A Step in the Right Direction: Surrogates and Optimization Workshop

October 28, 2025

This interactive blog demonstrates how data preparation, surrogate modeling, active learning, and optimization methods can be applied to real engineering design problems. Interactive demos show:

• Mooring Line Optimization — data-science workflow (EDA, surrogate models, validation).
• Hull Form Optimization — multi-fidelity workflow combining low- and high-fidelity simulations.
• Adaptive Learning — Bayesian optimization that selects new samples based on model uncertainty.

Note: If the embedded Huggingface Spaces are not loaded properly or if you are viewing this on a mobile device, please click on the usecase links directly or refresh page.

---

Usecase 1: Mooring Line Optimization (Data Science Workflow)

Interactive demo: Usecase 1 Link

Why It Matters

• Mooring layout affects operability and safety.
• Data cleaning and surrogate modeling reduce costly simulation cycles.

---

Objective

Accelerate optimization of mooring line parameters by replacing repeated simulations with surrogate models.

Design variables

• feature_diameter_mm
• feature_anchorRadius_m
• feature_f_ml_length

Performance targets

• target_MooringMass_t
• target_MaxLoads_N
• target_MaxLoads_norm_by_displ

Goal: minimize load while keeping mass low.

---

Workflow Overview

1. Exploratory Data Analysis

   • Inspect distributions, correlations, and outliers.

2. Modeling

   • Compare Linear, Ridge, RF, and GP models (R², RMSE, residuals).

3. Cross-Validation

   • 5–10 fold CV to assess generalization.

4. Optimization

   • GA (single objective) or NSGA-II (multi-objective).
   • Visualize Pareto trade-offs.

---

Key Observations

• Diameter and anchor radius strongly influence mass and load.
• GP models perform well for small datasets.
• CV results show high accuracy (R² > 0.95, RMSE < 5%).
• NSGA-II identifies balanced mass–load trade-offs.

---

Learning Outcomes

• Surrogates enable rapid design exploration.
• Cross-validation increases confidence in predictions.
• Multi-objective optimization clarifies trade-offs.
• Workflow generalizes to many engineering applications.

---

Usecase 2: Hull Form Optimization (Multi-Fidelity Workflow)

Interactive demo: Usecase 2 Link

Why It Matters

• High-fidelity CFD is expensive.
• Multi-fidelity strategies reduce cost by mixing LF and HF samples.

---

Overview

Minimize calm-water resistance using 14 design parameters with bounds ([-1, 1]).

Three fidelity levels differ in mesh density, cost, and accuracy. Multi-fidelity learning uses many LF samples plus a few HF samples to improve prediction quality within a fixed compute budget.

---

Simulation Fidelity Summary

Fidelity	Total Nodes	Resistance [N]	Grid Error [%]	NCC
High	16.5k	41.5	1.16	1.00
Medium	5.7k	43.5	5.97	0.11
Low	2.1k	48.9	19.3	0.03

Diagnostics include Pearson r, BPB, and Kendall τ to judge MF suitability.

---

Workflow Overview

1. Reference & DoE

   • Generate HF/LF samples that meet the set computational budget.

2. Modeling

   • GP surrogate with Matern kernel + WhiteKernel.

3. Diagnostics & Comparison

   • Compare MF vs HF-only vs LF-only at equal cost using RMSE and parity plots.

---

Key Observations

• MF often achieves better accuracy per computational cost than HF-only.
• Too little HF weakens LF→HF correlation; too much HF reduces MF benefit.
• Diagnostics help determine when MF is appropriate.

---

Learning Outcomes

• MF modeling balances accuracy and cost.
• Proper diagnostics guide sampling allocation.
• Valid comparisons require equal computational budgets.

---

Usecase 3: Active Learning (Bayesian Optimization Workflow)

Interactive demo:

Usecase 3 Link

Why It Matters

• Expensive simulations limit dataset size.
• Adaptive sampling focuses effort where uncertainty or expected improvement is highest.

---

Objective

Introduce adaptive learning / Bayesian optimization:

• Single-objective: GP + Expected Improvement (EI).
• Multi-objective: GP + Expected Hypervolume Improvement (EHVI).
• ** Propeller Optimization: Real-world case for multi-objective EHVI application

---

Workflow Overview

1. Initialize

   • Small DoE (5–10 samples).
   • GP with Matern kernel.

2. Acquire

   • Evaluate EI or HVI to select next samples.

3. Evaluate

   • Run simulation or function at selected point.

4. Update

   • Retrain and repeat until convergence or budget limit.

---

Key Observations

• EI efficiently finds global minima with few evaluations.
• HVI steadily expands the Pareto front and tracks progress via hypervolume.
• Same approach applies to expensive engineering simulations.

---

Learning Outcomes

• Bayesian optimization intelligently balances exploration vs. exploitation.
• Matern kernels suit smooth engineering responses.
• EI works well for small datasets; HVI extends BO to multi-objective problems.
• Adaptive workflows support efficient experimental or simulation-based design.

---