Fitting Methods for Experimental Data and Theoretical Models

In the study of thermal convection, experimental observations and theoretical modeling exist in a symbiotic relationship. While experimental data provides a high-fidelity record of physical phenomena, theoretical models—or empirical correlations—aim to distill these observations into underlying physical laws. The process of curve fitting serves as the bridge between these two domains. It is not merely a tool for verifying model accuracy; it is a critical procedure for determining unknown parameters, thereby enabling the construction of mathematical expressions capable of predicting performance in real-world engineering applications.

At its core, curve fitting involves finding a mathematical function $f(x, \beta)$ that best represents a set of experimental data points $(x_i, y_i)$ by minimizing the discrepancy between the function and the observations under a specific criterion. Here, $\beta$ represents the set of model parameters to be determined.

In heat transfer research, a quintessential objective is to establish the relationship between the Nusselt number ($Nu$), the Reynolds number ($Re$), and the Prandtl number ($Pr$). A common form for such empirical correlations in forced convection is the power-law expression:

$$Nu = C \cdot Re^m \cdot Pr^n$$

In this context, the goal of the researcher is to use experimental data to solve for the constants $C, m,$ and $n$.
Depending on the mathematical structure of the model, fitting techniques are generally categorized into linear and non-linear approaches.

1. The Least Squares Method

The Least Squares Method is the most widely utilized criterion in statistical modeling. The fundamental principle is to minimize the Residual Sum of Squares (RSS), which represents the sum of the squared differences between the observed experimental values $y_i$ and the values predicted by the model $f(x_i, \beta)$:

$$RSS = \sum_{i=1}^{n} [y_i - f(x_i, \beta)]^2 \to \min$$

By minimizing this sum, we ensure that the resulting model provides the best possible "average" fit to the dataset.

2. Linearization via Logarithmic Transformation

Many fundamental correlations in fluid mechanics, such as the power-law model mentioned above, are inherently non-linear. However, performing direct non-linear regression can sometimes be computationally intensive or sensitive to initial guesses. A highly efficient alternative is linearization.

By applying a natural logarithm to both sides of the power-law equation, we transform it into a linear form:

$$\ln(Nu) = \ln(C) + m \ln(Re) + n \ln(Pr)$$

This transformed equation is linear with respect to the parameters $\ln(C), m,$ and $n$. This allows researchers to use multiple linear regression to find the parameters quickly, often yielding an analytical solution with high computational efficiency.

3. Non-linear Regression

In cases where the model cannot be linearized—or where linearization distorts the error distribution (making it non-homoscedastic)—non-linear regression algorithms must be employed.

• Levenberg-Marquardt (LM) Algorithm: This is the industry standard for non-linear least squares problems. The LM algorithm is a hybrid approach that interpolates between the Gauss-Newton algorithm and the method of gradient descent. It is prized for its robustness and its ability to converge even when the initial parameter estimates are relatively far from the optimal values.
• Application Scenarios: This method is essential for complex heat transfer correlations that involve exponential terms, fractional exponents, or more sophisticated transcendental functions that do not yield to simple logarithmic transformations.

A Standardized Workflow for Data Fitting

To ensure that the resulting models are scientifically rigorous and reproducible, a systematic approach should be followed:

1. Data Preprocessing: Before fitting, it is vital to identify and remove outliers—data points resulting from experimental error rather than physical phenomena. Furthermore, data normalization or scaling is often necessary. For instance, if $Re$ is on the order of $10^5$ while $Pr$ is $0.7$, the vast difference in magnitude can lead to numerical instability or failure to converge during optimization.
2. Model Selection: The choice of model must be grounded in physical intuition. For example, in convective heat transfer, the scaling of $Nu$ with $Re$ typically changes depending on the flow regime; one might use $Nu \propto Re^{1/2}$ for laminar flow and $Nu \propto Re^{0.8}$ for turbulent flow.
3. Parameter Initialization: For non-linear solvers, the "starting guess" is critical. Poor initialization can lead to local minima rather than the global optimum. It is best practice to use values derived from existing literature or from a preliminary linearized fit.
4. Computational Execution: Modern computational tools, such as scipy.optimize.curve_fit in Python or nlinfit in MATLAB, are typically used to perform the heavy lifting of the optimization process.
5. Model Validation: A model should never be judged solely on how well it fits the data used to create it. A robust approach involves testing the fitted model against a validation dataset—a separate set of experimental data not used during the fitting process.

Quantitative Evaluation Metrics

Visual inspection of a curve passing through data points is insufficient for scientific validation. Quantitative metrics must be employed to assess the model's performance:

• Coefficient of Determination ($R^2$): This indicates the proportion of the variance in the dependent variable that is predictable from the independent variables. An $R^2$ value approaching 1 suggests a high degree of correlation.
• Root Mean Square Error (RMSE): This provides a measure of the absolute deviation between the predicted and observed values. It is expressed in the same units as the dependent variable, making it easy to interpret in terms of physical magnitude.
  $$RMSE = \sqrt{\frac{1}{n}\sum_{i=1}^{n}(y_i - \hat{y}_i)^2}$$
• Mean Absolute Percentage Error (MAPE): In heat transfer studies, where the Nusselt number can vary across several orders of magnitude across different flow regimes, MAPE is often more informative than RMSE. It provides a relative measure of error, which is more meaningful when comparing accuracy across different scales of operation.

Critical Pitfalls: Overfitting and Extrapolation

While the pursuit of a "perfect fit" is tempting, researchers must remain vigilant against two common errors:

Overfitting occurs when a model is made overly complex (e.g., using a high-degree polynomial) to capture every minor fluctuation and noise in the experimental data. While this results in an exceptionally high $R^2$, the model loses its physical significance and fails to generalize to new data. A good model should capture the trend, not the noise.

Extrapolation Risk is perhaps the most dangerous error in empirical modeling. A correlation derived from a specific range of Reynolds numbers is only valid within that range. Because the underlying physics of fluid flow can change fundamentally—such as the transition from laminar to turbulent flow—applying a model to a regime outside its experimental bounds can lead to wildly inaccurate and physically impossible predictions.