TL;DR
This paper introduces a novel approach using symbolic regression (specifically genetic programming) to automatically discover heat transfer correlations from experimental data, bypassing the need to pre-define the correlation's functional form. This method yields more accurate predictions than traditional approaches, even for complex systems, by finding optimal functional forms and constants simultaneously.
What Problem Does This Paper Address?
In engineering design, predicting heat transfer is crucial. While first-principles calculations are ideal, they are often too complex due to factors like turbulence and temperature-dependent properties. Consequently, engineers rely on experimental data compressed into correlations. The traditional method involves assuming a functional form (often power laws) and then using regression to find the best constants. However, this approach has limitations: the assumed form might not be optimal, leading to larger predictive errors than experimental uncertainties, and regression can get stuck in local minima. This paper addresses the challenge of finding accurate heat transfer correlations by developing a method that automatically discovers both the functional form and the constants, thus avoiding the limitations of pre-defined forms.
Methods
The core of this paper is the application of Symbolic Regression (SR), specifically using Genetic Programming (GP). GP is an evolutionary computation technique that evolves computer programs (represented as parse trees) to solve a problem. In this context, GP searches for the best mathematical expression that fits the experimental data.
Key aspects of their methodology include:
- Tree-based Representation: Functions are represented as tree structures, where branch nodes are operators (+, -, *, /, ^) and leaf nodes are variables (like Reynolds number, Prandtl number) or constants.
- Fitness Function: The fitness of a correlation is defined as the reciprocal of the variance of the error between predicted and experimental values. A higher fitness means a better correlation.
- Penalty Function: To prevent overly complex correlations (which are harder to interpret and use), a penalty function is introduced. This function penalizes correlations based on their "size" (complexity, measured by the dimensions of their array representation). This ensures that the algorithm favors simpler, yet accurate, forms.
- Genetic Operators: Standard GP operators like crossover (exchanging parts of tree structures between two parent correlations) and mutation (randomly altering nodes in a tree) are used to evolve the population of correlations.
- Constant Determination: Since GP primarily evolves the structure of the correlation, the constants within the evolved functions are optimized separately using a Genetic Algorithm (GA) and local optimization (Nelder-Mead algorithm). This is done periodically to refine the best correlations found.
Datasets
The methodology was tested on two publicly available experimental datasets:
- Compact Heat Exchanger Data: Data from McQuiston [25] for a fin-tube compact heat exchanger, focusing on air-side heat transfer (Colburn j-factors) under dry-surface conditions. This dataset involves variables like Reynolds number (Re) and an air-side area ratio (Ar).
- Pipe Flow Data: Data from Sieder and Tate [26] for heating and cooling of liquids in pipes. This dataset includes Nusselt number (Nu) as the target variable, with inputs such as Reynolds number (Re), Prandtl number (Pr), viscosity ratio (μ/μw), and the diameter-to-length ratio (D/L).
Key Insights and Ideas
- Form-Free Correlation Discovery: The most significant insight is that GP can effectively discover functional forms for correlations without any prior assumptions. This liberates the process from the biases and limitations of human-chosen functional forms.
- Trade-off Between Complexity and Accuracy: The introduction of a penalty function highlights the inherent trade-off between correlation complexity and predictive accuracy. The sigmoidal penalty function allows for a controlled balance, preventing overly complex solutions while still achieving high accuracy.
- Multiplicity of Solutions: The study demonstrates that multiple functional forms can yield similar predictive accuracy, a phenomenon also observed in previous work. This suggests that there isn't always a single "best" correlation, but rather a set of good candidates.
- Improved Accuracy: The GP-derived correlations consistently outperformed existing, published correlations in terms of RMS percentage error for both datasets. This underscores the power of the form-discovery approach.
Key Results
- Compact Heat Exchanger: The GP-derived correlations showed smaller RMS percentage errors compared to the published correlation by McQuiston [23] and a globally regressed correlation by Pacheco-Vega et al. [3]. For example, the best GP correlation (Eq. 10) achieved an RMS error of 6.18%, compared to 14.74% for McQuiston's correlation.
- Pipe Flow: The GP-derived correlation (Eq. 14) also outperformed established correlations by Sieder and Tate [26] (13.79% error) and Levenspiel et al. [27] (11.19% error), achieving an RMS error of 10.69%.
- Complexity: While the GP-derived correlations were sometimes slightly more complex (e.g., Eq. 10 had a size of 24, compared to 10 for some others), the increase in complexity was justified by the significant improvement in predictive accuracy.
Limitations and Future Work
- Computational Cost: Symbolic regression, especially with genetic programming, can be computationally intensive, requiring multiple runs and statistical analysis to account for its probabilistic nature.
- Parameter Sensitivity: The performance of the penalty function depends on the choice of its parameters (a₁ and a₂). While the paper analyzes their effect, finding optimal values might require further investigation or heuristics.
- Interpretability: While the method finds accurate correlations, the resulting functional forms might not always be as intuitively interpretable as simpler, human-designed correlations, although the penalty function aims to mitigate this.
The authors suggest that this methodology is broadly applicable to other areas of engineering where empirical correlations are used. Future work could involve exploring different penalty functions, optimization strategies, and applying the method to more complex phenomena or different types of experimental data.
Summary updated