Systematic iterative refinement of machine-learned universal force fields trained on a diverse database of finite-temperature materials

AI Technology: Machine Learning. 

Density functional theory (DFT) is the primary workhorse in computational materials science.

Implementations of DFT, such as the Vienna ab-initio Simulation Package (VASP), have facilitated the growth of large-scale databases of materials properties, finding widespread application in several subfields of physics, chemistry, and engineering. 

Recently, these databases have been used to develop universal machine-learned force-fields (UFF). These UFFs are trained to predict a very large phase-space of multi-component inorganic materials at the accuracy of DFT, but at a fraction of the computational cost. Reliably training these UFFs still poses a major challenge. 

The datasets used to train UFFs suffer from systematic redundancies in phase space, leading to bulky models that are challenging to fine-tune. In this proposal, we leverage recent developments in VASP to generate a high-quality dataset consisting of finite temperature material properties. 

The project team trains a diffusion-model to generate novel structures, apply a single-shot model to generate finite-temperature structures, compute their ground state properties and use this dataset of DFT calculations to fine-tune UFFs. 

The expected outcomes of our proposal are an accurate UFF, a high-quality finite-temperature materials database, and a diffusion model that predicts novel stable structures.