| Abstract Scope |
Inverse design is a novel materials science approach that starts with desired properties and identifies materials that exhibit them, reversing the traditional test-and-characterize process. This method has advanced through AI and machine learning, especially generative models like variational autoencoders (VAEs). VAEs, trained on materials datasets, create a latent space structured by property values and can generate new materials with target properties by sampling this space. When combined with property prediction models, VAEs enable efficient discovery of materials with desired traits.
While generative models have been used for molecules and solid-state materials, they have not yet been applied to molten salts—materials of growing interest for energy applications. Molten salts are challenging to study experimentally due to high temperatures, toxicity, and corrosivity. Inverse design, powered by AI, offers a route to efficiently explore their vast chemical space and reduce research costs. However, the lack of appropriate data representations has slowed progress in this area.
To address this, we developed a workflow for molten salt inverse design. We assembled a dataset of molten salt properties, using density and viscosity data from MSTDB-TP and NIST-Janz, and material descriptors from JARVIS-CFID. A generative VAE model and a property prediction neural network were trained on this dataset. New molten salts were generated and evaluated using molecular dynamics (MD) simulations.
The workflow proved effective: the predictive model estimated properties with reasonable accuracy, the generative model produced a property-ordered latent space, and generated salts matched target properties in MD validation. |