Estimation of boiling point, refractive index, and dielectric constant by using machine learning [Courtesy of Industrial Technology Center of Wakayama Prefecture]

Machine learning methods such as GCN and XGBoost were used to construct models that predict boiling point, refractive index, and relative permittivity from chemical structure. Transfer learning improved prediction accuracy for silicon-based compounds, showing promise for applications in materials informatics.

Use Cases Highlights

Development of property prediction models using machine learning, transfer learning, and integration with simulations
High-accuracy prediction of material properties
Improvement of learning models using transfer learning

Development of property prediction models using machine learning, with examples of transfer learning and integration with simulations

A correlation between predicted and experimental boiling points from a machine learning model is shown. The GCN-based model shows high correlation for both training and test data, confirming the capability for reliable prediction.

Results of a machine learning model for boiling point

Evaluation of thermal properties

A comparison between predicted and measured boiling points for new compounds is shown. Predictions were made for two types of compounds, both of which showed good agreement with measured values.

Comparison between predicted and measured values for a new compound

Evaluation of optical properties

A learning model was constructed using 658 refractive index data from the literature, yielding good prediction accuracy for hydrocarbon compounds. In contrast, errors were large for silicon-based compounds, so transfer learning was performed with an additional 28 data points, significantly improving prediction accuracy. Transfer learning enabled efficient model enhancement with a small amount of data.

Learning results using transfer learning

Reference

[1] (in Japanese) https://www.wakayama-kg.jp/docs/keisankagaku-jirei-12.pdf
[2] (in Japanese) https://www.wakayama-kg.jp/docs/keisankagaku-jirei-16.pdf;
https://www.wakayama-kg.jp/docs/keisankagaku-jirei-17.pdf
[3] Presentation Material of Industrial Technology Center of Wakayama Prefecture (WINTEC) included in J-OCTA Special Lecture "Introduction of Materials Development by Materials Informatics"　
[4] “AIST Materials Gate Data Platform” (DPF)

Details of analysis