Skip to content

Machine Learning: Predict New Properties

In the present tutorial page, we will explore how the results of the Train Model derived from Machine Learning (ML) can be used to predict new material properties by linear regression, such as implemented by the Exabyte Machine Learning Engine.

In the present example, we consider the Electronic Band Gap calculated in the previous tutorial for the case of Si/Ge-based materials, however the general approach exposed herein can work for many different target properties.


We follow the below steps, by making use of our Web Interface.

  1. Pre-requisite: trained model
  2. Create "ML Predict" job
  3. Select trained model as workflow
  4. Select target properties
  5. Execute "ML Predict" job
  6. View results

1. Pre-requisite: Trained Model

The present tutorial assumes that an ML model contained in the workflow called "ml_predict" has already been trained to predict the band-gap of Si/Ge-based materials, by following the steps outlined in this other tutorial.

2. Create "ML Predict" Job

The general instructions for creating a new Job can be followed for setting up a new "ML Predict" Job, after opening the relevant interface.

3. Select Trained Model as Workflow

The aforementioned "ml_predict" workflow should be selected as the main Workflow for the "ML Predict" Job being designed, so that it can be applied to predict the properties of a new set of target materials similar to the ones used originally to train the model.

4. Select Target Properties

The properties which will be predicted by a trained model are the target properties which have been ticked and selected under the unit editor interface of the "input" unit of the "ml_predict" workflow, under the "Targets" section of the interface.

5. Execute "ML Predict" Job

The reader should follow these instructions in order to finally execute the "ML Predict" job, following its creation with Job Designer.

6. View Results

The newly predicted properties can finally be inspected under the results tab of job viewer.


In the following animation, we demonstrate how the above steps can be followed to predict the band-gap of a new set of Si/Ge-based materials, using the model trained in a previous tutorial. For the sake of this example, we predict the bang-gap for the Si4Ge12 stochiometric composition. The results of the ML prediction for both the direct and indirect band gaps (0.525 and 0.490 eV respectively) are in very good agreement with the values of their direct computation using DFT (0.517 and 0.441 eV respectively).