Scalable Inference-Time Annealing with Surrogate Likelihood Estimators.

Area of Science:

• Computational chemistry and biophysics
• Generative modeling
• Molecular dynamics

Background:

• Efficiently sampling the Boltzmann distribution of molecules is a key challenge.
• Generative models offer a simulation-free approach but often require computationally expensive divergence calculations.
• Current methods like diffusion models with importance sampling struggle with scalability.

Purpose of the Study:

• To develop a scalable method for efficient molecular Boltzmann distribution sampling.
• To overcome the computational intractability of existing generative sampling techniques for larger systems.
• To introduce a novel approach that avoids costly score field divergence computations.

Main Methods:

• Introduced Scalable Inference-Time Annealing (SITA), a method that retrains flow-based models.
• Utilized an energy-based model to provide fast surrogate likelihoods.
• Employed iterative retraining along a temperature ladder for progressively lower temperature sampling.

Main Results:

• Achieved state-of-the-art performance on Alanine Dipeptide and Alanine Tripeptide.
• Demonstrated scalability by avoiding computationally expensive divergence terms.
• Successfully generated samples at progressively lower temperatures efficiently.

Conclusions:

• SITA offers a computationally tractable and scalable solution for molecular Boltzmann distribution sampling.
• The method advances generative modeling applications in computational chemistry and biophysics.
• SITA provides an efficient alternative to traditional simulation-based sampling techniques.