The impact of low-cost molecular geometry optimization in property prediction via graph neural network.

Machine-learning (ML) algorithms have demonstrated the potential to tackle several material science challenges, ranging from predicting quantum molecular properties to screening novel molecules with tailored properties via inverse design. In this realm, molecular descriptors and graph neural networks (GNNs) based on molecular geometry have yielded promising results for a variety of ML tasks. Nevertheless, the majority of these studies trained and validated their model with geometries optimized through density functional theory (DFT), implying that geometries at the same level of theory will be available for unseen molecular data. Unfortunately, generating these 3D geometries is computationally expensive, limiting their application to explore a myriad of molecular candidates. In contrast, universal function approximators such as neural nets (NNs) can learn any desired function, meaning that GNNs can map non-equilibrium geometries to predict ground-state properties. In this sense, this work investigates the impact of a computationally fast (but less accurate) method to predict molecular properties calculated at the DFT B3LYP/6-31G(2df, p) level. Precisely, we assess the predictive performance of the enn-s2s model for twelve molecular properties from the QM9 dataset with geometries optimized via the Merck Molecular Force Field (MMFF94) and DFT B3LYP/6-31G(2df, p) framework. As a result, 9 out of 12 properties demonstrated a low error gap between feeding the enn-s2s with the MMFF94- and DFT-optimized geometry, thus confirming NNs as a feasible strategy to overcome the before-mentioned limitation for practical application.