High-throughput material and drug discovery based on density functional theory (DFT) has achieved tremendous success in recent decades. However, the power of DFT on real-life molecules like organic semiconductors (OSC), metal–organic frameworks (MOF), and ribonucleic acids (RNA) can suffer significantly from its high computational cost. In this talk, we will discuss two of our efforts of removing the computational bottleneck of DFT with minimal accuracy loss using state-of-the-art machine learning (ML) algorithms. In the first part, we will introduce a new exchange–correlation (XC) functional developed in our group, ML-ωPBE, which evaluates the molecule-specific range-separation parameter (ω) in a range-separated hybrid (RSH) functional using the stacked ensemble algorithm and composite molecular descriptors. Compared to first-principles OT-ωPBE, a well-trained ML-ωPBE reaches a mean absolute error (MAE) of 0.00504 a₀^–1 for optimal ω’s, reduces the computational cost by 2.66 orders of magnitude, and achieves comparable predictive power in optical properties to OT-ωPBE. In the second part, we will show a new version of the many-body expansion (MBE) theory that integrates advanced double-layer graph convolutional networks (GCNs). This GCN-MBE theory divides a large system into multiple fragments, evaluates one-fragment properties using DFT and many-fragment interactions using GCN-learned structure–property relationships. Our preliminary studies on water clusters, phenol clusters, and water–phenol mixtures demonstrate that GCN-MBE can reproduce DFT-evaluated two-fragment and three-fragment energies within 0.1 kcal/mol with simple geometric and electronic descriptors, while reducing the computational cost by more than 99.9%. These two projects set the stage for developing versatile, robust, physics-based, and data-driven computational models for high-throughput material and drug discovery.

Yili Shen

Appearances