Autonomous Discovery of Low-Dimensional Physical Invariants from Real-World Data

Modern machine learning systems excel at prediction but rarely explain why complex systems behave as they do. Scientific progress, however, depends on identifying low-dimensional organizing principles that compress messy observations into interpretable structure. This work presents a general computational architecture for autonomous discovery of physical invariants from real-world data. Rather than fitting target variables directly, the system searches for derived quantities that collapse heterogeneous system behaviors onto shared functional relationships. The objective isn't accuracy, but abstraction. The approach is validated across multiple settings, including synthetic physical systems with hidden dynamics and a real NASA lithium-ion battery aging dataset. In the battery domain, the system autonomously identifies a candidate thermodynamic efficiency invariant, constructed from delivered electrical energy and thermal rise, which collapses degradation trajectories across distinct cells without referencing capacity directly. The results demonstrate that autonomous abstraction and theory formation are tractable computational problems, distinct from regression, curve fitting, or black-box prediction. This repository serves as a methods-level proof of concept for machine-assisted scientific discovery, emphasizing explanation as compression rather than performance on predefined targets. The work is intentionally conservative in scope. It doesn't claim a complete theory of any physical domain, but instead illustrates how machines can discover organizing variables that humans traditionally identify through intuition, dimensional analysis, and long empirical study.