1. Double-click the Materials Science icon
   • (No icon? See Starting Maestro)

Figure 2-1. Change Working Directory option.

2. Go to File > Change Working Directory
3. Find your directory, and click Choose
4. Pre-generated files are included for running jobs or examining output. Download the zip file here: schrodinger.com/sites/default/files/s3/release/current/Tutorials/zip/ml_formulations_applied.zip
5. After downloading the zip file, unzip the contents in your Working Directorythe location where files are saved for ease of access throughout the tutorial

Figure 2-2. Save Project panel.

6. Go to File > Save Project As
7. Change the File name to ml_formulations_applied_tutorial, click Save
   • The project is now named ml_formulations_applied_tutorial.prj

Figure 3-1. The Formulation Machine Learning panel after opening.

The provided .csv file is loaded directly into the Formulation Machine Learning panel (not into MS Maestro):

1. Go to Tasks > Materials > Informatics > Formulations Machine Learning
   • The Formulation Machine Learning panel opens

Figure 3-2. Loading the training data into the panel.

2. Click Load CSV …
3. Navigate to the provided files, presumably in your working directorythe location where files are saved, select Section_03 > Pharmaceutical_Formulations-2025BAOTowards-train.csv and click Open
   • The training CSV data contains the drug as SMILES_0, first solvent as SMILES_1, and second solvent (if present) as SMILES_2. The composition of the drug (comp_0) is fixed at 50%, and relative compositions of the solvents (comp_1 and comp_2) are set such that they sum up to 50%.
   • The panel is populated with the training data

Figure 3-3. Viewing the training data in the panel.

Take a moment to view the data in the panel. Note that the data is editable within the panel. To alter any molecules, click on the corresponding SMILES string to open the 2D Sketcher. Click on the row to edit the values. Importantly, editing the data in the panel will not edit the imported .csv file. If you edit the structure, you can create a new CSV file through the Export Data button to save your changes. For this example, we will not edit the data.

Figure 3-4. Showing the LogS property.

The panel offers a useful way to visualize the distribution of the data set prior to model construction.

4. On the right side of the panel, for Properties, choose LogS
   • Statistical parameters such as R², root-mean-squared error (RMSE) and Pearson’s r correlation coefficient are printed, and can be toggled on and off with the checkboxes above the plot

Figure 3-5. Setting the parameters on the Build tab.

5. Go to the Build tab
6. Select Fingerprint and Graph Representation as the Featurizers
   • Deselect all other options
7. Select Dense Neural Networks, Set2Set, and Graph-based Models for the Machine learning models
8. Change the Target property to LogS
   • The target property will be  predicted by the machine learning model
9. Select Temperature (K) as the Descriptors
   • If you have additional descriptors relevant to the target property, you always add them as additional inputs into the machine learning model. In this example, solubility was measured as a function of temperature, which is why temperature is included in the model.
10. Change the Hyperparameter tuning steps to 20

Figure 3-6. Opening the Advanced Options.

11. Open the Advanced Options
    • If additional input features like temperature are missing for data points, you can check the Enable descriptor imputation option to train a model and infer missing features.
12. Check Out-of-sample splitting
    • Instead of randomly splitting, we chose out-of-sample splitting because we have temperature-dependent data that can have the repeated drug-solvent combinations at different temperatures. Out-of-sampling splits the training and testing data such that the test set has unique combinations of structures outside of the training set. Out-of-sampling is a better approach to rigorously measure whether a model can predict new drug-solvent combinations.
13. Click OK to close the panel

Figure 3-7. Naming and running the training calculation.

14. Change the Job name to formulation_ml_train_logS
15. Adjust the job settings () as needed
    • This job requires a CPU host. The job will be completed in about 1 day on a CPU host
16. If you would like to perform the calculation, click Run. Otherwise, we will import pre-generated results in the next step.

By default, the panel will automatically perform a 90:10 train:test split on your dataset, which means we use 90% of the data for training the model and leave out 10% of the data as the test set to evaluate model generalizability to unseen formulations. Performance on the train and test sets are reported after building a model.

Figure 3-8. Loading the model file.

If you performed the calculation, the results will automatically be incorporated into the panel when the job is complete. Here, we will assume that you are proceeding with the provided files:

17. Go to the Performance tab
18. Click Load Model …
19. In the provided files, go to the Section_03 > formulation_ml_train_logS directory, choose the formulation_ml_train_logS.mlform file
20. Click Open
    • If you get a warning saying that the uploaded ML model was trained on an older version, feel free to click okay and proceed. If you wish to retrain the model, you can use the ML Model Manager panel

 

Note: If you performed the calculations yourself, you should expect slight variance in the results arising from the randomness of the train/test split and machine learning model hyperparameters.

v

Figure 3-9. Viewing the model.

The performance tab is populated with the corresponding scatterplot.

Above the plot, the Featurizers and Machine learning models used in building the model are listed.

The plot contains predicted versus actual values from the train and test set, with corresponding R² and RMSE values in a table below.

In this case, we can clearly see that the model generalized well on the test set with an R² of 0.85 (ideal model would have R² of 1).

Figure 3-10. The Predict tab. 

21. In the Formulation Machine Learning panel, go to the Predict tab
22. Ensure that next to the Load Model, the .mlform file is shown. If not, load the formulation_ml_train_logS.mlform file again

Figure 3-11. Loading the prediction input data. 

We can load new formulation data into the panel and use our model to predict solubility values.

23. Make sure the dropdown shows Prediction Input and click Load …
24. Navigate to the provided tutorial files and choose the Pharmaceutical_Formulations-2025BAOTowards-test.csv file. Click Open
    • Note that this test CSV file has unique drug-solvent combinations outside of model training to truly evaluate new formulations.

Figure 3-12. Viewing the prediction input data. 

The panel is updated with the data in the Pharmaceutical_Formulations-2025BAOTowards-test.csv file. Note that the experimental solubilities are included only to compare the quality of the predictions; the practical application of this model would be to predict the solubility of these formulations using only SMILES, composition, and temperature information (i.e. without experimental solubility).

Similar to the “Training Data” tab, the data is editable within the “Predict” tab. To alter any molecules, click on the corresponding SMILES string to open the 2D Sketcher. To edit any values, click on the row to enable editing. For this example, we will not edit the data.

Figure 3-13. Naming and running the job. 

With the trained model and input data loaded, we can run the job to predict solubility for each formulation.

25. Change the Job name to formulation_ml_test_logS
26. Adjust the job settings () as needed
    • This job requires a CPU host. The job will be completed in about 1 minute on a CPU host
27. If you would like to perform the calculation, click Run. Otherwise, we will import pre-generated results in the next step.

Figure 3-14. Loading the prediction output.

If you performed the calculation, the results will automatically be incorporated into the panel when the job is complete. We will assume that you are proceeding with the provided files:

28. Change the dropdown from Predict Input Data to Prediction Output
29. Click Load …
30. Navigate to the formulation_ml_test_logS_predict.csv file in the Section_03 > formulation_ml_test_logS > formulation_ml_test_logS_predict.csv
31. Click Open

Figure 3-15. Viewing the prediction output.

The blue text data columns contain the experimental temperature and solubility. The black text data columns contain the outputs from the machine learning model, such as the predicted solubility and the uncertainty of the predictions.

Figure 3-16. Plotting a scatter plot of the data.

The panel enables quick generation of a scatter plot to assess the performance of the model.

32. Click on the LogS column header to pull up the plotter tool
33. Set the Properties from the dropdowns to LogS and LogS_predict

A scatterplot of predicted solubility versus provided solubility is shown. Statistical parameters such as R², RMSE and Pearson’s r are printed, and can be toggled on and off with the checkboxes above the plot.

Figure 4-1. The Formulation Machine Learning panel after opening.

1. Reset the Formulation Machine Learning panel

Figure 4-2. Loading the training data into the panel.

2. Click Load Training Data …
3. Navigate to the provided files, presumably in your working directorythe location where files are saved, select Section_04 > Viscosity-Bilodeau2023-train.csv and click Open
   • The training CSV file contains the first solvent as SMILES_0 and second solvent (if present) as SMILES_1. Their corresponding compositions are stored as comp_0 and comp_1 such that the compositions sum up to 100%.
   • The panel is populated with the training data

Figure 4-3. Viewing the training data in the panel.

Once again, take a moment to view the data in the panel.

Figure 4-4. Viewing the logV values of the training data.

4. On the right side of the panel, for Properties, choose logV
   • Statistical parameters such as R², root-mean-squared error (RMSE) and Pearson’s r correlation coefficient are printed, and can be toggled on and off with the checkboxes above the plot

Figure 4-5. Setting the parameters on the Build tab.

5. Go to the Build tab
6. Select Fingerprint and Graph Representation as the Featurizers
   • Deselect all other options
7. Select Dense Neural Networks, Set2Set, and Graph-based Models for the Machine learning models
8. Change the Target property to logV
   • The target property is that which we wish to predict with the machine learning model
9. Select Temperature (K) as the Descriptors
   • If you had additional descriptors available with your dataset, you could refer to them here
10. Change the Hyperparameter tuning steps to 10

Figure 4-6. Opening the Advanced Options.

11. Open the Advanced Options
12. Check Out-of-sample splitting
    • Similar to the previous example, this dataset has temperature-dependent viscosity data. Out-of-sampling splits the training and testing data such that the test set has unique combinations of structures outside of the training set. Out-of-sampling is a better approach to rigorously measure whether a model can predict new pure or binary solvent combinations.
13. Click OK to close the panel

Figure 4-7. Naming and running the training calculation.

14. Change the Job name to formulation_ml_train_viscosity
15. Adjust the job settings () as needed
    • This job requires a CPU host. The job will be completed in about 10 hours on a CPU host
16. If you would like to perform the calculation, click Run. Otherwise, we will import pre-generated results in the next step.

By default, the panel will automatically perform a random 90:10 train:test split on your dataset. The panel will use the training set for model training and testing set to evaluate whether the model can generalize to unseen formulations. Performance on the train and test sets are reported after building a model.

Figure 4-8. Loading the model file.

If you performed the calculation, the results will automatically be incorporated into the panel when the job is complete. Here, we will assume that you are proceeding with the provided files:

17. Go to the Performance tab
18. Click Load Model …
19. In the provided files, go to the Section_04 > formulation_ml_train_viscosity directory, choose the formulation_ml_train_viscosity.mlform file and click Open
    • If you get a warning saying that the uploaded ML model was trained on an older version, feel free to click okay and proceed. If you wish to retrain the model, you can use the ML Model Manager panel

Note: If you performed the calculations yourself, you should expect slight variance in the results.

Figure 4-9. Viewing the model.

The performance tab is populated with the corresponding scatterplot.

Above the plot, the Featurizers and Machine learning models used in building the model are listed.

The plot contains predicted versus actual values from the train and test set, with corresponding R² and RMSE values in a table below.

In this case, we can clearly see that the model generalized well on the test set.

Figure 4-10. The Predict tab. 

20. In the Formulation Machine Learning panel, go to the Predict tab
21. Ensure that next to Load Model, the .mlform file is shown. If not, load the formulation_ml_train_viscosity.mlform file

Figure 4-11. Loading the prediction input data. 

We can load new formulation data into the panel and use our model to predict viscosity values.

22. Make sure the dropdown shows Prediction Input and click Load …
23. Navigate to the provided tutorial files and choose the Viscosity-Bilodeau2023-test.csv file. Click Open

Figure 4-12. Viewing the prediction input data. 

The panel is updated with the data in the Viscosity-Bilodeau2023-test.csv file. Note that the experimental viscosities are provided only to compare the quality of the predictions; the machine learning model will be able to predict the viscosity of these formulations with only SMILES, composition, and temperature information.

Similar to the “Training Data” tab,  the data is editable within the “Predict” tab. To alter any molecules, click on the corresponding SMILES string to open the 2D Sketcher. To edit any values, click on the row to enable editing. For this example, we will not edit the data.

Figure 4-13. Naming and running the job. 

With the trained model and input data loaded, we can run the job to predict viscosity for each formulation.

24. Change the Job name to formulation_ml_test_viscosity
25. Adjust the job settings () as needed
    • This job requires a CPU host. The job will be completed in about 1 minute on a CPU host
26. If you would like to perform the calculation, click Run. Otherwise, we will import pre-generated results in the next step.

Figure 4-14. Loading the prediction output.

If you performed the calculation, the results will automatically be incorporated into the panel when the job is complete. We will assume that you are proceeding with the provided files:

27. Change the dropdown from Predict Input Data to Prediction Output
28. Click Load …
29. Navigate to Section_04 > formulation_ml_test_viscosity > formulation_ml_test_viscosity_predict.csv directory
30. Click Open

Figure 4-15. Viewing the prediction output.

The first data column contains the provided temperature and the second data column contains the provided logV. The third and fourth columns contain the outputs from the machine learning model, specifically the predicted viscosity and associated uncertainty.

Figure 4-16. Plotting a scatter plot of the data.

The panel enables quick generation of a scatter plot to assess the performance of the model.

31. Set the Properties from the dropdowns to logV and logV_predict

A scatterplot of predicted logV versus provided logV is shown. Statistical parameters such as R², RMSE and Pearson’s r are printed, and can be toggled on and off with the checkboxes above the plot.

Figure 5-1. The Formulation Machine Learning panel after opening.

1. Reset the Formulation Machine Learning panel

Figure 5-2. Loading the training data into the panel.

2. Click Load Training Data …
3. Navigate to the provided files, presumably in your working directorythe location where files are saved, select Section_05 > Copolymer_Tg_penzel1997Glass.csv and click Open
   • The training CSV file contains binary copolymer systems, where the repeat unit of monomer 1 and 2 is stored as SMILES_0 and SMILES_1, respectively. The compositions (comp_0 and comp_1) denote the extent of these monomers present in the copolymer system.
   • The panel is populated with the training data

Figure 5-3. Viewing the training data in the panel.

Take a moment to view the values in the panel. Each input structure contains a monomer repeat unit of a copolymer. The compositions represent the relative ratios of each monomer unit for a random copolymer system.

Note: Th and Lr are dummy atoms that serve as the head and tail groups of the monomer, and to mark where to connect the monomers to form a polymer system.

Figure 5-4. Viewing the training data.

4. On the right side of the panel, for Properties, choose Tg(K)
   • Statistical parameters such as R², root-mean-squared error (RMSE) and Pearson’s r correlation coefficient are printed, and can be toggled on and off with the checkboxes above the plot

Figure 5-5. Setting the parameters on the Build tab.

5. Go to the Build tab
6. Select Fingerprint and MACCS Keys as the Featurizers
7. Select Dense Neural Networks and Set2Set for the Machine learning models
8. Change the Target property to Tg(K)
9. Change the Hyperparameter tuning steps to 20

Figure 5-6. Opening the Advanced Options.

10. Open the Advanced Options
11. Check Out-of-sample splitting
    • Instead of randomly splitting, we chose out-of-sample splitting because we have a copolymer dataset with varying compositions per copolymer combination. Out-of-sampling splits the training and testing data such that the test set has unique combinations of copolymer structures outside of the training set. Out-of-sampling is a better approach to rigorously measure whether a model can predict new copolymer combinations.
12. Click OK to close the panel

Figure 5-7. Naming and running the training calculation.

13. Change the Job name to formulation_ml_train_Tg
14. Adjust the job settings () as needed
    • This job requires a CPU host. The job will be completed in about 20 minutes on a CPU host
15. If you would like to perform the calculation, click Run. Otherwise, we will import pre-generated results in the next step.

By default, the panel will automatically perform a 90:10 train:test split on your dataset. The panel will use the training set for model training and testing set to evaluate whether the model can generalize to unseen formulations. Performance on the train and test sets are reported after building a model.

Figure 5-8. Loading the model file.

If you performed the calculation, the results will automatically be incorporated into the panel when the job is complete. Here, we will assume that you are proceeding with the provided files:

16. Go to the Performance tab
17. Click Load Model …
18. In the provided files, go to the Section_05 > formulation_ml_train_Tg directory, choose the formulation_ml_train_Tg.mlform file and click Open
    • If you get a warning saying that the uploaded ML model was trained on an older version, feel free to click okay and proceed. If you wish to retrain the model, you can use the ML Model Manager panel

Note: If you performed the calculations yourself, you should expect slight variance in the results.

Figure 5-9. Viewing the model.

The performance tab is populated with the corresponding scatterplot.

Above the plot, the Featurizers and Machine learning models used in building the model are listed.

The plot contains predicted versus actual values from the train and test set, with corresponding R² and RMSE values in a table below.

In this case, we can clearly see that the model generalized well on the test set.

Figure 6-1. The Formulation Machine Learning panel after opening.

1. Reset the Formulation Machine Learning panel

Figure 6-2. Loading the training data into the panel.

2. Click Load Training Data …
3. Navigate to the provided files, presumably in your working directorythe location where files are saved, select Section_06 > Geopolymer_Concrete-Rao2018Quantitative.csv and click Open
   • The training CSV file contains 240 mixtures of fly ash (FA) and ground granulated blast-surface slag (GGBFS) used to create the geopolymer systems, where their compositions are labeled as comp_0 and comp_1, respectively. Chemical structures are ignored since these systems cannot be easily represented as a SMILES; hence, they are denoted as “MISSING” in SMILES_0 and SMILES_1. We use the composition of each ingredient as inputs to machine learning models.
   • The panel is populated with the training data

Figure 6-3. Viewing the training data in the panel.

Take a moment to view the data in the panel. Since no SMILES structure is available, each ingredient is labeled as “comp_0” and “comp_1” to represent their individual compositions.

Figure 6-4. Viewing the training data.

4. On the right side of the panel, for Properties, choose target_compression_strength_MPa
   • Statistical parameters such as R², root-mean-squared error (RMSE) and Pearson’s r correlation coefficient are printed, and can be toggled on and off with the checkboxes above the plot

Figure 6-5. Setting the parameters on the Build tab and running the calculation.

5. Go to the Build tab
6. Select Composition only as the Featurizers
   • Deselect all other options
   • Composition only option will ignore chemical structure and use only the composition as inputs to machine learning models
7. Select All for the Machine learning models
8. Change the Target property to target_compression_strength_MPa
9. Select Powderkg, Liquidkg, WC, Admixturekg, Aggregateskg, temperature as the Descriptors. The additional descriptors are experimental features that were varied and summarized below:
   • Powderkg: Powder of fly ash + GGBFS content in kg/m³
   • Liquidkg: Alkaline solution in kg/m³
   • WC: Water-to-cement ratio / Alkali-binder ratio
   • Admixturekg: Total superplasticizer in kg/m³ (4% of binder)
   • Aggregateskg: Total aggregate in kg/m³
   • temperature: Curing temperature in Celsius
10. Change the Hyperparameter tuning steps to 20
11. Change the Job name to formulation_ml_train_geopolymer
12. Adjust the job settings () as needed
    • This job requires a CPU host. The job will be completed in about 10 minutes on a CPU host
13. If you would like to perform the calculation, click Run. Otherwise, we will import pre-generated results in the next step.

By default, the panel will automatically perform a random 90:10 train:test split on your dataset. The panel will use the training set for model training and testing set to evaluate whether the model can generalize to unseen formulations. Performance on the train and test sets are reported after building a model. Note that for this example, we select random splitting since the compositions of the same ingredients are being varied across the dataset (e.g. extent of FA versus GGBFS). If there were distinct, unique formulations with varying compositions, out-of-sample splitting would be useful to evaluate whether the model can generalize to new formulations as shown in the previous examples.

Figure 6-6. Loading the model file.

If you performed the calculation, the results will automatically be incorporated into the panel when the job is complete. Here, we will assume that you are proceeding with the provided files:

14. Go to the Performance tab
15. Click Load Model …
16. In the provided files, go to the Section_06 > formulation_ml_train_geopolymer directory, choose the formulation_ml_train_geopolymer.mlform file and click Open
    • If you get a warning saying that the uploaded ML model was trained on an older version, feel free to click okay and proceed. If you wish to retrain the model, you can use the ML Model Manager panel

Note: If you performed the calculations yourself, you should expect slight variance in the results.

Figure 6-7. Viewing the model.

The performance tab is populated with the corresponding scatterplot.

Above the plot, the Featurizers and Machine learning models used in building the model are listed.

The plot contains predicted versus actual values from the train and test set, with corresponding R² and RMSE values in a table below.

In this case, we can clearly see that the model generalized well on the test set.

17. Go to the Feature Importance tab

Figure 6-8. Running the Feature Importance.

18. Change the Job name to formulation_ml_feature_importance_geopolymer
19. Click Calculate Feature Importance
    • This calculation will run in the panel in just a couple minutes

 

Figure 6-9. Viewing the Feature Importance.

If you performed the calculation, the results should automatically be incorporated into the panel when the job is complete. Here, we will assume that you are proceeding with the provided files:

20. Go to the Performance tab
21. Click Load Model …
22. In the provided files, go to the Section_06 > formulation_ml_feature_importance_geopolymer directory, choose the formulation_ml_train_geopolymer.mlform file and click Open
    • This is an updated formulation_ml_train_geopolymer.mlform file that includes the calculated feature importance
    • If you get a warning saying that the uploaded ML model was trained on an older version, feel free to click okay and proceed. If you wish to retrain the model, you can use the ML Model Manager panel