Flexible uncertainty calibration for machine-learned interatomic potentials.

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Statistical Modeling

Background:

• Reliable uncertainty quantification (UQ) is crucial for trustworthy machine-learned interatomic potentials (MLIPs) in atomistic simulations.
• Conformal prediction (CP) offers formal coverage guarantees for UQ but often struggles with accuracy, scalability, and adaptability to complex atomic environments.

Purpose of the Study:

• To develop a flexible and efficient uncertainty calibration framework for MLIPs.
• To improve the accuracy and adaptability of predictive intervals in MLIPs.
• To enable more robust and reliable atomistic simulations.

Main Methods:

• Reformulated conformal prediction (CP) as a parameterized optimization problem for direct learning of environment-dependent quantile functions.
• Developed a flexible uncertainty calibration framework compatible with diverse MLIP architectures and baseline UQ schemes.
• Validated the framework using the MACE-MP-0 foundation model across various benchmarks.

Main Results:

• Achieved sharper and more adaptive predictive intervals with negligible computational overhead.
• Demonstrated substantial improvements in uncertainty-error correlation and detection of high-error configurations for active learning.
• Showcased reliable transferability across different exchange-correlation functionals and system types (ionic crystals, catalytic surfaces, molecular systems).

Conclusions:

• The proposed framework offers a practical and data-efficient approach to robust and transferable uncertainty quantification in MLIPs.
• This method enhances the reliability of atomistic simulations by providing accurate and adaptive uncertainty estimates.
• The framework's generality and compatibility pave the way for broader adoption in materials science and related fields.