This presentation will detail the steps used to build a TensorFlow machine learning model to predict the band gap of semiconducting materials. The band gap is a fundamental property determining photovoltaic-device efficiency for applications such as solar panels. The optimal band gap range for these applications is 0.5eV - 2.5eV. Our dataset was generated from the open source Materials Project and the Python Materials Genomics (pymatgen) code and database. After generation of the dataset and elemental physical descriptors (features), we proceeded to create a “pipeline” of models to facilitate feature engineering and feature selection. First, we developed a support vector machine regression model with a radial-basis (gaussian) kernel and employed a grid search to get a RMSE of 0.45eV. Although this model gave good results, each dataset instance consisted of 128 features, and we sought to reduce this initial feature space to avoid model complexity and overfitting.
As part of feature engineering through our “model pipeline”, we employed Random Forest and its feature importance measures to reduce the feature space dimensionality. Finally, with the reduced feature space, we build a machine learning model in TensorFlow with L2 regularization (ridge regression). Through judicious hyperparameter tuning of our TensorFlow model and using element-specific atomic descriptors, we predict the band gaps of more than 75 AB binary compounds suitable for photovoltaic applications. A root mean squared error (RMSE) of 0.2 eV is achieved using the TensorFlow model with only a 10-dimensional set of descriptors, reduced from the 128-dimensional feature space we started with!
The resulting TensorFlow model can be deployed to discover new materials that can be used in applications with the goal of reducing the world’s carbon footprint such as solar panels for generating electricity and water splitting devices for hydrogen fuel cells.
You need this ticket from Eventbrite to sign up:
Applied Machine Learning Conference.
