Calibrating DFT Formation Enthalpy Calculations by Multifidelity Machine Learning.

Area of Science:

• Computational Materials Science
• Machine Learning in Materials Discovery
• Predictive Modeling of Material Properties

Background:

• Predicting material properties experimentally is crucial but hindered by limited data.
• Existing computational methods like density functional theory (DFT) have accuracy limitations.
• Machine learning offers potential but struggles with sparse experimental datasets.

Purpose of the Study:

• To develop a machine learning model for predicting experimental formation enthalpy of materials.
• To improve upon the accuracy of established DFT functionals and deep learning approaches.
• To calibrate DFT predictions and discover thermodynamically stable materials.

Main Methods:

• Utilized a multifidelity random forest model to learn from limited experimental data.
• Compared model performance against Perdew-Burke-Ernzerhof (PBE), PBEsol, and meta-generalized gradient approximation (meta-GGA) functionals (SCAN, r²SCAN).
• Employed the model for data mining to analyze deviations between DFT and experimental results.

Main Results:

• The multifidelity random forest model achieved higher prediction accuracy for formation enthalpy than PBE, PBEsol, and meta-GGA functionals.
• The model surpassed the performance of deep neural network-based representation learning and transfer learning methods.
• Model application led to the calibration of DFT formation enthalpy in the Materials Project database and the identification of stable materials.

Conclusions:

• Multifidelity random forest models are effective for predicting material properties with limited experimental data.
• This approach offers superior accuracy compared to traditional DFT functionals and current deep learning techniques.
• The model serves as a valuable tool for materials discovery and understanding DFT-experimental discrepancies.