Abstract
This paper presents a hybrid approach to solving the Reaction-Diffusion (RD) equation, a fundamental model for physical processes where substances spread and react. The authors specifically address sulfate attack in concrete, a chemical process leading to structural degradation. By combining deep neural networks with dimensional analysis (the Damkohler number), they developed a mesh-free alternative to traditional numerical solvers capable of accurately estimating chemical concentrations across varying reaction and diffusion rates.
Introduction
Modeling how chemicals penetrate and react within materials like concrete is vital for infrastructure longevity. This is typically achieved using Partial Differential Equations (PDEs). However, engineers face a "Goldilocks" problem: analytical methods provide exact solutions but only for simple geometries, while numerical methods (e.g., Finite Difference) are computationally heavy and mesh-dependent. This study bridges the gap using Deep Learning (DL) to create an "intelligent" solver that is mesh-free, fast, and generalizable.
Methodology
The study employs a three-pronged methodological approach:
- Analytical Solution: The Danckwerts method was used to derive an exact solution for the 1D RD equation, serving as a ground-truth benchmark.
- Data Generation: A synthetic dataset of 3 million points was generated using a second-order finite difference numerical method, covering a wide distribution of concentration ($C$), time ($T$), and diffusion coefficients ($D_e$).
- Deep Learning Architecture: A 3-layer Deep Neural Network (DNN) was designed using LeakyReLU and Sigmoid activation functions with the Adam optimizer.
The network integrates the Damkohler number ($Da$)—a dimensionless ratio of reaction rate to diffusion rate—to handle "edge cases" where either diffusion or reaction dominates.
Results and Discussion
The researchers evaluated the model using a "Threshold Concept," where a prediction is considered successful if it falls within a specific range ($\theta$) of the analytical solution.
Batch Sensitivity and Performance
A key finding was the sensitivity of the model to the number of training batches. As shown in the performance data, increasing the training batches significantly improved accuracy and reduced Mean Square Error (MSE).
The best-performing model achieved a 91.71% accuracy (at a threshold of 2) with the training MSE reduced to 0.5782 after 1000 batches. This demonstrates that increasing the data volume allows the model to better capture the complex behavior of the PDE.
Parameter Robustness
The model remained reliable across a wide range of diffusion coefficients ($2.6 \times 10^{-7}$ to $2.6 \times 10^{-12}$). However, accuracy decreased when the Damkohler number fell below specific values, indicating regimes where the neural network requires more specialized training data to handle the dominance of reaction over diffusion or vice versa.
Conclusion
The proposed DL framework offers a promising mesh-free tool for civil engineering applications. While currently limited to 1D space, it demonstrates high accuracy and computational efficiency once trained. Future work will focus on extending the model to 3D complex geometries and higher-dimensional spaces to better simulate real-world concrete structures.
Summary updated