Learning the Electron Density Fields Predicted by Density Functional Theory Using Physics-Inspired Machine Learning Approaches

Density functional theory (DFT) is one of the common simulation methods used in understanding the feasibility of the potential atomic structures for various chemical compounds. Unfortunately, DFT methods suffer from poor scaling of computational cost with simulation size, limiting both the types of physical phenomena and number of systems that can be studied. To overcome these limitations, machine learning (ML) has emerged as a promising solution. Here, I propose a novel machine learning approach to connect the atomic structure to its DFT predicted electron density field. The ML approach will employ a combination of Convolutional Neural Networks (CNNs) as well as the recently developed Materials Knowledge System (MKS) framework at Georgia Tech (GT). The advantage of learning electron density is twofold – it can be used to directly derive other DFT outputs such as energy and atomic forces, and also to dramatically reduce computation time by serving as an excellent starting point for DFT calculations on completely new atomic structures. There has been limited research on using ML in this context. The advantage of the proposed approach is that it will aim to combine the best attributes of CNN and MKS methods. Typical CNN models suffer from a vast parameter space and ‘black box’ nature that makes model optimization a difficult time-consuming process and restricts our ability to extract physical intuition from the results. The MKS framework overcomes these challenges by casting the governing physics in the forms of Green’s function derived convolution kernels, which can guide the development, optimization and interpretation of CNN models. To demonstrate the significance of the proposed ML model, I identify two potential applications relevant to Technology Area 12.1 “Materials”: the prediction of DFT interatomic forces for efficient and accurate molecular dynamics simulations, and the rapid discovery and simulation of novel multifunctional materials with tailored properties.

Density functional theory (DFT) is one of the common simulation methods used in understanding the feasibility of the potential atomic structures for various chemical compounds. Unfortunately, DFT methods suffer from poor scaling of computational cost with simulation size, limiting both the types of physical phenomena and number of systems that can be studied. To overcome these limitations, machine learning (ML) has emerged as a promising solution

This is early/mid-stage (TRL 3) — the most realistic path in is NASA SBIR/STTR, which funds small businesses and research institutions to develop technology aligned with NASA's needs (equity-free, phased funding). Check whether a current SBIR/STTR solicitation topic overlaps with this project's technology area, or contact the project directly (above) to ask.

None of these are guaranteed paths for this specific project — TechPort itself doesn't have an "apply" button. Reaching out to the contact(s) above with a specific question is usually the fastest way to find out what's actually open.