A Joint Semi-Supervised Variational Autoencoder and Transfer Learning Model for Designing Molecular Transition Metal Complexes

Deep generative models (DGMs) have shown great promise in the generation of organic molecules and inorganic materials with chemical sensible structures and optimized properties. However, there is a lack of their applications in transition metal (TM) complexes due to their flexible coordination environment, multiple accessible oxidation and spin states, despite the importance of these complexes in fine chemical synthesis, commodity production, and optical applications. Herein, we propose a joint semi-supervised junction-tree variational autoencoder (SSVAE) and artificial neural network (ANN) classifier model, coined as LiveTransForM (Ligand variational auto-encoder and Transfer learning For transition Metal complexes), for the design of octahedral TM complexes. LiveTransForM allows the design of ligands that build up TM complexes and the prediction of the spin states of the assembled complexes. We show that the accuracy of the classifier is improved when the latent variables from the SSVAE are used as input for the ANN model compared to those from the unsupervised VAE. Input augmentation using the three molecular axes also improves the accuracy of the classifier. 58 complexes with predicted spin states are then generated by LiveTransForM and the accuracy of their spin state labels are validated by density functional theory methods. Two design strategies, single mutation and seeded generation, are also introduced to allow the directed evolution of a parent complex towards a desirable spin state and local modification of seed complexes with similar spin states, respectively.