningful law exists in noisy data, as shown by R² = 0.00 for Bitcoin prices, indicating robust failure detection
Summary
ProtoScience, a new deterministic pipeline for autonomous scientific discovery, has demonstrated remarkable success in rediscovering fundamental physics laws directly from raw numerical data without relying on language models for the discovery process. The system employs sparse regression, power-law fitting, and statistical validation to identify governing equations, using an LLM only for final interpretation. Validation results include recovering Kepler's Third Law with R² = 0.998 from 3,519 NASA exoplanet observations, achieving 93% accuracy on the Sun's 27-day rotation period from solar wind data, and perfectly predicting all five General Relativity predictions (R² = 1.000) from simulated black hole observables.
The pipeline architecture processes raw data through feature extraction, candidate law generation, fitting, and verification stages, demonstrating robustness by correctly identifying when no meaningful law exists—as evidenced by returning R² = 0.00 for Bitcoin daily prices. Additional validation includes discovering the T ~ v^3.40 power law in solar wind data and extracting the chirp mass relationship from 219 LIGO gravitational wave events with R² = 0.998. All experiments are fully reproducible with code publicly available on GitHub.
- LLMs are deliberately excluded from the discovery mechanism and used only for final natural language interpretation, challenging assumptions about AI's role in scientific breakthrough
Editorial Opinion
ProtoScience represents an important philosophical and methodological shift in AI-assisted scientific discovery—demonstrating that deterministic mathematical approaches can rival or exceed more recent LLM-centric methods in recovering known physics laws from raw data. By explicitly avoiding language models in the discovery pipeline and focusing on sparse regression and statistical validation, the project challenges the notion that large neural networks are necessary for scientific breakthroughs, while the perfect recovery of General Relativity predictions suggests genuine mathematical insight rather than pattern memorization. This work opens intriguing questions about the optimal role of different AI paradigms in science and suggests that well-designed classical techniques may be underutilized in the rush toward foundation models.
