Artificial Intelligence for Accelerating Time Integrations in Multiscale Modeling.

Area of Science:

• Computational biology
• Biophysics
• Artificial intelligence in scientific computing

Background:

• Complex biophysical simulations, such as platelet dynamics in blood flow, are computationally demanding.
• Standard time stepping algorithms often involve redundant calculations, leading to inefficiencies.
• Accurate modeling of multiscale spatiotemporal dynamics is crucial for understanding biological processes.

Purpose of the Study:

• To develop a novel Artificial Intelligence-enhanced Adaptive Time Stepping (AI-ATS) algorithm for efficient biophysical simulations.
• To accelerate the solution of computationally intensive problems like platelet dynamics in shear blood flow.
• To adapt timestep sizes dynamically based on underlying biophysical dynamics.

Main Methods:

• Integration of a novel AI framework, including recurrent neural network autoencoders (LSTM, GRU) and fully-connected neural networks, into the simulation process.
• Development of AI algorithms to memorize dynamic states and optimize timestep sizes and step jumps.
• Comparison of the AI-ATS algorithm against the standard time stepping algorithm (STS) for accuracy and computational speed.

Main Results:

• The AI-ATS algorithm successfully adapted timestep sizes to biophysical dynamics, enabling significant speedup.
• A reduction of 40% in unnecessary calculations was achieved compared to STS.
• Errors in mechanical and thermodynamic properties were bounded to within 3%.

Conclusions:

• The AI-ATS algorithm offers a computationally efficient approach for solving complex biophysical problems.
• Dynamic adaptation of timestep sizes by AI can significantly reduce computational cost without compromising simulation accuracy.
• This AI-driven method holds promise for accelerating multiscale simulations in various scientific domains.