The optimal scientific workflow in EDOS relies on three main stages:
1
DoE
Plan a robust set of initial experiments to optimally explore the parameter space
→
2
BO
Refine conditions and optimize towards targets over successive physical lab campaigns
→
3
SA
Review the data to uncover underlying trends, interactions, and physical insights
You should start by planning a new experiment using the DoE module. Once you have initial data, you can continuously refine the conditions using one or several campaigns of Bayesian Optimization (BO). Finally, you can review the resulting comprehensive data using the Statistical Analysis (SA) module to extract physical insights and build predictive estimators.
However, the three modules—DoE, BO, and SA—can be used separately if you only have specific projects or aims.
1. Design of Experiments (DoE) PLANNING
Use this module when starting from scratch to create a statistically sound plan for your first experiments.
Step-by-Step Instructions
Step 1: Define Features
Go to the Features tab. Add your input variables. Choose Continuous for smooth numeric ranges, Regular for fixed-step grids (e.g., `[20, 100]5` for 5 points), Discrete for specific values, or Categorical for non-numerical options (e.g., Catalyst A, B, C).
Step 2: Add Objectives
Switch to the Objectives tab. Enter the names of the metrics you plan to measure (e.g., Yield, Purity, Cost).
Step 3: Configure Tweaks
In the DoE Tweaks tab, choose your design model. Box-Behnken is great for secondary surfaces, while Definitive Screening is efficient for many variables. Latin Hypercube Sampling (LHS) generates a space-filling design that works with any number of factors — ideal when you want maximum coverage of the parameter space with a free choice of N experiments. Set your preferred number of experiments.
Step 4: Generate & Export
Click "Create table of experiments". Review the suggested matrix and click "Export CSV". The app will calculate quality metrics:
Orthogonality: How independent the features are from each other (closer to 100% is better).
D-Efficiency: The overall statistical efficiency of the parameter space coverage.
Resolution: Indicates what types of interaction and quadratic effects can be safely estimated without confounding.
Curvature: Whether the design permits modeling of non-linear (curved) responses.
2. Bayesian Optimization (BO) OPTIMIZATION
The "brain" of the app. It analyzes your current data and recommends the next best settings to reach your goals.
Step-by-Step Instructions
Step 1: Load your Data
Drag and drop your experimental CSV into the upload zone. Ensure your results are filled in the spreadsheet before uploading.
Step 2: Configure Goals
In the Objectives tab, tell the AI what to do for each column: Maximize, Minimize, or hit a specific Target. Click the "Include" checkboxes for the goals you want to focus on now.
Step 3: Review Pareto Trade-offs
Look at the Pareto Front chart. Orange markers show experiments that achieve the best balance of all your objectives simultaneously.
Note: The Pareto plot displays raw unweighted data. Even if objectives have different importance, the chart shows the physical units of your measurements to keep the visualization clear and grounded in reality.
Step 4: Run the Optimizer
Click "Run Bayesian Optimization". The AI will use GPU acceleration (if available) to simulate thousands of possibilities and pick the top recommendations based on your configurations.
Configuring BO Tweaks
Batch Size: How many new experiments the AI should propose for your next lab campaign.
Optimization Strategy:
Gradient-based (Default): Fast and precise for continuous parameters.
Exhaustive Grid (EDBO+): Tests every possible combination of a defined grid. Ideal for highly discrete spaces or when you want to "snap" to specific equipment settings.
Acquisition Function: The strategy the AI uses to score candidates (e.g., LogEI for confident improvements, or qUCB for balanced exploration).
Adaptive Multi-Objective Strategy:
Discovery Mode: Active when all objectives have equal importance (e.g., 50/50). Uses NEHVI to find the best overall trade-offs.
Priority Mode: Active when you set unequal importance (e.g., 90/10). Uses Weighted Scalarized EI to strictly prioritize your "High Importance" goals.
Kernel: The mathematical assumption about the physical surface (Matern 5/2 is standard; RBF is smoother).
Exploration vs. Exploitation: A slider to tell the AI whether to take risks (focus on unknown areas) or be greedy (focus near known optimal areas).
Noiseless: Check this only if your lab measurements have absolutely zero variance/error.
Avoid Re-evaluations: Prevents the AI from proposing an experiment that has the exact same parameters as one you've already done.
The Commitment Flow (Data Management)
When the BO module suggests experiments, they appear in a green-highlighted table.
Perform those experiments in the lab and type the outcomes directly into the cells in that table.
Once all result boxes are filled with numbers, click "Add all to Main Dataset".
This merges your new data with the old, and instantly refreshes all plots, ready for the next round!
3. Statistical Analysis (SA) INSIGHTS
Go beyond optimization to understand why your experiments are behaving the way they do.
Step-by-Step Instructions
Step 1: Select Features & Targets
Load your data, then go to the SA module. Check the boxes for the features you want to analyze and the objectives you want to predict.
Step 2: Choose your Model
In SA Tweaks, select Linear Regression for simple trends, or Neural Network/MLP if you suspect complex, non-linear interactions.
Step 3: Run Analysis
Click "Run the Statistical Analysis".
Step 4: Exporting Results
Click the "Export Analysis Report" button to download a fully interactive HTML dashboard of your current results. This standalone file embeds all your analysis data and dynamic plots, allowing you to zoom, pan, and explore categorical interactions offline without needing the app.
Understanding the Results
Correlation Matrix: A heatmap showing how strongly each numeric feature affects your goals.
Categorical Interactions: Interaction matrices showing how different combinations of levels (e.g., specific catalysts at specific pressures) impact the success score.
Feature Impact (SHAP Analysis): A dual-panel interactive plot revealing exactly how each parameter drives the outcomes.
Left Panel (% Impact): Bar charts showing the macro-level relative importance of each variable across the whole dataset.
Right Panel (SHAP Beeswarm): A colored swarm of your actual experiments. Dots pushed to the right (+ SHAP) increased the objective, dots pushed to the left (- SHAP) decreased it. Red dots mean the parameter setting was high, Blue means it was low.
Estimator Tool: Use the sliders at the bottom to predict What-If scenarios. The tool gives you a predicted value and a "Success Score" based on your defined targets.