PAL - parallel active learning for machine-learned potentials.

Area of Science:

• Computational science
• Machine learning
• Scientific computing

Background:

• Effective machine learning requires representative datasets.
• Active learning (AL) enhances models by iteratively acquiring data, minimizing costs.
• Current AL methods are often manual and lack parallelism, hindering computational efficiency.

Purpose of the Study:

• Introduce PAL, an automated, modular, and parallel active learning library.
• Enable efficient execution and communication of AL tasks on shared- and distributed-memory systems using MPI.
• Provide flexibility for users to customize AL components like models, oracles, and exploration strategies.

Main Methods:

• Developed PAL, a library integrating AL tasks with MPI for parallel execution.
• Implemented features for customizable machine learning models with uncertainty estimation.
• Designed oracles for ground truth labeling and strategies for target space exploration.

Main Results:

• PAL significantly reduces computational overhead and improves scalability.
• Achieved substantial speed-ups via asynchronous parallelization on CPU and GPU hardware.
• Demonstrated effectiveness across diverse applications: biomolecular systems, molecular dynamics, inorganic clusters, and thermo-fluid dynamics.

Conclusions:

• PAL enables efficient utilization of high-performance computing resources in active learning.
• Accelerates machine learning model development in scientific research and engineering.
• Fosters advancements through parallelized, automated active learning workflows.