Bayesian Reaction Optimization

Optionally, upload data

If you have already run some of the experiments in this design space, upload them by pressing “Upload Experiments”. The model will be trained with this data before suggesting subsequent experiments.

Upload a CSV file with the following format:

Define parameter space

Choose the bounds of your optimization campaign. If you uploaded data, the features and objective(s) will already be filled in. Add additional feature values that you want to include in the parameter space. If you did not upload data, type in the feature name, select if it is a numerical or categorical feature, then type in unique feature values and press enter after typing each one.

Define target variable(s)

If you uploaded data, you don’t need to do this. If you didn’t upload data, click the objectives tab. Type in a target variable name, the target value, and if you want to maximize or minimize this variable. Press “Add Objective” if you want to add more target variables for multiobjective optimization.

Select a batch size and press generate experiments

Select a batch size and press generate experiments.

Run suggested experiments, input your data, and update the model

At any point this data can be downloaded by pressing download predictions.

Edit experiment sample space after initial round

Modify the sample space to consider additional reaction conditions or parameters.

Prediction Value

Pressing the prediction value button will add additional columns to your data describing the model’s predicted outcome for the experiment, the variance associated with the prediction, and the expected improvement parameter.

Analyzing predictive power

Pressing predictive power provides SHAP values for each feature. This is a measure of how important that feature is for the model’s prediction of that exact experiment.

Model interpretation (SHAP)

A SHAP beeswarm plot visualizes how each feature contributes to a model's predictions across all data points in a dataset. The plot displays features on the y-axis, ranked by importance. The most important are at the top and % SHAP importance is displayed next to feature names. The X-axis is SHAP value, where positive values indicate the feature pushes the prediction higher and negative values push it lower. Each data point represents a single observation, with the point’s color indicating the feature's value for that observation. The horizontal position of each point shows the magnitude and direction of that feature's impact on the prediction for that specific instance. You can hover over a data point to view details about that exact experiment.

Interactive contour plot

The contour plot will allow you to visualize the surface the model is predicting. First, drag and drop the features and objective you want to analyze into the “plot axes” section. You can use SHAP analysis to guide which features are most important. Press graph.

To set the fixed feature values to values of interest, modify the right panel. To explore the predictions of specific feature combinations on the plot, adjust the left panel or hover over the plot.

Model Interpretation (Surrogate Model)

The image below represents the Gaussian Process which models the objective function based on past reactions. The light blue shaded areas represent uncertainty (variance). The taller the shaded region, the greater the uncertainty.

Model Interpretation (Acquisition Function)

The acquisition function selects the next sample by balancing high potential (mean) and high uncertainty (variance). It estimates how much better on average a new point could be compared to the best one observed so far.