Data-driven Progressive Discovery of Physical Laws

This paper introduces Chain of Symbolic Regression (CoSR), a novel framework that mimics the progressive, hierarchical nature of scientific discovery to overcome the limitations of conventional end-to-end symbolic regression by successfully uncovering interpretable physical laws and improving scaling theories across diverse fluid dynamics and engineering applications.

The Gist

Imagine you are trying to figure out the secret recipe for a perfect cake.

The Old Way (Traditional Symbolic Regression):
Most AI tools today try to guess the entire recipe in one giant leap. They look at a pile of ingredients (data) and try to spit out a single, massive sentence like: "Mix 3.42 cups of flour, 0.007 eggs, and a pinch of salt, then bake at 341.2 degrees for 12.04 minutes, but only if the moon is full."

This approach has two big problems:

1. It's a mess: The recipe is so long and complicated that no human can understand why it works.
2. It fails: If you try to bake a cake with slightly different ingredients, the recipe breaks because it memorized the noise instead of the logic.

The New Way (This Paper's "CoSR"):
The authors of this paper, Mingkun Xia and Weiwei Zhang, realized that scientists don't actually discover laws of physics in one giant leap. They build them up, step-by-step, like stacking blocks.

They created a new AI framework called Chain of Symbolic Regression (CoSR). Think of CoSR not as a magician pulling a rabbit out of a hat, but as a master chef teaching an apprentice.

Here is how CoSR works, using simple analogies:

1. The "Lego" Approach (Progressive Discovery)

Instead of trying to build a whole castle at once, CoSR builds it layer by layer.

• Step 1: The Foundation (Invariance Learning): First, it strips away the confusing units. It doesn't care if you measure distance in miles or kilometers; it just looks at the ratio. It's like realizing that a "big" cake and a "small" cake are actually the same recipe, just scaled up.
• Step 2: Building the Walls (Multi-layer Compression): It finds small, simple patterns first. Maybe it discovers that "heat + time = cooking." Then it takes that simple rule and combines it with another simple rule to make a bigger rule. It builds a "knowledge chain."
• Step 3: Polishing the Finish (Scaling Transformation): Finally, it looks at the whole structure and says, "Hey, this part is too complicated. Let's simplify it." It turns a wiggly, messy curve into a clean, straight line that is easy to understand.

2. Real-World Examples from the Paper

The team tested their "Master Chef" AI on four different "kitchens" to prove it works:

• The Celestial Kitchen (Gravity):

  • The Goal: Figure out how planets move.
  • The Result: The AI didn't just guess Newton's Law of Gravity. It first rediscovered Kepler's law (how long a planet takes to orbit), then figured out how mass works, and finally combined them to "invent" the Law of Universal Gravitation all by itself. It was like watching the AI re-live the history of science in seconds.

• The Boiling Pot (Turbulent Convection):

  • The Goal: Understand how heat moves through a pot of boiling water.
  • The Result: Scientists usually see this as a messy, non-linear mess. CoSR found a hidden "correction term" (a secret ingredient) that made the relationship perfectly linear. It turned a chaotic storm into a straight, predictable highway.

• The Rusty Pipe (Viscous Flow):

  • The Goal: Predict how much resistance water faces in a rough pipe.
  • The Result: Old models needed different rules for smooth pipes vs. rough pipes. CoSR found a single, unified "Golden Rule" that worked for both smooth and rough pipes, bridging the gap between them.

• The Laser Welder (Laser-Metal Interaction):

  • The Goal: Predict how deep a laser burns into metal.
  • The Result: The AI discovered a new "Material Characteristic Number." It realized that Aluminum behaves very differently from Steel, not just because of the laser, but because of how the metal stores heat. It created a new formula that predicted the depth much more accurately, especially for Aluminum.

Why This Matters

The authors call the old method the "Rashomon Gate Dilemma." (Named after a movie where everyone tells a different story). The old AI gives you a thousand different math formulas that all fit the data, but none of them make sense to a human.

CoSR is different. It forces the AI to tell a story that makes sense. It builds the law from simple, understandable blocks (like "force," "mass," "time") rather than a giant, unintelligible wall of numbers.

In a nutshell:
This paper teaches AI to stop trying to memorize the whole textbook at once. Instead, it teaches the AI to learn like a human scientist: start with simple observations, build small theories, combine them, and refine them until you discover the universal laws of the universe. It's the difference between a random guess and a structured, logical discovery.

Technical Summary

1. Problem Statement

Conventional Symbolic Regression (SR), a "white-box" machine learning method used to extract interpretable mathematical expressions from data, typically employs an end-to-end, "one-step" discovery strategy. This approach faces significant limitations when applied to complex physical systems characterized by high dimensionality, multivariate interactions, and multiscale features:

• Combinatorial Explosion: Searching the entire space of mathematical expressions for a single complex law leads to an intractable search space.
• Overfitting vs. Parsimony Trade-off: High-fidelity models often result in lengthy, uninterpretable expressions that fail to generalize. Conversely, enforcing parsimony often omits critical physical information, reducing accuracy.
• Deviation from Scientific Logic: Real scientific discovery is rarely a single leap; it follows a hierarchical and progressive path from simplicity to complexity (e.g., Kepler's laws $\to$ Newton's gravity). Current SR methods ignore this cognitive progression, treating discovery as a static, one-off event.

The core problem is how to automate the discovery of physical laws in a way that mimics the human scientific process: building complex theories from simpler, interpretable knowledge units.

2. Methodology: Chain of Symbolic Regression (CoSR)

The authors propose Chain of Symbolic Regression (CoSR), a framework that models physical law discovery as a dynamically growing knowledge chain. Instead of searching for a global function $y = F(x)$ directly, CoSR decomposes the problem into a structured, multi-stage assembly process.

The framework operates through three sequential tiers:

A. Invariance Learning (Dimensionality Reduction)

• Mechanism: Applies the Buckingham $\pi$ theorem to normalize raw physical quantities into dimensionless groups.
• Goal: Reduces the input space while maintaining dimensional consistency.
• Implicit Discovery: Uses symbolic regression to uncover intrinsic constraint relationships (implicit equations) within the data, laying the groundwork for subsequent modeling.

B. Multi-layer Compression (Hierarchical Discovery)

• Mechanism: Treats the discovery process as a deep neural network-like hierarchy. It progressively extracts intermediate knowledge units with clear physical meanings from the dimensionless parameters.
• Process:
  • Inputs from the previous layer serve as features for the next.
  • Uses Hierarchical Symbolic Regression to find composite functions $y = f_1(f_2(\dots))$.
  • Employs a loss function based on multivariate polynomial fitting to ensure intermediate variables are parsimonious and influential.
• Goal: Constructs hidden layers of physical concepts (e.g., identifying the Reynolds number or Prandtl number as intermediate steps) rather than jumping to the final law.

C. Scaling Transformation (Refinement)

• Mechanism: Applies mathematical transformations to simplify complex nonlinear relationships into linear or unified scaling laws.
• Techniques:
  • "Curves to Lines": Transforming nonlinear power laws into linear relationships (e.g., $\log(y) = k \log(x)$).
  • "Multiple Lines to One": Unifying different asymptotic regimes (e.g., smooth vs. rough flow) into a single scaling equation.
• Goal: Achieves a trade-off between formal parsimony, physical interpretability, and predictive fidelity.

Key Technical Distinction: Unlike Symbolic Neural Networks (which are pre-specified), CoSR's knowledge chain is dynamically adapted and bidirectionally driven (bottom-up for refinement, top-down for simplification).

3. Key Contributions

1. Epistemological Shift: Moves symbolic regression from a "global search" paradigm to a "progressive assembly" paradigm, aligning AI discovery with the historical trajectory of scientific theory development.
2. Framework Innovation: Introduces CoSR, a structured framework integrating invariance learning, hierarchical compression, and scaling transformation to discover laws from data without prior theoretical assumptions.
3. Algorithmic Implementation: Develops specific loss functions and strategies for:
   • Implicit Discovery: Enforcing constant constraints and variable sensitivity.
   • Transformation Discovery: Reformulating complex relations into simpler forms via parallel subfunction search.
4. Validation: Demonstrates the framework's ability to recapitulate classical discoveries and uncover new physical insights across diverse domains.

4. Results and Case Studies

The authors validated CoSR on five distinct problems, demonstrating its ability to improve upon classical scaling theories:

Case Study	Classical Context	CoSR Discovery & Improvement
Universal Gravitation	Kepler's 3rd Law $\to$ Newton's Law	Recapitulated the historical path. Automatically discovered reduced mass ($\mu$) and a generalized Kepler's law for binary systems, unifying planet-star and binary-star dynamics.
Rayleigh-Bénard Convection	Nonlinear $Nu \sim Ra^\beta$	Discovered a linear scaling law by introducing a correction term ($\alpha \Delta T$) previously overlooked. Revealed a synergistic interaction between thermal expansion, Nusselt, and Rayleigh numbers.
Rough-Wall Pipe Flow	Piecewise laws (Blasius/Strickler)	Uncovered a Widom-type scaling law (Goldenfeld's scaling) that unifies smooth and rough regimes. Further improved this with a novel correction term, reducing prediction error in the transitional regime from 3.55% to 2.88%.
Laser-Metal Interaction	Keyhole depth prediction	Discovered a Material Characteristic Number ($X$) that naturally classifies materials (Al, Ti, SS). Derived a modified Keyhole Number ($Ke^*$) that significantly improved prediction accuracy for Aluminum alloys (reducing upper error bound by ~43%).
Aerodynamic Coefficients	Scaling for different aircraft	Established a unified non-power-product scaling function for axial force. Identified that the functional form remains identical across geometries (sharp cone vs. blunt body), with only constants ($c_1, c_2, c_3$) varying to reflect shock wave, friction, and pressure drag differences.

5. Significance and Future Outlook

• Bridging AI and Physics: CoSR addresses the "Rashomon gate dilemma" (multiple uninterpretable solutions) by enforcing a progressive, physics-guided structure. It transforms AI from a curve-fitting tool into a mechanism for structured knowledge assembly.
• Interpretability: The resulting laws are not just black-box predictions but transparent mathematical expressions composed of physically meaningful intermediate variables.
• Generalizability: The framework successfully handles complex, high-dimensional systems where traditional dimensional analysis or intuition fails to find the optimal scaling laws.
• Limitations & Future Work:
  • Selection among mathematically equivalent but physically inconsistent expressions still requires some human intuition.
  • The coordination between discovery modes currently relies on heuristic strategies; future work aims for autonomous pathway optimization.
  • Current CoSR discovers relationships within predefined concepts; the ultimate goal is for AI to autonomously generate new physical concepts and paradigms.

Conclusion: The paper presents CoSR as a transformative step toward "AI for Science," offering a methodology that not only discovers physical laws from data but also mimics the cognitive process of scientific discovery, moving from simple building blocks to complex, unified theories.