Putting Density Functional Theory to the Test in Machine-Learning-Accelerated Materials Discovery.

Area of Science:

• Computational Materials Science
• Machine Learning Applications
• Electronic Structure Theory

Background:

• Machine learning (ML) accelerates computational materials design, but current workflows face limitations.
• Density functional theory (DFT) derived training data introduces biases, and many calculations fail for complex materials.
• Materials with strained bonds, radicals, or metal-organic bonds pose significant challenges for traditional electronic structure methods.

Purpose of the Study:

• To outline necessary advances in accuracy, efficiency, and methodology for ML-accelerated materials discovery beyond conventional DFT-based workflows.
• To address the challenges posed by complex materials and improve the reliability of computational predictions.
• To introduce 'decision engines' for autonomous workflows in materials discovery.

Main Methods:

• Developing ML models trained to predict outcomes of multiple electronic structure methods or their differences.
• Implementing ML models for quantitative sensitivity analysis of DFT calculations.
• Creating ML models to predict the likelihood of calculation success and detect strong correlation effects.

Main Results:

• ML models can predict the success rate of DFT calculations, enabling better screening of potential materials.
• Quantitative sensitivity analysis is achievable through ML models predicting method differences.
• ML-based 'decision engines' can diagnose potential issues and suggest adaptation strategies for calculations.

Conclusions:

• Advanced ML models are crucial for overcoming limitations in current DFT-based materials discovery.
• Predictive ML models enhance the robustness and efficiency of computational materials design.
• The development of 'decision engines' paves the way for autonomous and reliable materials discovery workflows.