Machine Learning Discovery of Computational Model Efficacy Boundaries.

Area of Science:

• Computational Physics
• Materials Science
• Machine Learning

Background:

• Computational models are often built in hierarchies of variable fidelity.
• Quantitative rules for defining fidelity boundaries between models are lacking.
• This limits the reliable selection and application of appropriate computational methods.

Purpose of the Study:

• To establish a quantitative, data-driven method for determining accuracy boundaries in computational models.
• To identify a symbolic rule governing the transition between high-fidelity and low-fidelity models.
• To enhance the understanding of factors influencing model fidelity in atomistic simulations.

Main Methods:

• Construction of a dataset using radial distribution functions from seven high-fidelity atomistic computational methods.
• Inclusion of diverse features such as element, density, and temperature.
• Optimization of a support vector machine (SVM) through iterative feature engineering to discover a symbolic decision boundary.

Main Results:

• Discovery of a symbolic decision boundary that quantitatively defines the accuracy limit between high-fidelity and simple pair-potential models.
• Emergence of an algorithmic-independent symbolic rule governing model fidelity.
• Identification of the central role of atomic physics in determining the accuracy of computational models.

Conclusions:

• The data-driven approach successfully reveals a fundamental rule for computational model fidelity.
• The discovered symbolic rule provides deeper insights into the boundaries of model accuracy.
• This work facilitates more informed choices in selecting computational models for scientific inquiry.