TL;DR
This paper introduces an intelligent model using a Back-Propagation Neural Network (BPNN) to predict the dynamic shear modulus (G) and damping ratio (D) of undisturbed marine clay. This approach overcomes limitations of traditional empirical models by accurately capturing the complex relationships between soil properties, shear strain, and depth, while avoiding error propagation issues.
What Problem Does This Paper Address?
Marine engineering projects, such as offshore wind farms and tunnels, are increasingly common. The seismic performance of these structures is heavily influenced by the dynamic properties of the underlying marine soil. Specifically, the dynamic shear modulus (G) and damping ratio (D) are crucial parameters that describe how soil behaves under dynamic loading. These properties are known to vary significantly with shear strain (γ) and are also influenced by factors like soil depth, which translates to mean effective confining pressure (σm) in laboratory tests.
Existing empirical models, often based on hyperbolic functions, struggle to accurately represent these complex, non-linear relationships across different soil depths. They typically require manual fitting of parameters, which can introduce errors and limit their universality. Furthermore, these models often fail to explicitly account for the influence of soil depth (or confining pressure) on the G and D characteristics, making it difficult to develop a robust predictive tool for diverse marine clay conditions. This paper aims to bridge this gap by developing a more intelligent and universally applicable model.
Methods
The core of this research is the development of an intelligent model based on a Back-Propagation Neural Network (BPNN). The BPNN architecture consists of an input layer, a hidden layer with 'n' neurons, and an output layer.
- Input Layer: Takes in two key parameters: mean effective confining pressure (σm, representing soil depth) and shear strain (γ).
- Hidden Layer: Processes the input data through neurons activated by a hyperbolic tangent sigmoid transfer function.
- Output Layer: Produces the predicted values for dynamic shear modulus (G) and damping ratio (D).
The model employs a forward propagation mechanism to predict G and D based on the input parameters. Crucially, it incorporates a back-propagation algorithm to adjust the network's weights and thresholds based on the error between predicted and actual values. This iterative learning process, utilizing Bayesian regularization and gradient descent momentum, allows the BPNN to "learn" the complex, non-linear relationships within the data.
A key advantage highlighted is that this BPNN approach avoids the "forward propagation of errors" inherent in fitting empirical formulas, where errors in one parameter's fit can cascade to others. It also eliminates the need for manual intervention in setting fitting parameters.
Datasets
The study utilized experimental data from resonant-column (RC) tests conducted on undisturbed marine clays collected from two distinct locations in China: Bohai Bay and Hangzhou Bay.
- Bohai Bay: 17 samples were collected from depths ranging from 6.3 m to 70 m.
- Hangzhou Bay: 10 samples were collected from depths ranging from 18.3 m to 68.5 m.
These samples were classified as lean clay (CL) according to the Unified Soil Classification System. The RC tests were performed under mean effective confining pressures (σm) ranging from 43 to 466 kPa.
For model training, approximately 70.7% of the Bohai Bay data and 70.0% of the Hangzhou Bay data were used. The remaining data served as a prediction set for model validation and error calculation.
Key Insights and Ideas
- Depth Matters (Significantly): The study strongly confirms that not only shear strain but also soil depth (represented by mean effective confining pressure, σm) profoundly influences the dynamic shear modulus (G) and damping ratio (D) of marine clays. Deeper soils (higher σm) exhibit higher G and lower D for a given strain.
- BPNN as a "Black Box" Learner: The BPNN model effectively "learns" and predicts these complex, non-linear relationships without explicit mathematical formulation of the depth dependency, a major advantage over traditional empirical models.
- Error Mitigation: By using back-propagation, the model inherently avoids the cascading error issues common in fitting multi-parameter empirical equations, leading to more robust predictions.
- Universality Potential: The prediction accuracy of the BPNN model was found to be relatively insensitive to the number of neurons in the hidden layer, suggesting a degree of universality and robustness.
- Damping Ratio Prediction Improvement: The BPNN model shows a particularly strong improvement in predicting the damping ratio (D) compared to existing mathematical functions, which often struggle with this parameter.
Key Results
- High Prediction Accuracy: The intelligent BPNN model achieved high prediction accuracy for both G and D. For the prediction set, the relative errors between predicted and measured values were generally within ±8% for G and D.
- Statistical Performance: Statistical indicators like Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) showed desirable levels, with MAE decreasing as the number of hidden neurons increased up to a certain point. Coefficient of correlation (R) values were consistently above 0.9.
- Superior to Existing Models: When compared to established models by Ishibashi and Zhang [20] and Zhang et al. [24], the proposed BPNN model demonstrated superior performance, particularly in predicting the damping ratio (D). The Ishibashi and Zhang model tended to overestimate stiffness, while Zhang et al.'s model showed poorer D prediction.
- Effective Curve Fitting: The model accurately predicted the characteristic G/Gmax-γ and D-γ curves, capturing the non-linear decay of G and the growth of D with increasing shear strain, and how these curves shift with increasing soil depth.
Limitations and Future Work
- Limited Geographic Scope: The study focused on marine clays from only two specific sea areas (Bohai Bay and Hangzhou Bay). The authors acknowledge that research on undisturbed marine soils is scarce, and more data from diverse locations and depths are needed.
- Future Work: The authors suggest expanding the study to include a wider variety of marine soils and locations. They also propose constructing an "intelligent database" based on the developed model to further advance the understanding and prediction of marine soil dynamics.
Summary updated