Machine Learning-Aided Property Prediction of Hybrid Organic-Inorganic Perovskites Using Hirshfeld Surface Representations of Crystal Structures

This paper describes a lightweight neural network (NN)to predictthermodynamic, electric, and electronic properties of hybrid organic-inorganicperovskites (HOIPs) using Hirshfeld surfaces as novel material representationfor HOIPs. The neural network utilizes only a few Hirshfeld surfacefeatures (e.g., volume, surface area, globularity, and effective radius),along with qualitative and quantitative (mixed) variables, to predictthe properties of HOIPs in a highly accelerated manner. Our use ofHirshfeld surface-based descriptors of HOIP crystals leads to a newmetric for measuring the effective radius of an organic molecule withina given structure, which are proven to be highly effective featuresfor efficient machine learning of crystalline materials' properties.A detailed comparison between the crystal graph convolutional neuralnetwork (CGCNN) and the Hirshfeld surface-based neural network analysisvia UMAP and HDBSCAN clustering is provided to assess the efficacyof these methods for different compound chemistries. It is shown thata combination of lower-order feature representation and a shallowlightweight neural network is capable of predicting material propertiesfor HOIPs. Benchmarking against well-established denser crystal propertyprediction techniques such as the CGCNN and deeper graph attentionlayer graph neural network (deeper GATGNN) shows that our approachprovides comparable and, in some cases, even superior predictive performanceof properties such as formation energy, band gap, and electronic dielectricconstant but all at much lower computational cost.