Introduction
Predicting the spread of contaminants in natural streams is a critical task for environmental engineering. This process is governed by the Longitudinal Dispersion Coefficient (LDC). Traditionally, researchers have relied on analytical models (often too simple), statistical regressions (limited generalizability), or "black-box" machine learning models (highly accurate but opaque). This paper introduces the Evolutionary Symbolic Regression Network (ESRN), a novel approach that bridges the gap between high-performance machine learning and transparent, physics-based equations.
Methodology
The ESRN framework combines Genetic Algorithms (GA) to evolve equation structures with Neural Networks (NN) to optimize constant parameters. To ensure physical validity, the authors employed Dimensional Analysis (Buckingham $\pi$ theorem), constraining the AI to produce dimensionally consistent results.
A key part of the methodology involved rigorous data refinement. The researchers used the Inter Quartile Range (IQR) method to identify and remove outliers, followed by Subset Selection of Maximum Dissimilarity (SSMD) to ensure the training and testing sets were representative of global river diversity.
Datasets
The study utilized a comprehensive global database of 1,094 samples of hydraulic and channel properties. After cleaning, 660 samples remained. Key features included channel width ($w$), depth ($d$), flow velocity ($U$), and shear velocity ($U^*$), with the target variable being the LDC ($D_l$).
Key Insights and Ideas
The primary breakthrough of the ESRN is its ability to "distill" complex data into an elegant, interpretable formula: $D_l = 13.89 w U^*$. This simplicity is a major advantage over previous symbolic models.
As shown in the table above, earlier attempts using genetic algorithms or neural networks (e.g., Models 24 and 25) resulted in extremely complex, multi-line equations that are difficult to interpret and prone to overfitting. In contrast, the ESRN-derived formula recovered fundamental physics, mirroring G.I. Taylor’s classic work on turbulent pipe flow. This suggests that turbulent mixing remains the dominant mechanism even in complex natural river systems.
Key Results
The ESRN model demonstrated superior performance compared to existing symbolic models, achieving an $R^2$ of 0.34 and an RMSE of 32.29 on the test set. While these metrics reflect the inherent difficulty of modeling natural streams, they represent a significant leap over previous benchmarks (like Models 22 and 23, which had $R^2$ values near 0.1). Crucially, the model showed robust generalization across different river types, making it a practical tool for field engineers.
Limitations and Future Work
The model's precision decreases for very high LDC values ($D_l > 100$), likely because the current four-variable feature set has reached its theoretical limit. Future research could incorporate additional parameters such as river sinuosity or bed roughness to better capture high-dispersion scenarios and further expand the dataset to refine the evolutionary process.
Summary updated