01 · A
Problem
With only a small set of public jacket-design examples, the team could not treat a larger neural model as a substitute for evidence. The system had to connect data experiments, structural simulation, search and engineering review.
All projects05 / 09
Sparse data to inspectable trade-offs
Offshore Wind Design Platform
A research platform that connects sparse design data, OpenSees evaluation, evolutionary search and two-space Pareto analysis for offshore jacket concepts.
- Period
- 2024 — 2026
- Focus
- Multi-objective Optimization · Structural Engineering · Decision Support
- Core stack
- Python · Flask · Socket.IO · OpenSeesPy
TL;DR
- Problem
- With only a small set of public jacket-design examples, the team could not treat a larger neural model as a substitute for evidence. The system had to connect data experiments, structural simulation, search and engineering review.
- What I did
- Explored REaLTabFormer/GAN synthesis and structured MLP/GPT-2 SFT approaches.
- Result / limit
- Converted a large candidate set into separate performance and structural views without claiming a globally optimal or certified design.
01 · B
Context
The research moved from tabular-data synthesis and structured prediction experiments toward an interactive Flask/Socket.IO platform for OpenSees-backed NSGA-II and GA runs.
Why it matters
Jacket design has no single best answer across weight, utilization, dynamics and constructability. Engineers need feasible, explainable candidate families—not an opaque optimum label.
Explore technical detailsWhat made it hard · System path · Key trade-offs
What made it hard
- 01
Public design examples were scarce, so synthetic rows needed field-level distribution checks.
- 02
Multi-output predictions could violate relationships between structural fields.
- 03
OpenSees evaluation is expensive and failed candidates need deterministic handling.
- 04
The four objectives are normalized weight, a resonance penalty when natural frequency approaches the turbine's rated operating frequency, leg stress level and brace stress level; stress levels closer to zero are safer.
- 05
A large Pareto set still requires interpretable post-analysis.
Research choices
Explored models and adopted models are not the same story
Sparse data required several experiments. I distinguish my own exploratory work from the team's final model.
01explored
REaLTabFormer / GAN
Synthetic tabular-data experiments
02explored
Structured MLP
Joint multi-output prediction
03explored
GPT-2 SFT
Table-to-text inspired prediction
04team-adopted
Autoregressive path
Sequentially predicts structural fields
My approach
System path
Prepare
Compare synthetic-data variants with field-level distribution distances
Propose
Explore structured multi-output and table-to-text prediction paths
Evaluate
Build jacket geometry and calculate OpenSees responses
Decide
Search, stream generations and inspect Pareto families in two spaces
Workflow illustration
Turning a long optimization run into an inspectable workspace
I use this workflow illustration to bring parameters, a structural preview, generation progress and Pareto candidates into one view, while keeping the candidates explicitly exploratory.
Optimization loop
Every candidate travels through geometry, simulation and interpretation
The web layer does not replace the engineering loop. It makes inputs, failures, generations and trade-offs visible while OpenSees remains inside evaluation.
01configure
Design parameters
Geometry, material and environmental inputs
02model
Parametric jacket
Build a candidate structure
03simulate
OpenSees response
Evaluate structural objectives and failures
04search
NSGA-II / GA
Generate the next population
05interpret
Pareto analysis
Cluster two spaces and select representatives
Key trade-offs
Keep failed experiments visible
Structured MLP and GPT-2 SFT explorations are recorded as investigated paths, not relabeled as the adopted team model.
Stream progress, not authority
A WebSocket connection makes long evolutionary runs inspectable; streamed candidates are not certified designs.
Separate two kinds of similarity
Objective-space clusters explain trade-offs; design-space clusters explain structural families.
04 · A
My contribution
- Explored REaLTabFormer/GAN synthesis and structured MLP/GPT-2 SFT approaches.
- Developed the Flask/Socket.IO workflow connecting predicted structures, OpenSees and evolutionary search.
- Added generation-by-generation visualization and objective/design-space clustering for candidate review.
04 · B
What was implemented
- Implemented OpenSees-backed optimization, live progress and Pareto post-analysis paths.
- Converted a large candidate set into separate performance and structural views without claiming a globally optimal or certified design.
Lessons learned
01A research platform creates value by exposing failures, progress and trade-offs—not only an algorithm name.
02Prediction can seed a search, but engineering decisions need inspectable simulation and human judgment.
What I can share
I built this research prototype to explore the application, OpenSees/optimization workflow and two-space clustering. It supports candidate review; engineering validation and convergence certification remain outside its scope.
Public sources