Group Contribution-Based Graph Convolution Network: Pure Property Estimation Model

Properties data for chemical compounds are essential information the design and operation of processes. Experimental values reported in literature, but that too scarce compared with exploding demand data. When not available, various estimation methods employed. The group contribution method is one standards simple techniques used today. However, these have inherent inaccuracy due to simplified representation molecular structure. More advanced emerging, including improved representations handling experimental such processes also suffer from a lack valid adjusting many parameters. We suggest compromise between complex machine learning algorithm linear this contribution. Instead representing molecule using graph atoms, we employed bulkier blocks—a functional groups. new approach dramatically reduced number adjustable parameters learning. result shows higher accuracy than conventional methods. whole process was examined aspects—incorporating uncertainties data, robustness fitting process, detecting outlier