Generative Models for Extrapolation Prediction in Materials Informatics.

Area of Science:

• Materials Informatics
• Computational Chemistry
• Machine Learning

Background:

• Materials informatics often faces challenges with incomplete experimental datasets.
• Predicting material properties from limited data is crucial for discovering new functional materials.
• Extrapolation prediction, vital for novel material exploration, remains difficult with conventional methods.

Purpose of the Study:

• To develop a deep generative model for regression tasks within materials informatics.
• To create a data imputer capable of predicting missing experimental properties of organic molecules.
• To enhance the prediction accuracy and extrapolation capabilities for materials data.

Main Methods:

• Implementation of a deep generative model as a component of a data imputer.
• The imputer "imagines" missing data points to utilize incomplete material databases.
• Evaluation of the model's performance on predicting over 20 diverse experimental properties of organic molecules.

Main Results:

• The data imputer maintains high prediction accuracy even when 60% of the data is removed.
• The model demonstrates superior performance in extrapolation prediction, handling data outside the training range.
• Prediction performance improved by over 30% compared to traditional linear regression and boosting models, especially for properties with limited data (<100 cases).

Conclusions:

• The deep generative data imputer effectively handles incomplete materials data and improves prediction accuracy.
• The approach facilitates efficient exploration of functional materials and overcomes limitations of conventional methods.
• This method shows significant promise for accelerating the discovery of novel materials with desired properties.