LFaB: Low Fidelity as Bias for Active Learning in the Chemical Configuration Space

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Mechanics

Background:

• Active learning aims to optimize training data selection for machine learning models.
• Current methods often focus on minimizing model variance, which can be inefficient and even less effective than random sampling.
• The bias-variance trade-off is a fundamental concept in machine learning model performance.

Purpose of the Study:

• To introduce a new active learning strategy that minimizes model bias instead of variance.
• To demonstrate the effectiveness of this bias-minimization approach in improving sample selection efficiency.
• To reduce the computational cost and data requirements for machine learning models in scientific applications.

Main Methods:

• Developed a bias approximation technique using low-fidelity data, drawing from multifidelity machine learning (Δ-ML).
• Implemented a greedy sample selection procedure focused on minimizing the approximated bias.
• Applied and evaluated the method on quantum chemistry tasks, including excitation energy prediction and ab initio potential energy surface calculations.

Main Results:

• The proposed bias-minimization strategy closely matched the best-case error achievable by any greedy selection method.
• Achieved significant reductions in training data consumption, up to an order of magnitude, compared to standard active learning.
• Demonstrated the method's efficacy across various quantum chemistry applications.

Conclusions:

• Minimizing bias is a more effective strategy for active learning sample selection than minimizing variance, particularly in scientific domains.
• The use of low-fidelity data provides an efficient way to approximate bias for active learning.
• This approach offers substantial improvements in data efficiency for machine learning in quantum chemistry.