(169cj) Molecular Design of High Performance Electrolytes with Generative Algorithm

In recent years, the field of computational chemistry has witnessed a surge in novel approaches aimed at designing novel materials for energy storage applications. One such technique involves harnessing the power of Genetic Algorithms (GA) to optimize molecular structure for target properties by evolving and selecting candidate molecules. The goal of this study is to design solvents for FIB electrolytes that concurrently exhibit robust chemical stability against fluoride-ion attacks and facilitate efficient ion transport, making them promising candidates for use in fluoride ion batteries using a machine learning augmented genetic algorithm.

The molecular optimization process is guided by carefully crafted fitness functions that encapsulate the key criteria for stability against fluoride ion attacks as well as the ability of the solvent to solvate fluoride ions. This loss function incorporates two critical components: a stability score and a solvation score. The stability score (optimized using GA to be higher) is derived from the application of the Bell–Evans–Polanyi (BEP) principle, which allows for the evaluation of activation barrier trends by computing the heat of a potential reaction using the GFN2-xTB method. Enumeration of all conceivable products arising from the reaction of a molecule with fluoride ion is executed using its chemical graph, from which the lowest activation barrier is determined based on this stability score. This principle serves as a robust metric in predicting the likelihood of a molecular structure remaining stable under the influence of fluoride ions.

Concurrently, the solvation score plays a crucial role in the molecular optimization process. Departing from conventional methodologies that rely on computationally expensive simulations, the solvation score (optimized using GA to be lower) in this approach leverages machine learning techniques to predict solvation-free energy efficiently. We train a random forest (RF) model to predict the solvation-free energy of a designed solvent molecule for fluoride ions. Inexpensive molecular features such as Morgan Fingerprint and Solvent Accessible Surface Area (SASA) were calculated which serves as input featurization to this RF model. RF models were trained and tested on an 80:20 split of a dataset consisting of 643 data points in the model. This machine learning model streamlines the computational process but also enables the rapid assessment of the solvation effects of the designed molecules.