Abstract
Designing effective maritime structures requires balancing the maximization of wave protection with the minimization of construction costs. Traditional optimization approaches are often hindered by the extreme computational expense of high-fidelity physics simulations and the difficulty of representing variable geometric configurations. This paper presents a Surrogate-Assisted Evolutionary Algorithm (SaEA) that utilizes Deep Convolutional Neural Networks (CNNs) as fast proxies for wave height prediction. By representing breakwater layouts as 2D image masks, the system enables generative design without fixed-input constraints. Experimental results demonstrate that this approach achieves a 42% reduction in cost and a 17% improvement in wave defense compared to standard evolutionary baselines within a fixed time budget.
Introduction
Coastal engineering faces a critical challenge in the design of breakwaters: protecting harbors and shorelines from wave energy while managing the high costs of maritime construction. The design process typically involves iterative optimization, where each candidate configuration must be evaluated using numerical models like SWAN (Simulating WAves Nearshore). However, these simulations are computationally intensive, often requiring days to explore a meaningful design space. Furthermore, traditional optimization models struggle with variable geometry—situations where the number and arrangement of breakwater segments are not fixed.
To address these bottlenecks, we propose a generative design pipeline that combines evolutionary strategies with deep learning surrogates. This approach replaces slow physics simulations with a fast AI "proxy," allowing for the rapid exploration of thousands of potential designs.
Methodology
Generative Design Pipeline
The proposed framework, the Surrogate-Assisted Evolutionary Algorithm (SaEA), integrates several key components to streamline the design process:
- Evolutionary Strategy: The system employs the Strength Pareto Evolutionary Algorithm 2 (SPEA2) to evolve breakwater designs. This multi-objective approach identifies the optimal trade-offs (the Pareto front) between construction costs and wave attenuation.
- Image-Based Representation: To overcome the limitations of coordinate-based inputs, breakwater layouts are represented as 2D image masks. This allows the system to process any number of structures or segments, treating the design task as a computer vision problem.
- CNN Surrogate Model: A Deep CNN, utilizing an encoder-decoder architecture, is trained to predict wave heights across the study area. This surrogate acts as a high-speed replacement for the SWAN physics engine.
- The Surrogate Assistant: To mitigate the risk of "AI hallucinations" or inaccurate predictions, a secondary classifier—the "Assistant"—estimates the surrogate's confidence. If the predicted error is high, the system defaults to the high-accuracy SWAN simulation, ensuring the integrity of the optimization process.
Online Learning and Data Generation
Unlike models that require massive pre-existing datasets, this system utilizes an "online learning" approach. It begins with a random initialization phase to gather diverse spatial data, then trains the surrogate model on the fly. This allows the tool to adapt to specific coastal environments and bathymetry without prior training data.
Experimental Setup
The methodology was validated using a synthetic water area with varying bathymetry, where water depth increased progressively from one corner to another. The SWAN model served as the ground truth for wave height data. The performance of the SaEA was compared against standard evolutionary algorithms and random search methods under a strict 10-hour computational time limit.
Results and Discussion
Efficiency and Performance
The surrogate-assisted approach significantly outperformed traditional methods. Within the 10-hour window, the SaEA achieved:
- 42% reduction in construction costs.
- 17% improvement in wave defense effectiveness.
The designs discovered by the surrogate-assisted model consistently "dominated" those found by baseline methods, providing solutions that were simultaneously more affordable and more protective.
Surrogate Accuracy
The CNN surrogate demonstrated high reliability, achieving a Mean Absolute Percentage Error (MAPE) of just 4.24%. This level of accuracy confirms that deep learning models can effectively capture the complex spatial relationships between maritime structures and wave dynamics.
The Role of the Assistant
The inclusion of the Surrogate Assistant was vital for stability. By acting as a gatekeeper, it prevented the evolutionary algorithm from converging on unrealistic designs based on erroneous surrogate predictions, a common failure mode in surrogate-assisted optimization.
Conclusion and Future Work
This research demonstrates that treating coastal maps as images and utilizing CNN surrogates can revolutionize maritime structural design. The approach bypasses the rigid constraints of traditional modeling and significantly reduces the time required to find high-performance designs.
Future research will focus on transitioning from synthetic maps to real-world coastal environments with complex bathymetry and weather patterns. Additionally, the authors suggest exploring Generative Adversarial Networks (GANs) or Variational Autoencoders (VAEs) to propose even more creative candidate solutions. Finally, extending the model from 2D layouts to 3D dynamics—considering the volume and material cross-sections of the walls—will be essential for practical engineering applications.
Summary updated