4. Machine Learning

At Alchemy, our machine learning (ML) models use input and output variables for training to recognize certain types of patterns and make predictions. Alchemy AI (Figure 4.1) uses two types of ML algorithms depending on the type of output variables:

• Regression machine learning algorithms - For training models with continuous numerical output variables.
• Classification machine learning algorithms - For training models with predefined (categorical) output variables. 

Alchemy’s AutoML module consists of 13 different algorithms for regression and 10 different algorithms for classification, which are trained in parallel in order to significantly shorten the training duration.

Figure 4.1 Alchemy AI Diagram

4.1 Dataset

A dataset consists of input and output variables.

Input variables are independent variables whose values are measured and input by the user. In Alchemy, input variables are:

• Materials (e.g. resin A, solvent B)
• Processing steps (e.g. temperature, mixing speed)

Output variables are variables which depend on input variables. In Alchemy, output variables are:

• Numerical properties (continuous numerical values, e.g. Viscosity, Drying time)
• Predefined numerical properties (numerical categorical values)
• Predefined alphanumeric properties (text or numerical categorical values, e.g. pass/fail)

A dataset for training AI in Alchemy is surfaced through Alchemy’s Scan and Score or Alchemy AI functionality. These tools give information about the:

• Number of matching trials with respect to the requirements added in the material constraints table on the Lab Book Overview record
• Number of relevant trials with respect to the requirements added in the test table on the Lab Book Overview record
• Whether it is possible to train ML models based on the available dataset

The Show More Details button displays the number of available trials for each property separately.

4.2 Train AI

Train AI in Alchemy consists of:

• Hyperparameter Tuning
• Automatic Selection

4.2.1 Hyperparameter Tuning

Hyperparameter tuning is a process which includes searching for the hyperparameters that will produce the highest performance of the models for each ML algorithm.

In Alchemy, performance of ML models are evaluated through repeated k-fold cross validation.

In k-fold cross validation, the dataset is split into k numbers of sets and each time one set is held back and models are trained with the remaining sets. Held back sets are used for performance estimation. This means that a total of k models are fit and evaluated based on the performance of mean value of held back sets. This process is repeated l times with different splits, depending on how large the dataset is. In addition, l ✕ k number of models are fitted with repeated k-fold cross validation for estimating the performance of ML models.

The process of model training is shown in Figure 4.2. 

Figure 4.2. Process of Model Training

4.2.2 Automatic Selection 

In Alchemy, selection of the best model is automatically made for each target property. Automatic selection of the best model consists of the following steps:

1. Get the best performing model from all models for the same algorithm, but with different hyperparameters
2. Get the best model from all models from step 1, one model for each algorithm for regression (13) and/or one model for each algorithm for classification (10)

For automatically choosing the best model (Figure 4.3 and Figure 4.4), different performance metrics are used.

• Regression Models: combined performance metric which relies on mean absolute error (MAE) and root mean square error (RMSE)
• Classification Models:
  • If the predefined output values are balanced throughout the dataset, the system will choose accuracy for finding the best models 
  • If the predefined values are imbalanced throughout the dataset, the system will choose average precision for finding the best models   

Figure 4.3. Automatic Selection of the Best Model for Regression Algorithms
‍

Figure 4.4. Automatic Selection of the Best Model for Classification Algorithms

4.3 Performance Metrics

It is important that we track performance metrics to validate the models we generate in terms of the accuracy of predicted values. 

First, a couple definitions:

• Performance Metric: Validation of the model in terms of actual and predicted values. 
• Actual Value: The test result achieved for the property of a certain trial based on actual testing.
• Predicted Value: The test result which is predicted from machine learning models for the property of a certain trial.

4.3.1 Regression

4.3.2 Classification

4.4 Trained Model Goals

At Alchemy, we strive to achieve three goals when models are trained:

1. Recommend trials with the optimal test results for all target properties
2. Predict target property test result for trials with input variables defined by the user 
3. Get insights to the input variable’s importance

All predicted property values have associated predicted confidence intervals which will show how much deviation can be expected from the predicted property value for a certain trial.

4.5 How to use Machine Learning

Alchemy's Machine Learning functionality — exposed through the Use Alchemy AI button — trains predictive models on your historical trial data and uses them either to recommend new formulations that are most likely to meet your requirements (when model confidence is high) or to recommend exploratory experiments that increase model accuracy in regions where it is currently weak (when model confidence is low). The workflow covers four stages:

• Launching model training from the Lab Book Overview record.
• Reviewing model performance and reading the guidance message that appears below the performance table.
• Exploring how the models behave with the Model Explainer.
• Generating either Find Best Results trials (optimized formulations) or Improve Model Performance trials (exploratory formulations), selecting the ones to test in the lab, and proceeding to testing.

4.5.1 Prerequisites

Before launching Alchemy AI, the Lab Book Overview record must be set up so the system has both the inputs and the outputs needed to train predictive models. The same setup described in Section 2.1 Scan & Score Setup applies here.

Required inputs — Constraints table

• The Constraints table on the Lab Book Overview record must contain the materials and processing steps you want the models to learn from. These are the input variables (predictors) for training.
• Without any entries in the Constraints table, the Use Alchemy AI button is disabled.

Required outputs — Requirements table (Calculations and Tests)

• At least one measured or calculated property must be added to the Requirements table with a priority of Must Have or Nice to Have. Properties with a No target, rate only priority are not used in model training.
• These are the output variables (targets) the models will learn to predict.
• Both regression targets (continuous numerical properties such as Viscosity or Hardness) and classification targets (predefined categorical properties such as Pass / Fail or Crystalline / Powder) are supported. Alchemy automatically applies regression algorithms to numerical properties and classification algorithms to categorical properties (see Section 4.2.2).

Sufficient historical data

• Before model training is attempted, the available historical data must be sufficient for the properties to be predicted. The recommended path is to first run SCAN & SCORE ALL (Section 2.3) or ALCHEMY AI and confirm that the Data for predictive models box reports the data as sufficient or partially sufficient. If the data is insufficient, Alchemy recommends extending the dataset through Design of Experiments (Section 3) before returning to Machine Learning.

4.5.2 Launching Alchemy AI

At the bottom of the Lab Book Overview record is the Use Alchemy AI button (Figure 4.5). What happens when you click it depends on whether models have already been trained for this Lab Book:

• If no models have been trained yet, clicking Use Alchemy AI first opens the Scan & Score results page for the historical data that matches the Requirements and Constraints defined on the record. From there, click Train AI to start training. A "Model training in progress" message is displayed (Figure 4.6) while the system trains models in parallel across multiple algorithms as described in Sections 4.2.1 and 4.2.2, and automatically selects the best-performing model for each target property.
• If models have already been trained for this Lab Book, clicking Use Alchemy AI opens the Alchemy AI results page directly, showing the trained model performances and the Model Explainer (Section 4.5.3) without repeating the training step.

Figure 4.5. Use Alchemy AI button on the Lab Book Overview record.

Figure 4.6. Model training in progress — waiting message

4.5.3 Reviewing model performance

Once training is complete, the Alchemy AI results page opens (Figure 4.7). The top of the page displays the model performance table, which lists every Must Have and Nice to Have property together with a headline score for its best-performing model. The columns shown depend on the type of target:

• Model accuracy [%] — displayed for regression (numerical) properties. Higher is better; the metric is described in Section 4.3.1.
• Avg. precision — displayed for classification (categorical) properties. Ranges from 0 to 1 with higher values indicating better performance, particularly on imbalanced datasets.

Cells are color-coded to make it easy to assess model quality at a glance:

• Green — strong model performance, suitable for reliable predictions.
• Yellow / amber — moderate performance; predictions should be interpreted with caution.
• Red — poor performance; the dataset may need more trials or additional DOE before reliable predictions can be made for this property.

Figure 4.7 Alchemy AI results — model performance table with guidance message and FIND BEST RESULTS / IMPROVE MODEL PERFORMANCE buttons

Guidance message

Below the model performance table, Alchemy displays a guidance message that recommends the next action based on overall model confidence:

• If the performance for every property is at or above 80%, the message reads:

"High model confidence across all properties. Use FIND BEST RESULTS to generate optimized formulations and accelerate your optimization."

• If at least one property is below 80%, the message reads:

"Low or Moderate model confidence for {property_name}. Use IMPROVE MODEL PERFORMANCE to generate trials that explore the design space and increase prediction accuracy."

Where {property_name} is replaced with the name (or comma-separated names) of the underperforming property or properties. The two action buttons — Find Best Results and Improve Model Performance — are always shown together at the bottom of the page; the button matching the recommended action is rendered in primary (filled) style and the other in secondary (outlined) style, so the suggested next step is visually emphasized.

Detailed analysis — Regression and Classification properties

Clicking Show detailed analysis expands the headline table into two tables (Figure 4.8) that report the chosen model and its full set of metrics for each target property:

• Regression properties — columns: Property, Best performing model (e.g., LinearRegression, BayesianRidge, Lasso, ARDRegression), R², Mean absolute error (MAE), Root mean squared error (RMSE).
• Classification properties — columns: Property, Accuracy, F1 score, ROC AUC score, Average precision.

Definitions for these metrics are given in Section 4.3 Performance Metrics. Use these tables to confirm that the chosen model and its error/score values are acceptable for your use case before generating formulations. Clicking Hide detailed analysis collapses the tables again.

Figure 4.8 Detailed analysis — Regression properties and Classification properties tables

4.5.4 Model Explainer

The Model Explainer section below the performance tables lists every target property as a collapsible entry. Expanding a property opens a tabbed view that lets you inspect the model from several angles — from the most influential inputs through to per-trial predictions and "what if" scenarios. The exact set of tabs depends on whether the property is regression-based or classification-based:

• Regression properties (numerical targets) show: Feature Importances, Regression Stats, Individual Predictions, What If…, Feature Dependence.
• Classification properties (categorical targets) show: Feature Importances, Classification Stats, Individual Predictions, What If…, Feature Dependence.

The Feature Importances, Individual Predictions, What If…, and Feature Dependence tabs work the same way for both target types; only the Stats tab differs. A Download button in the top-right of the Model Explainer panel exports the current view for offline review or sharing.

4.5.4.1 Feature Importances

The Feature Importances tab answers the question “Which materials and processing steps matter the most?” and ranks every input variable by its overall impact on the model's predictions. Two scoring methods are available:

• SHAP values (default) — attribute each prediction to its input features using SHapley Additive exPlanations. The bar chart shows the average absolute SHAP value per feature — the higher the bar, the stronger the average influence of that feature on the predicted property.
• Permutation Importances — measure how much the model's performance metric drops when the values of a single feature are randomly shuffled. A larger drop means the feature is more important.

Switch between the two methods using the Importances type dropdown to cross-validate the ranking of influential features. The Depth dropdown can be used to limit how many features are shown.

Figure 4.9 Feature Importances tab — SHAP values / Permutation Importances controls and the Feature Descriptions panel

4.5.4.2 Regression Stats (properties with numerical non predefined values)

The Regression Stats tab (Figure 4.10) provides formal validation of model quality for the chosen target. It includes:

• Predicted vs. Observed scatter plot (Figure 4.10) — points should cluster along the diagonal (predicted ≈ actual). The closer the points to the diagonal, the more reliable the model.
• Residual plots (Figure 4.10) — show the residual (actual minus predicted) plotted against either the predicted value, a chosen feature, or the trial index. Residuals should look randomly scattered around zero; a clear trend, funnel, or curve indicates the model is systematically wrong in some part of the design space. Use the Horizontal axis dropdown to choose what is plotted on the X axis (Predicted, or any input feature) and the Residual type dropdown to switch between Difference, Ratio, and Log ratio residuals.

Figure 4.10. Regression Stats — residual plots with Horizontal axis and Residual type controls

4.5.4.3 Classification Stats (properties with predefined values)

For classification targets, the Stats tab is replaced with Classification Stats, which validates the model's ability to distinguish between classes. It contains:

• Metrics table (Figure 4.11) — Accuracy, F1, ROC AUC, and Average precision. Hovering on a metric reveals a short plain-English explanation (for example, "46% of predicted labels was predicted correctly" for Accuracy).
• Confusion Matrix (Figure 4.11) — a grid of observed vs. predicted classes. High percentages on the top-left to bottom-right diagonal indicate accurate predictions; high values off the diagonal indicate where the model confuses classes. Three controls let you adjust how the matrix is rendered: a Cutoff prediction probability slider sets the probability threshold above which a sample is classified into the positive class, a Highlight percentage checkbox toggles the percentage labels inside each cell, and Normalisation radio buttons (Overall / Observed / Predicted) change whether the percentages are normalized across all samples, within each observed class, or within each predicted class.

Figure 4.11. Classification Stats — Confusion Matrix with cutoff probability slider and normalisation controls

• Precision and Classification plots (Figure 4.12) — visualize how predicted-probability values map to actual outcomes; clear separation between the class distributions and a steady upward Precision line indicate that the model's confidence scores are well calibrated.
• ROC AUC plot (Figure 4.12) — the ROC curve traces the trade-off between True Positive Rate and False Positive Rate across decision thresholds. A curve that bows toward the top-left and an AUC value close to 1 indicate a strong model.
• PR AUC plot (Figure 4.12) — the Precision–Recall curve traces the trade-off between precision and recall. A curve that stays high toward the top-right indicates the model finds most positive cases while remaining accurate, which is especially informative on imbalanced datasets.

Figure 4.12. Classification Stats — Precision, Classification plots, ROC AUC and PR AUC curves

• Lift and Cumulative Precision curves (Figure 4.13) — quantify how much better the model performs than random sampling, and the precision you can expect when testing only the top X% of recommended trials.

Figure 4.13. Classification Stats — Lift and Cumulative Precision curves

4.5.4.4 Individual Predictions

The Individual Predictions tab lets you inspect the model's prediction for any specific trial in the training set, and explains why the model produced that prediction. The view differs slightly between regression and classification:

• For regression targets (Figure 4.14) — select a trial via the Select Index dropdown or the Random Index button. The Prediction box on the right reports the Observed value, the Predicted value, and the Residual (smaller residuals indicate a more accurate prediction). Two filters — Predicted range and Absolute residuals — narrow the index list to trials within a chosen predicted-value or residual range.
• For classification targets (Figure 4.15) — the Prediction box reports the predicted probability for each class as a small table, alongside a donut chart. An asterisk next to a class label indicates the observed class for that trial. The Observed State multi-select and Predicted probability range slider filter the index list to trials in the classes and probability ranges of interest.

Figure 4.14. Individual Predictions — regression target with prediction, observed and residual value 

Figure 4.15. Individual Predictions — classification target with class-probability table and donut chart

Two further plots help explain how the model arrived at the displayed prediction:

• Contributions plot (waterfall) (Figure 4.16) — starts from the model's average baseline and shows, step by step, how each feature pushed the prediction up or down. The size of each bar indicates the magnitude of that feature's effect on this specific trial.
• Partial Dependence Plot (PDP) (Figure 4.16) — selects a single feature (e.g., a material) and traces the predicted target as that feature's value sweeps across its range, while the other features are held at the values from the selected trial. This makes it easy to see whether the trend is monotonic, has a sweet spot, or plateaus.

Figure 4.16. Individual Predictions — Contributions waterfall plot (left) and Partial Dependence Plot (right) for a regression target

4.5.4.5 What If…

The What If… tab (Figure 4.17) lets you run virtual experiments without leaving the application. Pick a baseline trial via the Select Index dropdown, then edit any of the values in the Feature Input section — numeric inputs (e.g., material amounts, temperature, time) accept any numerical value, and categorical inputs (e.g., Atmosphere, Lab Equipment) use a dropdown.

Figure 4.17. What If… — baseline selector and editable Feature Input panel.

When the feature value is adjusted, the Prediction box updates instantly:

• For regression targets, the predicted value changes in real time.
• For classification targets, the predicted class probabilities update in real time, making it easy to find the “tipping point” at which the most likely class flips from one category to another (Figure 4.18).

Figure 4.18. What If… for a classification target — class-probability table and donut chart update in real time as feature values are edited.

This view supports both target-seeking (“what do I change to hit this spec?”) and sensitivity analysis (“how much does this property move when I push that input by 10%?”) without spending lab resources on a physical trial.

4.5.4.6 Feature Dependence

The Feature Dependence tab (Figure 4.19) shows how a chosen input feature relates to its impact on predictions across the whole dataset, and how that impact is modulated by a second feature. It contains two paired views:

• SHAP Summary (left) — ranks features by SHAP value and displays a beeswarm plot for each, where each point is a trial. Color encodes the feature's value (red = high, blue = low), so you can read the direction of effect at a glance — e.g., “high values of feature X push the prediction up.” Use the Summary Type dropdown to switch between detailed and aggregated views.
• SHAP Dependence (right) — plots the SHAP value of one feature against its raw value, optionally colored by a second feature. Use the Feature dropdown to choose the X axis feature and Color feature to choose the interacting feature. Distinct color bands or fan-shaped patterns reveal interactions — cases where the effect of one ingredient depends on the level of another.

Figure 4.19. Feature Dependence — Shap Summary (left) and Shap Dependence plot (right).

4.5.5 Find Best Results — generating optimized formulations

Use the Find Best Results path when overall model confidence is high (the guidance message recommends it) and you are ready to ask Alchemy AI for the formulations most likely to satisfy your defined targets.

Click the Find Best Results button at the bottom-left of the results page. A "Generating formulations — Please wait, this process can take a few minutes…" loading message is displayed (Figure 4.20) while Alchemy runs its genetic algorithm on top of the trained models to produce formulation recommendations that optimize toward the defined targets.

Figure 4.20. Find the best Results button and guidance message.

When generation is finished, the AI recommended trials page appears (Figure 4.21). The page is headed "AI recommended trials ranked by your requirements:" and is organized as follows.

Header row

• Matching Order — the rank of each formulation (#1, #2, #3, …), ordered from best to worst match against the defined targets.
• Trial Code — the auto-generated name of the recommended formulation (e.g., Formulation 1, Formulation 19).
• Select for Testing — a checkbox used to mark the formulations you want to carry forward to testing.

Properties table

The first table lists every Must Have and Nice to Have property, showing the target you defined alongside each formulation's predicted value and prediction interval:

• Property, Priority, Target, Unit — repeated from the Lab Book Overview record.
• Value — the model's predicted value for this property in the recommended formulation. Cells are green when the prediction meets the target and red when it does not.
• Prediction Interval [Low – High] — the range within which the true value is expected to fall, reflecting the model's uncertainty. Narrower intervals indicate higher-confidence predictions.

Materials and processing steps table

The second table lists every material and processing step from the Constraints table alongside each formulation's recommended weight or setting:

• Material — the name of the ingredient or processing condition.
• Material properties — the Type and Subtype defined for the material.
• Target — the constraint defined on the Lab Book Overview record.
• Unit — the unit of measurement for the value.
• Weight [g] (or formulating input unit) — the recommended quantity for the formulation. Values inside the defined constraint range are shown with a green background.
• Processing-step groups are expanded but can also be collapsed.
• A Total row at the bottom of the table reports the sum of the formulation's material weights.

Figure 4.21. Find Best Results — properties and materials per recommended formulation.

Selecting trials and proceeding to testing

Tick the Select for Testing checkbox for each formulation you want to test in the laboratory — more than one formulation can be selected. Once at least one formulation has been selected, the Proceed to Testing button becomes active (Figure 4.21). Clicking it closes the AI recommended trials page and creates a new Workspace record containing the selected formulations as actual trials, pre-populated with the recommended material weights and processing-step settings, so laboratory test results can be recorded against them.

4.5.6 Improve Model Performance — generating exploratory experiments

Use the Improve Model Performance path when at least one trained model has low or moderate confidence (the guidance message recommends it; see Figure 4.22). Instead of optimizing toward the targets, this path generates an adaptive design — a small set of new experiments placed in regions of the design space where the model is uncertain or where feature interactions are not yet well understood. Testing those experiments and re-training the models tightens predictions in the weakest areas, so that subsequent Find Best Results runs can be trusted.

How adaptive design works. Adaptive design replaces the traditional one-shot experimental matrix with a sequential strategy in which additional data points are placed where the predictive model shows the largest variance, so each new experiment gives the model the most information possible. 

Running Improve Model Performance

Click the Improve Model Performance button at the bottom of the results page (Figure 4.22). A "Planning next experiments — Please wait, this process can take a few minutes…" loading message is displayed while Alchemy plans the next set of experiments.

Figure 4.22. Improve Model Performance button and guidance message.

When the planning step is finished, the AI suggested experiments page appears (Figure 4.23). The page is organized differently from the Find Best Results page (Figure 4.21, Section 4.5.5): the Materials and processing steps table is shown first, followed by the Properties table with per-property Uncertainty values and a Global Uncertainty score for each experiment. This ordering reflects the purpose of the page — the suggested experiments are not the ones predicted to best meet the targets, but deliberately exploratory experiments designed to reduce uncertainty in the underperforming models.

Header row

The header identifies each suggested experiment:

• Experiment Order — the position of each experiment (#1, #2, #3, …), ordered from most to least informative.
• Experiment Code — the auto-generated name of the suggested experiment (e.g., Experiment 1, Experiment 2).
• Select for Testing — a checkbox used to mark the experiments you want to carry forward to testing.

Materials and processing steps table

The first table lists every material and processing step from the Constraints table alongside each experiment's recommended quantity or setting:

‍

• Material — the name of the ingredient.
• Material properties — the Type and Subtype defined for the material.
• Target — the constraint defined on the Lab Book Overview record.
• Unit — the unit of measurement.
• Weight [g] (or formulating input unit) — the recommended quantity for the experiment. Cells with a green background indicate that the value falls within the defined constraint range.
• Processing-step groups are expanded but can also be collapsed
• Total row at the bottom of the table reports the sum of the experiment's material weights.

Properties table 

The second table reports the model's uncertainty for each Must Have and Nice to Have property in each experiment, plus a single Global Uncertainty score per experiment:

• Target — the property name (Purity, Molar Mass, State, Yield Percentage, …).
• Uncertainty — one column per experiment, expressing how confident the model is in its prediction for that property–experiment pair. Higher values indicate that the experiment falls in a region where additional data would most improve the model. Note that uncertainty values are reported on the property's own scale, so they are not directly comparable between properties.
• Global Uncertainty — a single score per experiment, summarizing the model's overall uncertainty across all target properties. Experiments are ranked by Global Uncertainty, with the most informative ones (highest expected information gain) listed first.

Unlike the Find Best Results page, no prediction interval is shown on this page — the goal is to surface where the models are weakest, not to predict an outcome with confidence.

Selecting experiments and proceeding to testing

As with Find Best Results, tick the Select for Testing checkbox for each experiment you want to run. Any number of experiments can be selected. Once at least one experiment has been selected, the Proceed to Testing button at the bottom-left of the page becomes active. Clicking it closes the page and creates an AI Generated Workspace containing the selected experiments as actual trials, ready to be tested in the laboratory.

After the laboratory results are recorded, return to the Alchemy AI page and re-train the models. The new data should improve the performance of the previously underperforming properties; once their performance crosses the 80% threshold, the guidance message will switch to recommending Find Best Results, and you can move on to optimization.

Figure 4.23. Improve Model Performance — properties and materials per recommended formulation.

5. Interchangeable Materials

In formulation development, you often work with a set of similar ingredients that can be used in place of one another. For example, you might have several solvents that are functionally equivalent but have different costs or properties. Managing these substitutions can be complex, especially when designing large sets of experiments.

The Interchangeable Materials feature is designed to solve this problem by giving you more flexibility and control over your formulations. It allows you to group these similar materials and define clear rules for how they can be substituted for one another in your experiments.

With this feature, you can:

• Group similar materials: Create named groups of ingredients (e.g., "Resin", "Additive") based on their material type.
• Define substitution rules: Specify exactly how many materials from a group can be included in a single trial. For instance, you could require that exactly one solvent from a group is used, or a range of one to three different solvents.
• Combine with other constraints: These interchangeable groups can be used alongside other materials and processing steps in your experimental designs.

Using this feature impacts how the system generates a Design of Experiments (DOE), scores existing trials in Scan & Score, and makes AI recommendations, as all functions will adhere to the rules you define.

5.1. Creating an Interchangeable Material Group

Follow the steps to add an Interchangeable Material Group to your formulation table:

1. To begin, navigate to the Lab Book Overview page. At the bottom of the Constraints table, you will find a button labeled +INTERCHANGEABLE MATERIALS (Figure 5.1). Once you click the button, the "Interchangeable Materials" modal will appear. Here, you will define the materials and rules for your group.

Figure 5.1. Button for adding Interchangeable Materials Group

2. Define the Group: 
   
   • Give your group a unique Group name (Figure 5.2.).
   • Select a material Type and, optionally, one or more Subtypes. This will filter the list of materials you can add to the group in a later step (Figure 5.2.).

Figure 5.2. The Interchangeable Materials Modal set Group name and select Type and optionally Subtype

3. Set the Constraints: This is the core of the feature, where you define the rules for the group.

• First, choose how you want to Define Constraints On by selecting either Group or Material from the dropdown (Figure 5.3.).

Figure 5.3. The Interchangeable Materials Modal Defining Constraints

• If you chose "Group":
  
  • Set the Group target weight (Figure 5.4). You can select "Constant" for an exact value or "Between" for a range. The system will automatically calculate the allowed range for each individual material.

Figure 5.4. The Interchangeable Materials Modal Constraints Defined on a Group level

• If you chose "Material":
  
  • The Group target weight fields will become read-only. You will define the constraints for each material in the table below, and the system will automatically calculate the resulting total for the group.

Figure 5.5. The Interchangeable Materials Modal Constraints Defined on a Material level

4. Add Materials to the Group:
   
   • Click the "+ Material" button.
   • A dialog will appear showing all available materials based on the Type and Subtype you selected earlier.
   • Select the materials you want to include in the group and click "Add".
   • The selected materials will now appear in a table within the main dialog (Figure 5.6-5.8).

Figure 5.6. The Interchangeable Materials modal, showing constraints defined at the group level, including a constant target weight and the added materials.

Figure 5.7. The Interchangeable Materials modal, showing constraints defined at the group level, including a between target weight and the added materials.

5. Configure Individual Material Constraints (if applicable):
   
   • If you chose to define constraints at the Material level in step 3, you can now set the Target Weight for each material in the table. You can set each one to be "Constant" or "Between"  and enter the desired values (Figure 5.8).

Figure 5.8. The Interchangeable Materials modal, showing constraints defined at the material level, added materials with between or constant target weight.

6. Set the Number of Materials:
   
   • At the bottom of the dialog, use the "# of Materials" dropdowns to define the minimum and maximum number of materials from this group that must be used in any given trial. For example, if you have 3 materials in the group, you could specify that each trial must use between 1 and 2 of them.

Figure 5.9. The Interchangeable Materials Modal # of Materials to be used with defined constraints on group level.

Figure 5.10. The Interchangeable Materials Modal # of Materials to be used with defined constraints on material level

7. Add the Group to the Constraints table in Lab Book Overview:
   
   • Once you're satisfied with your configuration, click the "Add" button. The Interchangeable Material Group will be added as a new entry in your Constraints table on Lab Book Overview.

Figure 5.11. The Interchangeable Materials Group added into Constraints table in Lab Book Overview

5.2. Using Interchangeable Materials with Other Features

The rules you define for interchangeable materials directly guide how the platform generates a Design of Experiments (DOE), scores trials in Scan & Score, and creates AI recommendations.

5.2.1. DOE and AI Training with Interchangeable Materials

When an interchangeable material group is added, DESIGN EXPERIMENTS and ALCHEMY AI are adjusted to incorporate the formulating strategy. The formulating strategy determines how the interchangeable materials are varied and combined during experiment design and AI training, ensuring that the generated trials and model outputs respect the defined group relationships.

5.2.2. Scan & Score with Interchangeable Materials

When you use Scan & Score, the system will only show trials that follow the rules you defined for your interchangeable materials. This includes:

• Groups which do not contain 0 in their constraints must be present in the trials
• Trials must contain the correct number of materials from each interchangeable group (e.g., if you set a range of 1-2, only trials with 1 or 2 of those materials will be shown).

The values in the Scan & Score table are color-coded to indicate if they meet the constraints you set. Green means the value is within the defined range while red means it violates the rules.

6. Appendices

The following supplemental material is used to support the above documentation.

6.1. Appendix A — Detailed Statistical Analysis: Reference

This appendix provides a deeper reference for the Detailed Analysis tab of the AI-Powered Statistical Analysis report (Section 2.5.2). It explains, for each section of the report, what the metric or test means, why it matters for Machine Learning, and what action to take if the result is unfavorable.

6.1.1. Dataset Overview

Reports the size of the dataset: number of input variables (independent variables) and number of target variables (dependent variables). Use this as a quick sanity check that the dataset captured by Scan & Score matches what you defined on the Lab Book Overview record.

6.1.2. Variable Type Detection

Confirms which variables Alchemy treats as inputs (predictors) versus targets (outputs), and whether each variable is categorical or continuous. This classification drives the choice of regression vs. classification algorithms in training.

6.1.3. Dataset Quality Assessment

6.1.3.1. Target Coverage Analysis

Reports how many data points per feature are available in the dataset for each target variable. Low coverage relative to the number of features increases the risk of overfitting; the model has too few examples to learn from.

6.1.3.2. Variable Representation per Target

How many times each feature is present in the dataset with a value different than 0. Feature representation that is very low — i.e., the variable is rarely observed or is mostly missing — makes it difficult for ML models to learn a reliable relationship: the model either ignores the feature (treating it as low-variance and uninformative) or overfits based on the few rare samples. This significantly reduces robustness and usually requires manual intervention (deleting the underrepresented materials from the Constraints table on the Lab Book Overview record, which also excludes the few trials that contain them from the dataset).

6.1.3.3. Target Variation Analysis (Coefficient of Variation)

Reports the coefficient of variation (CV) for each target variable. A high CV indicates a large amount of intrinsic variability relative to the mean: the data is either noisy or has a vast, poorly modeled range. The signal (the actual trend) is obscured by the unexplained scatter, which makes it significantly harder for an ML model to achieve high predictive accuracy.

To resolve this, let ALCHEMY AI / TRAIN AI evaluate robust non-linear algorithms such as Gradient Boosting and Random Forests, which cope well with heterogeneous variance, and select the one that performs best on your dataset.

Note that with small sample sizes, the standard deviation is highly unstable and likely to change significantly as more data are added.

6.1.3.4. Constant or Low-Variance Variable Detection

Identifies variables with constant or very small variance. Such variables provide little to no discriminative information to the model: they cannot help distinguish between different outcomes or samples. Including them adds noise and computational overhead without contributing to predictive power. Constant variables are automatically excluded from model training, since they cannot improve predictions. For low-variance variables that are not strictly constant, consider whether the lack of variation is intentional or whether more diverse samples should be collected before continuing.

6.1.3.5. Multicollinearity Detection

Multicollinearity is the statistical phenomenon where two or more independent predictor variables are highly correlated with one another. Multicollinear pairs introduce redundancy and instability, making it difficult to accurately interpret individual feature importance or coefficient values. To resolve this, use ALCHEMY AI / TRAIN AI to evaluate a wider set of algorithms — including regularization techniques such as Ridge and Lasso regression that are designed to handle multicollinearity — and let the system select the algorithm that performs best on your dataset.

6.1.3.6. Normality Testing

Tests whether input variables and target variables are normally distributed. Failing the normality test for target variables primarily concerns Linear Regression (OLS): non-normal target variables usually lead to non-normal residuals, which violates a core OLS assumption and renders p-values and significance tests unreliable. Non-normal input features do not violate core OLS assumptions, but severe skewness still hinders finding clean linear relationships and may require feature transformation.

When target variables are clearly non-normal, use ALCHEMY AI / TRAIN AI to evaluate a diverse set of algorithms — including non-linear models that do not require the normality assumption — and let the system select the one that performs best on your dataset.

6.1.3.7. Outlier Detection

Tests for outliers using the modified z-score and informs the user which input or target values are flagged. Outliers in either inputs or targets typically reduce model performance and reliability: extreme values can disproportionately influence the fitted regression line. Input outliers can shift the slope of the line; target outliers increase the model's overall residual error (e.g., MSE).

Recommended action. For input outliers, design additional trials to fill the gap in the design space between the cluster of typical input values and the flagged extreme value. For target outliers, retest flagged trials as replicates to confirm whether the extreme value is reproducible; repeated tests help distinguish a real outlier from a one-time measurement error. If the outlier proves genuine, use ALCHEMY AI / TRAIN AI to evaluate robust algorithms such as Random Forests and Gradient Boosting, which are less sensitive to extreme values, and let the system select the one that performs best on your dataset.

6.1.4. Design Efficiency Criteria

6.1.4.1. G-efficiency

G-efficiency is a score that tells you how well spread out your data points are across the design space. G-optimality seeks to protect against the worst-case prediction variance — i.e., the model’s ability to predict well across the whole design space.

A higher G-efficiency value is better: it indicates a more optimal design where the variance of the predicted response is minimized across the design space, leading to more reliable and stable coefficient estimates. The direct impact on modeling is that the resulting regression model has greater statistical power to detect true factor effects and produce more precise predictions across all factor settings.

6.1.4.2. Distance-based criteria

Assess how uniformly spread out the experiment points are within the multidimensional design space. For ML, maximizing this distance ensures better coverage of the input domain, which increases the model’s ability to generalize and make accurate predictions across the entire range of possible recipes or conditions.

6.1.4.3. Discrepancy criteria

Quantify the non-uniformity or irregularity in the distribution of data points within the multidimensional design space. For ML, minimizing discrepancy means maximizing the space-filling quality of the data — crucial for building accurate predictive models that perform reliably across the entire range of possible input combinations. Low discrepancy (high uniformity) ensures that no large empty gaps exist in the experimental region, making predictions more robust and reducing uncertainty.

6.1.5. Cluster Analysis

Cluster Analysis groups your historical trials by similarity, identifying distinct "families" of formulations within the dataset. For each cluster, the report lists the input variables that are statistically higher or lower than the overall dataset average — these values describe what makes the cluster distinct, not whether it is "better" or "worse" than the others.

Neither many nor few clusters is automatically preferable. What matters is whether the clusters align with formulation strategies you recognize: a few well-defined clusters that map to families you've explored confirms the dataset has structure the model can learn from, while many small clusters or one large indistinct cluster usually indicates a dataset that lacks meaningful sub-structure.

The analysis also serves as a coverage check: if the design region you care about is sparsely represented in the surfaced clusters, the model will potentially struggle to predict there — consider extending the dataset with ALCHEMY AI / TRAIN AI (IMPROVE MODEL PERFORMANCE) to fill the gap.

6.1.6. Feature Importance Analysis

6.1.6.1. Correlation Analysis

Reports the linear relationship between each input variable and the target. Correlation analysis is important for ML because it quickly measures the strength and direction of simple associations between features and the target. From this you learn whether each input feature has high, moderate, or low predictive potential (strong, moderate, or low correlation with the target). The SHAP analysis (Section 6.1.6.2) then refines this picture by capturing non-linear effects and feature interactions that correlation alone can't see.

6.1.6.2. SHAP Analysis

SHAP measures how much each feature affects the prediction of a model. The SHAP value is a quantitative score derived from comparing predictions made when a feature’s information is present (its actual value) versus when its information is absent (replaced by a baseline or average). This comparison lets SHAP accurately attribute the final prediction score to each input.

6.1.7. Regression Model Results

6.1.7.1. Best Model Performance

Identifies the best-performing model, along with the coefficients, p-value, and significance for each predictor.

6.1.7.2. Residual Analysis

Validates the model’s assumptions through diagnostic plots and tests. If these tests fail, the resulting p-values and confidence intervals are unreliable, which prevents the user from trusting the model’s assessment of factor significance and the precision of its predictions. Consequently, the user cannot be confident in identifying the truly influential factors or in optimizing the process based on the model’s conclusions.

6.1.7.3. Unusual Observations

Flags individual data points with High Leverage, High Cook’s Distance, and High DFITS — i.e., observations that exert disproportionate influence on the regression model. High Leverage points have extreme input feature values that pull the fitted line toward themselves. High Cook’s Distance and High DFITS identify data points that, if removed, would significantly change the model’s coefficient estimates and/or predicted values. These points require careful review — they are often outliers or mismeasurements that distort the model.

‍

6.2. Appendix B - Mixture Design

Mixture Design, a subtype of Screening and Optimal Designs, is intended for experiments where the user is interested in the influence of varied dependent variables (weight or volume percentage of each material) on target properties. 

Mixture design is automatically proposed when the Formulating input is set to Target Weight [%] — i.e., when each material's quantity is expressed as a percentage of the total batch rather than as an absolute weight or volume.

Required constraints for each material are:

• Lower and upper bound for varied materials: 
• Constant for non varied materials (influence of this materials cannot be assessed)

Rules which need to be respected for properly setting the constraints:

• Sum of all lower bounds and constant materials should be lower than 100%
• Sum of all higher bounds and constant materials should be higher than 100%

‍

6.3. Appendix C - Mixture Process Design

Mixture Process Design, a subtype of Screening and Optimal Designs, is intended for experiments where the user is interested in the influence of varied dependent variables, materials (weight or volume percentage of each material) and processing steps, on target properties. 

Mixture Process design is automatically proposed when the Formulating input is set to Target Weight [%] and the design includes one or more processing steps in addition to the materials.

Required constraints for each material are:

• Lower and upper bound for varied materials or
• Constant for non varied materials (influence of this materials cannot be assessed)

Required constraints for each condition from processing step are:

• Lower and upper bound for conditions of number type
• Any of for conditions of predefined number or text (alphanumeric) type
• Set to for conditions of predefined number or text (alphanumeric) type which are not varied and
• Constant for non varied conditions  (influence of this conditions of processing steps cannot be assessed)

Rules which need to be respected for properly setting the constraints:

• Sum of all lower bounds and constant materials should be lower than 100%
• Sum of all higher bounds and constant materials should be higher than 100%

6.4 Appendix D - Factorial Design

Factorial Design, a subtype of Screening Design, is intended for experiments where the user is interested in the influence of varied independent variables, materials (values for weight and/or volume) and/or processing steps on target properties. 

Factorial design is automatically proposed when the user inputs weight and/or volume for materials and/or processing steps. 

Required constraints for each material are:

• Lower and upper bound or
• Constant 

Required constraints for each condition from processing step are:

• Lower and upper bound - for conditions of number type
• Any of - for conditions of predefined number or text (alphanumeric) type
• Set to - for conditions of predefined number or text (alphanumeric) type which are not varied and
• Constant - for non varied conditions (influence of this conditions of processing steps cannot be assessed)

Constraints are valid without any rules.

6.5 Appendix E - Response Surface Design

Response Surface Design, a subtype of Optimal Design, is intended for experiments where the user is interested in the influence of varied independent variables, materials (values for weight and/or volume) and/or processing steps, on target properties. 

Response surface design is automatically proposed when the user inputs weight and/or volume for materials. 

Required constraints for each material are:

• Lower and upper bound or
• Constant 

Required constraints for each condition from processing step are:

• Lower and upper bound for conditions of number type
• Any of for conditions of predefined number or text (alphanumeric) type
• Set to for conditions of predefined number or text (alphanumeric) type which are not varied and
• Constant for non varied conditions  (influence of this conditions of processing steps cannot be assessed)

Constraints are valid without any rules.

‍

6.6. Appendix F — Model Explainer Dashboard (regression): Plain-English Guide

This appendix is a non-technical reference for the Model Explainer tabs that are available for regression target properties (continuous numerical outputs such as Viscosity, Hardness, or Drying time). It describes, for each tab, the question the tab answers, what the user is looking at, how to read “good” vs. “bad,” and why the view matters.

Acronyms used in this appendix

RMSE (Root Mean Squared Error). Measured in the same units as the property being predicted; represents the typical distance between the prediction and the true value. Smaller numbers mean predictions are closer to the target.

MAE (Mean Absolute Error). Also in the same units as the property; shows the average distance between predictions and actual values. Lower values mean more consistently accurate predictions.

R² (Coefficient of Determination). Ranges from 0 to 1 and measures how well the predicted values follow the actual data. Values closer to 1 mean the points fall closer to the ideal diagonal line (predicted = actual).

SHAP (SHapley Additive exPlanations). Shown on a positive-to-negative scale where distance from zero indicates how strongly a feature pushes the prediction up or down, and larger magnitudes mean greater impact on reaching or missing the target.

6.6.1. Feature Importance

Question it answers: Which ingredients or process parameters matter the most?

What you’re looking at: A ranked list of features showing which ones have the biggest overall impact on the predicted property. Longer bars indicate bigger influence; shorter bars, smaller influence.

What “good” looks like: A clear set of top drivers that align with your domain intuition; the most important features are also measurable and controllable in production.

Why this matters: Tells you which variables deserve the tightest control in manufacturing, helps R&D focus effort on the inputs that actually move performance, and lets you deprioritize or simplify low-importance features. Rule of thumb: if changing a feature would meaningfully change the prediction, it should appear near the top.

6.6.2. Regression Statistics (Model Performance)

Question it answers: How good is this model overall?

What you’re looking at: Error metrics (RMSE, MAE), the R² score, and a Predicted vs. Actual plot. RMSE and MAE describe how far off predictions are on average — lower is better. R² describes how much of the real-world variation the model explains — higher is better. The Predicted vs. Actual plot shows visually how close predictions are to reality.

What “good” looks like: Low error values, R² close to 1, and points clustered along the diagonal of the Predicted vs. Actual plot.

Why this matters: These tell you whether the model is reliable enough to guide real decisions, not just to generate interesting charts.

6.6.3. Residuals Analysis

Question it answers: Is the model making consistent, unbiased errors?

What you’re looking at: Residuals vs. Predicted and Residuals vs. Feature plots. A residual is the actual value minus the predicted value; you want errors to look random rather than patterned.

What “good” looks like: Points scattered randomly around zero, with no curves, funnels, or trends.

Why this matters: Patterns here mean the model is systematically wrong in certain regions of the design space — risky for optimization or scale-up.

6.6.4. Individual Predictions & Contributions

Question it answers: Why did this specific formulation get this prediction?

What you’re looking at: The predicted versus actual value for one chosen sample, paired with a contribution (SHAP) breakdown that shows how each feature pushed the prediction up or down. The model starts from an average baseline; each feature either increases or decreases the prediction; the sum of those pushes equals the final predicted value.

What “good” looks like: The largest contributions come from known, meaningful features; the explanation matches scientific intuition.

Why this matters: This is where trust is built. Users can see exactly which ingredients or conditions caused a high or low prediction.

6.6.5. What-If Analysis (Scenario Planning)

Question it answers: What happens if I change this ingredient or process setting?

What you’re looking at: Interactive controls to change inputs, with predictions and explanations updating in real time. This is a digital experiment: adjust a concentration, temperature, or parameter, and instantly see how the predicted property changes — and why.

What “good” looks like: Smooth, predictable responses to small input changes, and a clear sense of which inputs the property is most sensitive to.

Why this matters: Enables fast exploration without physical experiments; supports target-seeking (“what do I change to hit this spec?”); and helps avoid risky or extreme combinations before committing to lab work.

6.6.6. Feature Dependence

Question it answers: How does changing one input generally affect the output?

What you’re looking at: Plots showing how a feature’s value relates to its impact on predictions. The X-axis is the feature value, the Y-axis is how much that feature pushes the prediction up or down, and color encodes another feature that may interact with it.

What “good” looks like: Clear trends — increase, decrease, or plateau — and interactions that make physical or chemical sense.

Why this matters: Helps identify non-linear effects, sweet spots, and interactions where two ingredients together behave differently than alone.

Feature Interactions (Synergy vs. Antagonism):

Synergy: two features together have a bigger positive effect than expected. Antagonism: one feature reduces the effect of another. No interaction: the two features behave independently.

What you're looking at: When points stratify into distinct color bands (different colors taking different SHAP-value paths along the X-axis), you have an interaction — the effect of the X-axis feature depends on the value of the color feature. When the colors are randomly mixed throughout the cloud, the two features behave independently and there is no interaction.

Why this matters: Critical for formulation science, where combinations matter more than single ingredients.

6.7. Appendix G — Model Explainer Dashboard (classification): Plain-English Guide

This appendix is a non-technical reference for the Model Explainer tabs that are available for classification target properties (predefined categorical outputs such as Pass/Fail, or any other set of named categories). It mirrors the structure of Appendix F but adapts the explanations to classification metrics.

Acronyms and key terms used in this appendix

ROC AUC (Receiver Operating Characteristic Area Under the Curve). A performance measurement for classification problems at various threshold settings. It tells how well the model is able to distinguish between classes.

PR AUC (Precision–Recall Area Under the Curve). Summarizes the trade-off between precision and recall using different probability thresholds.

Log Loss (Logarithmic Loss). Penalizes false classifications by taking into account the probability of the prediction. Lower values indicate better performance.

SHAP (SHapley Additive exPlanations). Distance from zero indicates how strongly a feature pushes the prediction toward one class or another; larger magnitudes mean greater impact.

Class (Category). The specific group or label the model is trying to predict (for example, “Pass” vs. “Fail”).

Probability. A number between 0 and 1 (or 0% to 100%) that represents the model’s confidence. A 0.85 probability means the model is 85% certain a sample belongs to a specific class.

6.7.1. Feature Importance

As in the regression case (Appendix F), Feature Importance ranks the inputs by their overall impact on the predicted class. Same interpretation, same rule of thumb: longer bars mean a feature has stronger influence on the predicted probability of a class.

6.7.2. Classification Statistics (Model Performance)

Question it answers: How reliable is this model overall?

What you’re looking at: Accuracy, Precision, Recall, F1 Score, ROC AUC Score, PR AUC Score, and Log Loss.

Accuracy. Percentage of total predictions that were correct — higher is better.

Precision. When the model predicts a category, how often is it actually correct — higher is better.

Recall. Out of all the actual instances of a category, how many did the model correctly identify — higher is better.

F1 Score. A balanced metric that combines precision and recall — closer to 1 is better.

ROC AUC Score. 0–1 score that measures the model’s ability to separate classes — closer to 1 is better.

PR AUC Score. Area under the Precision–Recall curve; higher values indicate better performance, particularly on imbalanced datasets.

Log Loss. Measures the “uncertainty” of predictions — lower values mean the model is more confident and correct.

What “good” looks like: High Accuracy, Precision, Recall, F1, ROC AUC, and PR AUC; low Log Loss; and high diagonal percentages on the Confusion Matrix (Observed class = Predicted class).

6.7.3. Interpretation of Plots

The Classification Stats tab includes several plots that go beyond aggregate metrics. Each is described below in plain English.

Confusion Matrix

What it is. A grid of observed vs. predicted classes — a scorecard showing where the model was right and where it got confused.

What “good” looks like. High numbers in the diagonal boxes (top-left to bottom-right) and near-zero numbers everywhere else.

Why this matters. It tells you exactly what kind of mistakes the model makes — does it miss real cases (false negatives) or sound false alarms (false positives)?

Precision and Classification plots

What they are. Two paired views of how “separation” works. The Classification plot shows the distribution of labels relative to a probability cutoff; the Precision plot tracks how the actual positive rate moves with the model’s predicted probability.

What “good” looks like. In the Classification plot, two distinct “humps” (one per category) with very little overlap. In the Precision plot, a steady upward line.

Why this matters. If the humps overlap, the model is “unsure” in that region. If the precision line goes up, you can trust a “90% confidence” score to be right roughly 90% of the time.

ROC AUC and PR AUC plots

What they are. “Stress test” curves that show the trade-off between being thorough and being accurate.

What “good” looks like. The ROC curve should bow deeply toward the top-left; the PR curve should stay high and trend toward the top-right.

Why this matters. These show the overall “strength” of the model. A high AUC means the model is excellent at ranking the best formulations at the top of the list.

Lift and Cumulative Precision curves

What they are. “Value add” curves that show how much better the model performs than random guessing.

What “good” looks like. Curves that start very high on the left and stay well above the diagonal baseline.

Why this matters. They tell you, for example, “If I only have the budget to test the top 10% of samples recommended by the model, how many successes will I find?” — a Lift of 3.0 means three times more successes than picking at random.

6.7.4. Individual Predictions & Contributions

Question it answers: Why did this specific formulation get this prediction?

What you’re looking at: A Contributions plot — a SHAP breakdown showing how each feature pushed the specific prediction up (toward one class) or down (toward another) — and a Partial Dependence Plot, which shows how the predicted probability of a class changes if you vary one specific feature while holding the others constant. The model starts from an average baseline; each feature increases or decreases the probability of the designated positive class; the sum of those pushes equals the final predicted probability.

What “good” looks like: The largest contributions come from known, meaningful features, and the explanation matches scientific intuition.

Why this matters: This is where trust is built. Users can see precisely which factors led to a high or low probability of achieving the positive class.

6.7.5. What-If Analysis (Scenario Planning)

Question it answers: What happens to the final category if I change a temperature, time, or concentration?

What you’re looking at: Interactive controls that update probabilities in real time. Adjust a concentration, temperature, or parameter; instantly see how the probability of a specific outcome shifts; and read the explanation of why it shifted.

What “good” looks like: Smooth, predictable probability shifts where small changes in inputs lead to logical changes in the probability of a category, plus a clear sense of which “tipping points” cause the model to switch its prediction from one category to another.

Why this matters: Enables fast exploration without physical experiments; supports target-seeking (“what do I change to hit this category?”); and helps avoid risky or extreme combinations before committing to lab work.

6.7.6. Feature Dependence

Same idea as the regression view (Appendix F), but the Y-axis here represents how much the feature pushes the predicted probability for the positive class up or down, rather than a numerical target value.

6.7.7. Feature Interactions (Synergy vs. Antagonism)

Same definitions as in Appendix F, Synergy = two features together have a bigger effect on the predicted class than expected; antagonism = one feature reduces the effect of another; no interaction = they behave independently. Reading the Feature Dependence colored scatter: distinct color bands indicate an interaction (the effect of the X-axis feature depends on the color feature); randomly mixed colors indicate no interaction.

‍