Machine Learning for Accelerated Metal Oxide Thermochemical Predictions.

Area of Science:

• Computational Materials Science
• Machine Learning in Chemistry
• Solid State Physics

Background:

• Accurate prediction of material properties like cohesive energies is crucial for discovering new materials.
• Traditional electronic-structure calculations, such as Density Functional Theory (DFT), are computationally expensive.
• Machine learning (ML) offers a potential pathway to accelerate these predictions.

Purpose of the Study:

• To develop a machine learning framework for predicting normalized clustering energies (NCE) of metal(II) oxides.
• To extrapolate NCE predictions to determine cohesive bulk energies.
• To assess the accuracy and efficiency of different ML models compared to DFT calculations.

Main Methods:

• A dataset of DFT-optimized M(II)O structures (alkaline earth and 3d transition metals) was compiled.
• Structures were encoded using Smooth Overlap of Atomic Positions (SOAP) descriptors.
• Kernel ridge regression (KRR), artificial neural networks (ANN), and tree-based models were trained and evaluated.

Main Results:

• KRR achieved the lowest prediction error for NCE (MAE = 0.619 kcal/mol, R² = 0.994), followed by ANN.
• Predicted bulk cohesive energies for several oxides were within ~1.5 kcal/mol of DFT values.
• Incorporating diverse structures improved model transferability, demonstrating ML's potential for accurate thermochemistry prediction.

Conclusions:

• Machine learning models trained on modest cluster data can accurately predict NCE and bulk thermochemistry.
• This ML approach significantly reduces the computational cost compared to traditional DFT calculations.
• The study highlights the effectiveness of encoding complex electronic structures for ML-driven materials property prediction.