Supervised Variational Autoencoders for the Inverse Design of Molten Salt Mixtures

In materials science, the approach of inverse design refers to a new paradigm in which one begins with desired material properties as an input and obtains a material that possesses these properties as an output. This contrasts with the traditional approach taken in materials science that begins with a material and performs experiments and computer simulations to explore its material properties. Several methods have been proposed to perform inverse design in materials, but recent years have seen several important advances through the application of artificial intelligence (AI)/machine learning (ML) methods, more specifically the generative models.<br/><br/>Generative models, such as variational autoencoders (VAEs), are trained on datasets of materials represented in an invertible way: the way the materials are represented in the algorithm allows for their output to be turned back into a new material. After training, these models are used to generate new data points, which themselves can be turned into new materials, similar to the ones present in the dataset. These models are generally coupled with property prediction models to generate a continuous lower-dimensional space, often called latent space, that the algorithms order according to the values of the predicted property. This approach allows for the material generation to be performed according to target values for the property.<br/><br/>There are several examples of different generative modeling methods applied towards the design of molecular and solid inorganic materials, but so far there have been no examples of these methods being applied towards the inverse design of molten salts. This is despite numerous applications of molten salts in energy production and energy storage, where such algorithms are expect to be effective to explore the enormous chemical design space of molten salts, and thereby reduce the computational expense of performing simulations such as molecular dynamics. The many properties that make experimentation on molten salts difficult, like their toxicity, high corrosivity, and the need to experiment on them at high temperatures, makes this approach particularly useful. The current lack of such a workflow might be due to the difficulty of developing suitable, invertible methods for the representation of molten salts in generative models.<br/><br/>For this reason, we have developed a workflow for the inverse design of molten salts through the compilation of an molten salt properties dataset using an invertible representation and associated information on macroscopic density and viscosity. We gathered the data from the Molten Salts Thermal Properties Database – Thermophysical (MSTDB-TP) and the NIST Properties of Molten Salts Database (NIST-Janz), with additional information given in the form of material property descriptors obtained from the Joint Automated Repository for Various Integrated Simulations – Classical Force-field Inspired Descriptors (JARVIS-CFID). We then implemented a generative model based on a VAE that is coupled with a property prediction neural network. The coupled generative model was trained on our compiled property dataset and used to generate new molten salts according to target values of material properties. We employed ab initio molecular dynamics (AIMD) simulations to compare against the predicted properties from our generative model.<br/><br/>Our results indicate that the inverse design workflow is successful, with the coupled predictive neural network predicting properties with reasonable accuracy. We find that the coupled model is able to generate a continuous lower-dimensional vector space with the values of the target properties ordered across the principal components, where the molten salts that were generated indeed possess physical property values that are in agreement with the results of AIMD simulations.