Uncertainty-Quantified Hybrid Machine Learning/Density Functional Theory High Throughput Screening Method for...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• High-throughput screening (HTS) using density functional theory (DFT) accelerates materials discovery but is computationally expensive due to DFT's cubic scaling.
• Existing machine learning models like CGCNN can predict material properties but lack uncertainty quantification, potentially leading to errors without structural relaxation.

Purpose of the Study:

• To develop a computationally efficient machine learning framework for materials discovery.
• To address the bottleneck of high computational cost in HTS by integrating ML with DFT.
• To enable uncertainty quantification in ML predictions for more reliable material screening.

Main Methods:

• Modified Crystal Graph Convolutional Neural Network (CGCNN) with hyperbolic tangent activation and dropout (CGCNN-HD) for formation energy prediction and uncertainty quantification.
• Hybrid ML/DFT-HTS approach: initial screening using CGCNN-HD followed by DFT validation of selected candidates.
• Benchmarking against all-DFT-HTS for discovering photoanode materials in Mg-Mn-O ternary compounds.

Main Results:

• The ML/DFT-HTS approach reduced DFT calculations by over 50 times compared to all-DFT-HTS for discovering the Mg2MnO4 photoanode material.
• CGCNN-HD with uncertainty quantification increased the discoverability of promising materials from 30% (CGCNN) to 68% compared to all-DFT-HTS.
• The proposed framework demonstrated a faster and more efficient exploration of vast chemical spaces.

Conclusions:

• The developed ML/DFT-HTS framework with uncertainty quantification offers a computationally efficient alternative to traditional DFT-HTS.
• This approach significantly accelerates the discovery of novel materials with desired properties.
• The integration of uncertainty measures enhances the reliability and discoverability of the screening process.