Hierarchical Dimensionless Learning (Hi-π): A physics-data hybrid-driven approach for discovering dimensionless parameter combinations

Mingkun Xia, Haitao Lin, Weiwei ZhangarXiv 2025 · arXiv · OpenReview

TL;DR

This paper introduces Hierarchical Dimensionless Learning (Hi-π), a novel physics-data hybrid approach that combines dimensional analysis and symbolic regression to automatically discover key dimensionless parameters in complex systems. Hi-π effectively reduces dimensionality, balances model accuracy and complexity, and improves the interpretability of physical models, outperforming existing methods in various fluid mechanics applications.

What Problem Does This Paper Address?

In many scientific and engineering fields, especially fluid mechanics, physical systems are described by a large number of variables. While dimensional analysis is a powerful tool for reducing this complexity by identifying dimensionless parameters, it often leads to redundant or unclear combinations, especially in high-dimensional systems. This makes it challenging to extract fundamental physical laws and build accurate, interpretable models. Traditional methods can be time-consuming, rely heavily on expert knowledge, and may struggle with complex, non-power-law relationships. This paper aims to address these limitations by developing an automated, data-driven method that can discover physically meaningful dimensionless parameter combinations, even for complex relationships.

Methods

The core of the proposed method, Hierarchical Dimensionless Learning (Hi-π), is a three-layer approach:

  1. Physical Layer (Dimensional Analysis): This layer uses dimensional analysis, particularly the Buckingham-Π theorem, to reduce the number of input variables by combining physical quantities into dimensionless parameters. This "dimensional embedding" accelerates subsequent steps, especially when dealing with many initial parameters.
  2. Data Layer (Symbolic Regression): Here, the method employs symbolic regression (specifically, the PySR algorithm) to discover mathematical expressions for parameter combinations. Unlike traditional methods, Hi-π does not assume these combinations are simple power laws, allowing for more complex and flexible relationships. PySR's multi-branch tree structure enables the discovery of multiple parameter combinations simultaneously.
  3. Mapping Layer (Polynomial Regression): Multivariate polynomial regression is used to map the discovered dimensionless parameter combinations to the output variable. This layer ensures the model is smooth, global, and can capture nonlinear relationships, providing a way to evaluate the quality of the discovered parameter combinations. The loss function, based on the error between the predicted and actual output, guides the symbolic regression search.

The method iteratively searches for optimal parameter combinations by adjusting the complexity of the polynomial fitting and evaluating the prediction error.

Datasets

The paper uses publicly available datasets from existing research for its case studies:

  • Rayleigh-Bénard convection: Data from three different papers ([29-31]) was collected and filtered for turbulent states.
  • Viscous flows in a circular pipe: Experimental data from Nikuradse [33] was used.
  • Compressibility correction in subsonic flow: Synthetic datasets were generated using the Karman-Tsien compressibility correction formula, with Latin hypercube sampling and added noise.

No new datasets were introduced.

Key Insights and Ideas

  • Physics-Data Synergy: Hi-π effectively bridges the gap between physics-based dimensional analysis and data-driven machine learning, leveraging the strengths of both.
  • Beyond Power Laws: The method breaks free from the restrictive assumption that dimensionless parameters must be simple power-law combinations, enabling the discovery of more complex, physically relevant relationships.
  • Hierarchical Structure Discovery: Hi-π can uncover hierarchical relationships between parameters, which is crucial for understanding complex systems.
  • Robustness to Noise and Data Range: The method demonstrates strong robustness in the presence of noisy data and is less affected by data range or parameter sensitivity compared to some existing methods.
  • Interpretability and Simplicity: By balancing accuracy and complexity, Hi-π aims to find parameter combinations that are not only predictive but also physically interpretable, offering a "white-box" approach.

Key Results

The Hi-π method was applied to three fluid mechanics examples:

  • Rayleigh-Bénard Convection: Accurately identified the two intrinsic dimensionless parameters: the Rayleigh number (Ra) and the Prandtl number (Pr), outperforming PyDimension, which sometimes failed to extract both. Hi-π also achieved higher prediction accuracy in both interpolation and extrapolation.
  • Viscous Flows in a Circular Pipe: Successfully rediscovered the Reynolds number (Re) and relative roughness (ɛ/D) as the optimal dimensionless parameters. The method demonstrated a trade-off between prediction accuracy and complexity, identifying the 4th-order polynomial fit as the best balance.
  • Compressibility Correction in Subsonic Flow: In the Karman-Tsien example, Hi-π achieved 70% accuracy in discovering complex parameter combinations, whereas direct symbolic regression failed. Hi-π's approach of first extracting optimal parameter combinations and then applying symbolic regression proved more robust and accurate for complex formulas.

Limitations and Future Work

  • Scarce Samples: The method requires a sufficient volume and accuracy of sample data, especially for extracting multiple parameter combinations, which can be more demanding than methods extracting a single combination.
  • Complex Mapping Relationships: While polynomial regression is effective, its ability to represent very complex mapping relationships in high-dimensional problems might be limited. The authors suggest exploring model-free methods from information theory for such scenarios.

The authors express hope that their work will inspire further development of similar methods and that Hi-π will find broader applications in various scientific fields.

Summary updated

Abstract

Dimensional analysis provides a universal framework for reducing physical complexity and reveal inherent laws. However, its application to high-dimensional systems still generates redundant dimensionless parameters, making it challenging to establish physically meaningful descriptions. Here, we introduce Hierarchical Dimensionless Learning (Hi-π), a physics-data hybrid-driven method that combines dimensional analysis and symbolic regression to automatically discover key dimensionless parameter combination(s). We applied this method to classic examples in various research fields of fluid mechanics. For the Rayleigh-Bénard convection, this method accurately extracted two intrinsic dimensionless parameters: the Rayleigh number and the Prandtl number, validating its unified representation advantage across multiscale data. For the viscous flows in a circular pipe, the method automatically discovers two optimal dimensionless parameters: the Reynolds number and relative roughness, achieving a balance between accuracy and complexity. For the compressibility correction in subsonic flow, the method effectively extracts the classic compressibility correction formulation, while demonstrating its capability to discover hierarchical structural expressions through optimal parameter transformations.
BibTeX
@misc{xia2025hierarchical,
  author = {Mingkun Xia and Haitao Lin and Weiwei Zhang},
  title = {{Hierarchical Dimensionless Learning (Hi-π): A physics-data hybrid-driven approach for discovering dimensionless parameter combinations}},
  year = {2025},
  eprint = {2507.18332},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.18332},
}
Figure