Physics-Constrained Neural ODEs for MXene Bandgap Prediction with Conformal Uncertainty

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• MXenes are promising photocatalysts due to tunable properties.
• Accurate prediction of electronic bandgaps is vital for MXene applications.
• Current high-fidelity bandgap calculations are computationally expensive.

Purpose of the Study:

• To develop a computationally efficient and accurate machine learning model for MXene bandgap prediction.
• To integrate advanced machine learning techniques for improved prediction accuracy and reliability.
• To identify novel MXene candidates for photocatalytic water splitting.

Main Methods:

• Developed a physics-constrained neural ordinary differential equation (PC-NODE) model.
• Integrated multi-fidelity learning, split-conformal calibration, and Pareto screening.
• Trained the model on the MXgap database and validated its predictive performance.

Main Results:

• Achieved a mean absolute error of 0.186 eV on MXene bandgap prediction.
• Demonstrated high accuracy in classifying materials as metallic or semiconducting (0.856 accuracy, 0.925 ROC-AUC).
• Identified 74 promising lanthanum-based MXene candidates for water splitting.

Conclusions:

• The PC-NODE framework provides a data-efficient and mathematically grounded alternative for MXene bandgap prediction.
• The developed model accurately predicts bandgaps and identifies suitable candidates for photocatalysis.
• This approach accelerates the discovery of novel MXene materials for sustainable energy applications.