Design of Experiments (DOE) - Interactive Expert

Overview

Master experimental design through interactive, goal-driven guidance that asks the right questions to recommend the best approach for your situation. This skill covers classical DOE, Bayesian optimization, model-driven designs, and active learning—helping you choose between batch and sequential strategies, screening and optimization, and exploration and exploitation.

Core value: Don't start with a method—start with questions. Based on your goals, budget, and constraints, get personalized recommendations for experimental design strategies that maximize information gain per experiment.

CRITICAL: Start with Questions, Not Methods

When a user mentions DOE or experimental design, ASK THESE QUESTIONS FIRST:

Primary Questions (Ask Before Recommending):

1. "What is your primary goal?"

   • Screening: Identify which factors matter (many factors → few important)
   • Optimization: Find best settings for known factors
   • Exploration: Understand the system/build model
   • Model discrimination: Choose between competing models
   • Robustness: Minimize sensitivity to noise

2. "Can you run experiments sequentially, or must they be done in a batch?"

   • Sequential: Run one (or few), get results, decide next experiment(s)
   • Batch: Must plan all experiments upfront
   • Hybrid: Sequential with occasional batches

3. "How expensive is each experiment?"

   • Very expensive: >$1000 or >1 day per experiment
   • Moderate: $100-$1000 or hours per experiment
   • Cheap: <$100 or minutes per experiment

4. "How many factors are you investigating?"

   • 1-5 factors: Full factorial, RSM feasible
   • 6-15 factors: Need screening or fractional designs
   • >15 factors: Definitive screening, regularization needed

5. "Do you have prior knowledge, existing data, or a mechanistic model?"

   • Have model: Use model-driven optimal designs
   • Have data: Warm-start Bayesian optimization
   • No knowledge: Need exploration, space-filling designs

Based on Answers, Recommend:

Sequential + Expensive + Unknown landscape
    → Bayesian Optimization (GP + Expected Improvement)

Sequential + Have model + Parameter estimation goal
    → Model-driven sequential optimal design

Batch only + Screening goal + Many factors
    → Fractional factorial or Definitive Screening Design

Batch only + Optimization + Few factors
    → Central Composite Design or Box-Behnken

Sequential + Moderate cost + Want model
    → Active learning with Gaussian Process

Batch only + No prior knowledge + Exploration
    → Latin Hypercube Sampling

Quick Decision Tree

Can experiments be run sequentially with feedback?
│
├─ YES (Sequential possible)
│   │
│   ├─ Expensive experiments (>$1000 or >1 day each)?
│   │   ├─ YES → Bayesian Optimization
│   │   │        • Expected Improvement for optimization
│   │   │        • Upper Confidence Bound for exploration
│   │   │        • Tools: BoTorch, scikit-optimize, Ax
│   │   │
│   │   └─ NO → Have a model?
│   │       ├─ YES → Model-driven sequential design
│   │       │        • D-optimal for parameter estimation
│   │       │        • Update design as parameters learned
│   │       │
│   │       └─ NO → Active learning
│   │                • GP with uncertainty sampling
│   │                • Tools: modAL, custom GP
│   │
└─ NO (Batch only)
    │
    ├─ What's your goal?
    │   │
    │   ├─ SCREENING (identify important factors)
    │   │   ├─ <20 runs available → Plackett-Burman, Fractional Factorial
    │   │   ├─ 20-50 runs → Definitive Screening Design
    │   │   └─ Tools: pyDOE3, dexpy
    │   │
    │   ├─ OPTIMIZATION (find best settings)
    │   │   ├─ 2-5 factors → Central Composite Design, Box-Behnken
    │   │   ├─ >5 factors → Sequential screening → then optimize
    │   │   ├─ Have model → D-optimal or I-optimal
    │   │   └─ Tools: pyDOE3, pycse, statsmodels
    │   │
    │   ├─ EXPLORATION (understand system)
    │   │   ├─ No prior → Latin Hypercube Sampling
    │   │   ├─ Build surrogate → Space-filling + modeling
    │   │   └─ Tools: pyDOE3, pycse, scipy
    │   │
    │   └─ ROBUSTNESS (minimize variability)
    │       └─ Control + noise factors → Taguchi / Robust Parameter Design

When to Use Each Approach

Classical DOE (Batch Experiments)

Use when:

• Must plan all experiments upfront
• Well-understood system or standard optimization
• Moderate number of factors (<15)
• Experiments are cheap enough for comprehensive coverage

Best for:

• Screening many factors (Plackett-Burman, fractional factorial)
• Response surface modeling (CCD, Box-Behnken)
• Standard process optimization
• Teaching/learning DOE concepts

Tools: pyDOE3, dexpy, pycse, statsmodels

Bayesian Optimization (Sequential)

Use when:

• Can run experiments one-at-a-time (or small batches)
• Experiments are expensive (time, money, resources)
• Black-box system (no mechanistic model)
• Want to minimize total number of experiments

Best for:

• Optimizing expensive processes ($1000+ per run)
• Materials discovery, drug screening
• Hyperparameter tuning for ML models
• Autonomous experimentation / self-driving labs

Tools: BoTorch, scikit-optimize, Ax

Model-Driven DOE (Optimal Designs)

Use when:

• Have mechanistic or empirical model
• Goal is parameter estimation or model discrimination
• Want statistically optimal experiments
• Can update designs based on current parameter estimates

Best for:

• Chemical kinetics (rate constant estimation)
• Pharmacokinetics (PK parameter estimation)
• Systems biology (model calibration)
• Comparing competing models

Tools: Custom implementation with Fisher Information Matrix, pyoptex

Active Learning (Sequential Model Building)

Use when:

• Want to build accurate surrogate model efficiently
• Can run experiments sequentially
• Need to balance exploration and exploitation
• Moderate experiment cost

Best for:

• Adaptive sampling for complex landscapes
• Building GP surrogates for later optimization
• Uncertainty quantification
• Smart data collection

Tools: modAL, scikit-learn, GPy

Quick Reference Table

Situation	Recommended Approach	Key Tool	Typical # Runs
Batch, <5 factors, screening	Full/Fractional Factorial	pyDOE3	8-32
Batch, 6-15 factors, screening	Plackett-Burman, DSD	pyDOE3, dexpy	12-50
Batch, optimization, 2-5 factors	CCD, Box-Behnken	pyDOE3, pycse	13-50
Batch, exploration, unknown	Latin Hypercube	pyDOE3, pycse	10×factors
Sequential, expensive, optimize	Bayesian Optimization	BoTorch, skopt, Ax	10-50
Sequential, build model	Active Learning + GP	modAL	20-100
Have mechanistic model	Model-driven D-optimal	Custom FIM	10-30
Parameter estimation	D-optimal, A-optimal	pyoptex	Varies
Model discrimination	T-optimal, KL-optimal	Custom	10-20
Multiple objectives	Multi-objective BO	BoTorch	30-100
Mixture components	Simplex designs	pyDOE3	10-30

Quick Start Examples

Example 1: Classical Response Surface (Batch)

Situation: Optimize 3 factors, batch experiments, moderate cost

import pyDOE3 as pyd
import pandas as pd

# Generate Central Composite Design
n_factors = 3
design = pyd.ccdesign(n_factors, center=(0, 4))  # With center points

# Create DataFrame with factor names
factors = ['Temperature', 'Pressure', 'Catalyst']
df = pd.DataFrame(design, columns=factors)

# Scale to actual ranges
ranges = {'Temperature': (300, 400),
          'Pressure': (1, 5),
          'Catalyst': (0.1, 1.0)}

for factor, (low, high) in ranges.items():
    df[factor] = df[factor] * (high - low)/2 + (high + low)/2

print(f"Generated {len(df)} experiments")
print(df.head())

# Export for lab work
df.to_csv('experimental_design.csv', index=False)

Next steps:

1. Run experiments and collect responses
2. Analyze with statsmodels or pycse
3. Optimize using fitted model

Example 2: pycse Surface Response

Situation: Quick RSM with integrated analysis

from pycse import design_sr, analyze_sr, sr_parity
import numpy as np

# Generate design
bounds = np.array([[300, 400],   # Temperature
                   [1, 5],        # Pressure
                   [0.1, 1.0]])   # Catalyst

design = design_sr(bounds,
                   inputs=['Temperature', 'Pressure', 'Catalyst'],
                   outputs=['Yield'])

print(design)

# After running experiments, add results
design['Yield'] = [78, 82, 75, 88, 91, 85, 79, ...]  # Your data

# Analyze
anova_table = analyze_sr(design)
print(anova_table)

# Parity plot (model fit)
sr_parity(design, show=True)

Advantages: Integrated workflow, automatic ANOVA, quick visualization

Example 3: Bayesian Optimization (Sequential)

Situation: Expensive experiments, 4 factors, want to minimize total runs

import numpy as np
from skopt import gp_minimize
from skopt.space import Real
from skopt.plots import plot_convergence

# Define parameter space
space = [
    Real(300, 400, name='Temperature'),
    Real(1, 5, name='Pressure'),
    Real(0.1, 1.0, name='Catalyst'),
    Real(10, 60, name='Time')
]

# Your black-box function (returns value to minimize)
def expensive_experiment(params):
    temp, pressure, catalyst, time = params

    # Run actual experiment here
    # For now, simulate
    yield_value = run_experiment(temp, pressure, catalyst, time)

    return -yield_value  # Minimize negative = maximize yield

# Bayesian optimization
result = gp_minimize(
    expensive_experiment,
    space,
    n_calls=30,          # Maximum experiments
    n_random_starts=5,   # Initial random exploration
    acq_func='EI',       # Expected Improvement
    random_state=42
)

print(f"Best parameters: {result.x}")
print(f"Best yield: {-result.fun:.2f}")

# Visualize convergence
plot_convergence(result)

Workflow:

1. Start with 5 random experiments (exploration)
2. Fit GP surrogate model
3. Suggest next experiment (maximize EI)
4. Run experiment, update model
5. Repeat until convergence

Example 4: Model-Driven D-Optimal (Parameter Estimation)

Situation: Have kinetic model, need to estimate 3 rate constants

import numpy as np
from scipy.optimize import minimize
from scipy.linalg import det

# Your mechanistic model
def model(t, k1, k2, k3):
    """Concentration vs time model"""
    return k1 * (1 - np.exp(-k2 * t)) + k3 * t

# Fisher Information Matrix for given design points
def fisher_information(t_design, k_guess):
    """Calculate FIM for design points t"""
    k1, k2, k3 = k_guess

    # Jacobian (sensitivity matrix)
    J = np.zeros((len(t_design), 3))

    for i, t in enumerate(t_design):
        J[i, 0] = 1 - np.exp(-k2 * t)
        J[i, 1] = k1 * t * np.exp(-k2 * t)
        J[i, 2] = t

    # FIM = J^T * J (assuming constant variance)
    FIM = J.T @ J
    return FIM

# D-optimal criterion: maximize determinant of FIM
def d_optimal_criterion(t_design, k_guess):
    FIM = fisher_information(t_design, k_guess)
    return -np.log(det(FIM))  # Maximize det = minimize -log(det)

# Find optimal design points
k_initial_guess = [1.0, 0.1, 0.05]
n_points = 6

result = minimize(
    lambda t: d_optimal_criterion(t, k_initial_guess),
    x0=np.linspace(0, 100, n_points),
    bounds=[(0, 100)] * n_points,
    method='L-BFGS-B'
)

optimal_times = result.x
print(f"Optimal measurement times: {optimal_times}")

# Run experiments at these times
# After getting data, update k_guess and re-optimize design if sequential

Sequential workflow:

1. Initial guess for parameters
2. Calculate D-optimal design
3. Run experiments
4. Fit model, update parameter estimates
5. Recalculate optimal design with new estimates
6. Repeat until parameters converge

Example 5: Active Learning (Build Surrogate)

Situation: Want accurate surrogate model, can run sequentially

from modAL.models import ActiveLearner
from modAL.uncertainty import uncertainty_sampling
from sklearn.gaussian_process import GaussianProcessRegressor
from sklearn.gaussian_process.kernels import RBF

# Initial random sample
X_initial = np.random.uniform([0, 0], [10, 10], size=(5, 2))
y_initial = [expensive_function(x) for x in X_initial]

# Create active learner with GP
regressor = GaussianProcessRegressor(kernel=RBF())
learner = ActiveLearner(
    estimator=regressor,
    query_strategy=uncertainty_sampling,  # Query where uncertain
    X_training=X_initial,
    y_training=y_initial
)

# Active learning loop
n_queries = 20
for i in range(n_queries):
    # Query next experiment (highest uncertainty)
    query_idx, query_instance = learner.query(X_candidate_pool)

    # Run experiment
    y_new = expensive_function(query_instance[0])

    # Update model
    learner.teach(query_instance, y_new)

    print(f"Iteration {i+1}: GP std = {learner.predict(query_instance)[1]:.4f}")

# Final model is in learner.estimator

Workflow: How to Use This Skill

Step 1: Answer Questions

When starting a DOE project, expect these questions:

1. What's your experimental goal?
2. Can you run sequentially or batch only?
3. How expensive are experiments?
4. How many factors?
5. Do you have a model or prior data?
6. Any constraints on factor combinations?
7. What's your total budget (# experiments)?

Step 2: Get Recommendation

Based on answers, receive:

• Recommended approach with rationale
• Alternative options with trade-offs
• Estimated number of runs needed
• Library/tool recommendations
• Example code to get started

Step 3: Generate Design

Use provided code to:

• Generate design matrix
• Visualize design in factor space
• Check design properties
• Export to CSV for lab work

Step 4: Run Experiments

Execute experiments according to design:

• Randomize run order (unless sequential)
• Record all results
• Note any deviations or issues

Step 5: Analyze Results

Analyze data with appropriate methods:

• Classical: ANOVA, regression, diagnostics
• Bayesian: Update GP, check convergence
• Model-driven: Fit model, assess parameters

Step 6: Make Decisions

Based on analysis:

• Classical: Optimize response, validate
• Bayesian: Continue or stop? Next experiment?
• Model-driven: Parameters converged? Need more data?

Detailed References

For deep dives into specific topics:

• references/DESIGN_SELECTION_GUIDE.md - Complete decision trees with questions
• references/CLASSICAL_DOE.md - Factorial, RSM, screening designs
• references/BAYESIAN_OPTIMIZATION.md - BO theory, acquisition functions, GP models
• references/MODEL_DRIVEN_DOE.md - Optimal designs, Fisher Information
• references/ACTIVE_LEARNING.md - Sequential strategies, uncertainty sampling
• references/ANALYSIS_METHODS.md - Statistical analysis, ANOVA, diagnostics
• references/PYCSE_INTEGRATION.md - Using pycse for RSM workflows

Common Patterns by Industry

Chemical Engineering:

• Reactor optimization → CCD or Bayesian Optimization
• Catalyst screening → Fractional factorial → BO on hits
• Process development → Sequential model-driven DOE

Materials Science:

• Composition optimization → Mixture designs or BO
• Property mapping → Latin Hypercube + GP surrogate
• Alloy discovery → High-throughput with active learning

Pharmaceutical:

• Formulation → Mixture designs, response surface
• Dose optimization → Bayesian optimization
• PK/PD modeling → Model-driven D-optimal designs

Manufacturing:

• Process parameter tuning → CCD or Box-Behnken
• Quality improvement → Taguchi, robust parameter design
• Continuous improvement → Sequential BO

Machine Learning:

• Hyperparameter tuning → Bayesian optimization
• Architecture search → BO with discrete/categorical variables
• Neural network training → Adaptive sampling

Installation

# Classical DOE
pip install pyDOE3          # Factorial, RSM, LHS
pip install dexpy           # Modern DOE library

# pycse for integrated RSM
pip install pycse

# Bayesian Optimization
pip install scikit-optimize # skopt - easiest to use
pip install ax-platform     # Meta's adaptive experimentation (uses BoTorch internally)
# pip install botorch torch # Advanced BO (requires PyTorch)

# Active Learning
pip install modAL           # Active learning framework

# Gaussian Processes
pip install GPy             # GP models
# scikit-learn has GP built-in

# Analysis
pip install statsmodels     # ANOVA, regression
pip install scipy           # Optimization, statistics
pip install scikit-learn    # ML models, cross-validation

# Visualization
pip install matplotlib seaborn plotly

# Sensitivity Analysis
pip install SALib

# All in one
pip install pyDOE3 dexpy pycse scikit-optimize statsmodels scipy scikit-learn modAL GPy matplotlib seaborn

Best Practices

1. Always Start with Questions

❌ Don't: "I'll use a Box-Behnken design" ✅ Do: "My goal is optimization, I have 3 factors, I can only batch, so Box-Behnken makes sense"

2. Match Method to Constraints

• Budget limited → Bayesian optimization
• Time limited → Batch classical DOE
• Knowledge limited → Space-filling exploration
• Have model → Model-driven optimal design

3. Sequential When Possible

Sequential designs (BO, active learning, model-driven):

• Adapt based on results
• Stop early if converged
• Avoid wasted experiments
• Total runs usually fewer than batch

When batch is better:

• Parallel equipment available
• Setup cost dominates run cost
• Well-understood system

4. Start Simple, Add Complexity

Initial phase: Simple screening (fractional factorial, LHS) Refinement: Focus on important factors (RSM, BO) Validation: Confirmation runs

5. Validate Assumptions

Classical DOE assumes:

• Normal residuals
• Constant variance
• Independent observations
• Linear/quadratic model adequate

Check diagnostics: Residual plots, Q-Q plots, lack-of-fit tests

Bayesian optimization assumes:

• GP model appropriate
• Kernel choice reasonable
• Enough initial exploration

Check: GP posterior uncertainty, kernel hyperparameters

6. Use Confirmation Experiments

After finding "optimum":

• Run additional experiments at predicted best settings
• Verify predictions match reality
• Account for prediction uncertainty

Common Pitfalls

1. Using Wrong Design Type

Problem: Applying batch RSM to expensive experiments Solution: Ask questions first, consider sequential approaches

2. Too Few Initial Points (BO)

Problem: GP needs diverse data to model landscape Solution: Start with 5-10 space-filling points (LHS)

3. Ignoring Constraints

Problem: Design suggests infeasible experiments Solution: Specify constraints upfront, use constrained optimization

4. Overfitting Surrogate Models

Problem: GP fits noise, poor predictions Solution: Cross-validation, hold-out test sets, regularization

5. Not Randomizing Run Order

Problem: Systematic effects confounded with factors Solution: Randomize execution order (except sequential designs)

6. Stopping Too Early (Sequential)

Problem: Declare convergence before truly converged Solution: Use stopping criteria (EI < threshold, GP uncertainty low)

Cost-Benefit Analysis: Sequential vs Batch

When Sequential is Worth It:

Expensive experiments:

• Batch CCD: 20 runs at $1000 = $20,000
• Sequential BO: 15 runs (converge early) = $15,000
• Savings: $5,000 + better result

Complex landscapes:

• Batch might miss global optimum
• Sequential adapts to findings
• Better final result

When Batch is Better:

Cheap experiments:

• Batch: 30 runs in 1 day (parallel)
• Sequential: 30 runs over 30 days
• Time savings dominate

Setup costs:

• If setup takes hours, run time minutes
• Batch amortizes setup
• Sequential overhead too high

Response Format

When helping with DOE, Claude should:

1. Ask questions first - Don't assume method
2. Explain recommendation - Why this approach for their situation
3. Provide complete code - Working examples, not pseudocode
4. Show alternatives - "BO is best, but if you must batch, try..."
5. Guide through workflow - Design → Execute → Analyze → Optimize
6. Interpret results - Statistical significance AND practical significance

Additional Resources

• pyDOE3 Documentation: https://pydoe3.readthedocs.io/
• pycse Examples: https://kitchingroup.cheme.cmu.edu/pycse/
• scikit-optimize: https://scikit-optimize.github.io/
• BoTorch: https://botorch.org/
• Bayesian Optimization Book: https://bayesoptbook.com/
• Design of Experiments (Montgomery): Classic textbook
• Statistics for Experimenters (Box, Hunter, Hunter): Comprehensive reference

Related Skills

• python-optimization - Response optimization after DOE
• pycse - Includes regression, ANOVA, confidence intervals for analysis
• python-multiobjective-optimization - Multi-response optimization
• python-plotting - Visualization of results and diagnostics