Machine learning transition temperatures from 2D structure.

Area of Science:

• Computational chemistry
• Materials science
• Data science

Background:

• Physicochemical properties like melting and boiling points are vital for materials discovery but challenging to model theoretically.
• Data science offers powerful tools for predicting material properties from chemical datasets, complementing traditional methods.

Purpose of the Study:

• To extend the Unified Physicochemical Property Estimation Relationships (UPPER) molecular representation for improved materials property prediction.
• To adapt UPPER descriptors for machine learning techniques to predict thermal transition temperatures.

Main Methods:

• Extended the UPPER molecular representation to include descriptors for sp2-bonded fragments.
• Constructed vector representations from UPPER descriptors for machine learning models.
• Utilized a gradient-boosting decision tree model for predicting transition temperatures.

Main Results:

• The enhanced UPPER representation combined with machine learning accurately predicts experimental transition temperatures across diverse chemical spaces.
• The method demonstrated predictive power for energetic materials, even with a limited dataset.
• Achieved competitive results on large public datasets of melting points, including OCHEM, Enamine, Bradley, and Bergström (over 47k structures).

Conclusions:

• The adapted UPPER representation provides an efficient and accurate framework for predicting material transition temperatures using machine learning.
• This data-driven approach accelerates materials discovery by enabling rapid virtual screening.
• Open-source software is available, facilitating broader adoption and further research.