Physics-Inspired Single-Particle Tracking Accelerated with Parallelism.

Area of Science:

• Computational Science and Engineering
• Statistical Physics and Biophysics

Background:

• Data modeling tools balance accuracy, computational cost, and flexibility.
• Physics-inspired likelihood-based methods offer high accuracy but are computationally expensive and difficult to scale.
• Existing parallelization strategies for these methods are often inefficient for algorithms requiring frequent inter-process communication.

Purpose of the Study:

• To develop a novel, scalable parallel computing framework for likelihood-based inference.
• To overcome the computational limitations of traditional single-threaded approaches in data modeling.
• To enable efficient high-performance inference on modern parallel architectures.

Main Methods:

• Developed a new strategy exploiting inherent parallelism in likelihood evaluation and posterior sampling.
• Operates on a single shared dataset with frequent, lightweight inter-thread and inter-processor communication.
• Framework designed for compatibility with modern parallel architectures like GPUs.

Main Results:

• Achieved up to a 50-fold speedup on a single GPU compared to single-threaded CPU implementations.
• Demonstrated scalability for large-scale problems in physics-inspired data modeling.
• Successfully applied to diffraction-limited single-particle fluorescence tracking.

Conclusions:

• The proposed framework offers a scalable and efficient solution for high-performance likelihood-based inference.
• This approach significantly reduces computational cost, making complex models more accessible.
• Enables faster and more efficient analysis of large datasets in scientific research.