No abstract

Vibrational spectroscopy is a nondestructive technique commonly used in chemical and physical analyses to determine atomic structures and associated properties. However, the evaluation and interpretation of spectroscopic profiles based on human-identifiable peaks can be difficult and convoluted. To address this challenge, we present a reliable protocol based on supervised manifold learning techniques meant to connect vibrational spectra to a variety of complex and diverse atomic structure configurations. As an illustration, we examined a large database of virtual vibrational spectroscopy profiles generated from atomistic simulations for silicon structures subjected to different stress, amorphization, and disordering states. We evaluated representative features in those spectra via various linear and nonlinear dimensionality reduction techniques and used the reduced representation of those features with decision trees to correlate them with structural information unavailable through classical human-identifiable peak analysis. We show that our trained model accurately (over 97% accuracy) and robustly (insensitive to noise) disentangles the contribution from the different material states, hence demonstrating a comprehensive decoding of spectroscopic profiles beyond classical (human-identifiable) peak analysis.

No abstract

No abstract

This paper describes a lightweight neural network (NN) to predict thermodynamic, electric, and electronic properties of hybrid organic–inorganic perovskites (HOIPs) using Hirshfeld surfaces as novel material representation for HOIPs. The neural network utilizes only a few Hirshfeld surface features (e.g., volume, surface area, globularity, and effective radius), along with qualitative and quantitative (mixed) variables, to predict the properties of HOIPs in a highly accelerated manner. Our use of Hirshfeld surface-based descriptors of HOIP crystals leads to a new metric for measuring the effective radius of an organic molecule within a given structure, which are proven to be highly effective features for efficient machine learning of crystalline materials’ properties. A detailed comparison between the crystal graph convolutional neural network (CGCNN) and the Hirshfeld surface-based neural network analysis via UMAP and HDBSCAN clustering is provided to assess the efficacy of these methods for different compound chemistries. It is shown that a combination of lower-order feature representation and a shallow lightweight neural network is capable of predicting material properties for HOIPs. Benchmarking against well-established denser crystal property prediction techniques such as the CGCNN and deeper graph attention layer graph neural network (deeper GATGNN) shows that our approach provides comparable and, in some cases, even superior predictive performance of properties such as formation energy, band gap, and electronic dielectric constant but all at much lower computational cost.