Accelerated Bayesian Inference for Molecular Simulations using Local Gaussian Process Surrogate Models.

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Statistical Modeling

Background:

• Bayesian inference is crucial for uncertainty quantification but computationally intensive.
• Modeling complex physical chemistry data, like spectra and scattering patterns, exacerbates computational challenges.
• Existing methods struggle with high-dimensional datasets.

Purpose of the Study:

• To introduce local Gaussian process (LGP) surrogate models for accelerating Bayesian inference.
• To overcome computational barriers in physical chemistry simulations.
• To enable efficient uncertainty quantification for complex thermophysical properties.

Main Methods:

• Employed local Gaussian process (LGP) surrogate models.
• Developed a method with linear time-complexity concerning independent variables.
• Trained an LGP on the radial distribution function of liquid neon.

Main Results:

• Achieved a 1,760,000-fold speed-up compared to molecular dynamics simulations.
• Outperformed conventional Gaussian processes by three orders of magnitude.
• Demonstrated linear scaling of LGP time-complexity with independent variables.

Conclusions:

• LGPs are robust and efficient surrogate models for Bayesian inference.
• LGPs significantly reduce computational cost in molecular simulations.
• This approach expands Bayesian inference applications to diverse experimental data in physical chemistry.