Parameterization Strategies for Intermolecular Potentials for Predicting Trajectory-Based Collision Parameters.

Area of Science:

• Computational Chemistry
• Chemical Physics
• Molecular Dynamics

Background:

• Accurate potential energy surfaces (PES) are crucial for simulating collisional energy transfer and reaction rates.
• Constructing full-dimensional PES for polyatomic systems remains computationally challenging.

Purpose of the Study:

• To systematically evaluate separable strategies for constructing PES for alcohol-bath gas interactions.
• To quantify the fitting efficiency of different functional forms and sampling strategies.

Main Methods:

• Investigated pairwise (exp6) and nonpairwise (permutationally invariant polynomials, PIPs) functional forms.
• Compared four sampling strategies, including biased Sobol quasirandom and unbiased pseudorandom sampling.
• Assessed fitting efficiency based on the number of ab initio data points required for desired accuracy.

Main Results:

• Fitting efficiency depends primarily on the number of adjustable parameters and sampling strategy, not functional form.
• Biased Sobol sampling is ~7x more efficient than pseudorandom sampling.
• The exp6 form is accurate for Ar but not N2 or H2O; PIPs offer higher accuracy but can overfit.

Conclusions:

• Validated a robust, automated strategy for generating accurate A + M PES with minimal human intervention.
• Highlighted the importance of biased sampling and the potential for overfitting with complex functional forms like PIPs.
• Identified areas for improvement, including pruning PIP expansions and coupling sampling with functional form descriptions.