Performance of uncertainty-based active learning for efficient approximation of black-box functions in materials...

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Approximating black-box functions is crucial for evaluating new materials.
• Active learning (AL) aims to enhance function approximation using minimal training data.
• Uncertainty sampling is a key AL strategy.

Purpose of the Study:

• To assess the efficiency of uncertainty-based active learning for approximating black-box functions in materials science regression.
• To compare AL performance against random sampling across diverse material datasets and dimensionalities.

Main Methods:

• Investigated uncertainty-based active learning (AL) in regression tasks.
• Utilized various material databases, including ternary systems, inorganic materials, small molecules, and polymers.
• Evaluated performance based on input data distribution (uniform vs. discrete/unbalanced) and feature space dimensionality.

Main Results:

• Uncertainty-based AL outperformed random sampling in low-dimensional, uniform input spaces (e.g., liquidus surfaces).
• AL's efficiency diminished in high-dimensional, unbalanced feature spaces typical of materials databases.
• Performance correlated with material descriptor dimensionality; lower dimensions favored AL.

Conclusions:

• Uncertainty-based active learning shows promise for materials science but is not universally efficient.
• The effectiveness of AL is highly dependent on data characteristics like dimensionality and distribution.
• Further research is needed to optimize AL strategies for complex materials data.