Redox potential and solubility of organic redox-active molecules are important parameters for electrochemical devices such as aqueous and non-aqueous redox flow batteries (RFBs). The redox potential is a property of the organic molecules and is by the solvent used, while the solubility of the active species determines the volumetric capacity. When the solubility and redox potential values are multiplied, they provide insight into the energy density of the RFB, making them important properties in the feasibility and efficiency of RFBs. Other factors, such as supporting electrolytes, solvents, temperatures, and pH levels, can also affect the redox potential and solubility, making this a multi-dimensional problem. As a result, the simulation or prediction of performance characteristics for newly synthesized organic redox-active molecules is not straightforward, requiring each possible configuration of parameters, i.e., solvents,electrolytes, and operating conditions, to be experimentally determined.Current methods to simulate redox potential and solubility require advanced modeling techniques such as COSMO-RS, density functional theory, or molecular mechanics. Recently, machine learning (ML) models are being used to capture the intricate relationship between all parameters in complex systems such as problems that involve molecular search and property prediction. Graph neural networks (GNNs) and variational auto-encoders (VAEs) are powerful ML tools for rapidly and accurately predicting the characteristics of chemical systems. One limitation of typical GNNs is that the input can only be a single graph. We are investigating this shortcoming by utilizing multiple GNNs as part of VAE models that encode the graph inputs, which represent molecular structures, to latent space vectors. The different latent space vectors are then concatenated into a single vector, which overcomes the limitation of only permitting a single graph per input. Additionally, the different latent space vectors can be scale to account for concentrations. Our method allows the different encoder models for the redox-active material, support electrolytes, and solubility to be individually optimized while simultaneously providing a more precise optimization.We will show how our approach allows GNNs to solve problems that consist of multiple molecular structures as inputs, such as redox potential and solubility. We then use this to predict the redox potential and solubility in various electrolyte systems to demonstrate the effectiveness of our approach. Additionally, we explore and analyze different classes of GNNs, such as graph convolutional network (GCN), crystal graph convolution neural network (CGCNN), and message passing neural network (MPNN), as the graph encoding model. Ultimately, we show that our approach provides fast and efficient predictions of complex systems and is easily transferable to other problems with multiple chemical components.
Read full abstract
Nov 25, 2024
Swapnil Arawade + 1
Open Access
