Speakers
Description
Deadline-based environment often degrades the quality of designs from a naval architect. This research proposes a non-human (A.I.) designer to automate basic marine designs, focusing on 2D topology optimization for ships' engine room. The study investigates the design automation and its potential impacts on the maritime industry. The methodology involved technical datasets from built ships to train Artificial Neural Networks (ANN) and Generative Adversarial Networks (GAN) using machine learning (M.L.) platforms. For the scarcity of data, three sets were utilized with data from 152 real ships, 750 virtual ships, and 6,996 synthesized ships. Quality was assured using a Conditional Tabular Generative Adversarial Network (CTGAN), with results validated with regression methods & confusion matrices. Research findings were presented with training sessions, ML results, data validation and industry impacts. In the unsupervised sessions, the success rate for design generation was nearly 0% across all modules. After the supervised sessions, success rates improved to 15% for ANN, 0.06% for GAN, and 0.15% for CTGAN. Result validation showed, the primary AI modules are on development trend despite low success rates. Research limitations include small dataset, feature loss in geometry and length of supervised learning. However, the ANN module demonstrated potentials for future research scopes.
| Topic Area | Design Optimization, Additive Manufacturing Technologies & Applications |
|---|