"Gold-Standard" Δ-Machine Learned Transferable Potential for Linear Alkanes.

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Physical Chemistry

Background:

• Conformational properties of linear alkanes (C_nH_2n+2) are of significant scientific interest.
• Previous studies focused on the minimum chain length for transitions between linear and hairpin conformations.
• Many-body permutationally invariant polynomial (MB-PIP) potentials have been developed for alkane simulations.

Purpose of the Study:

• To develop highly accurate MB-PIP potentials for linear alkanes.
• To improve upon existing B3LYP-based potentials using a delta-machine learning (Δ-ML) approach.
• To validate new potentials against high-level quantum chemical calculations.

Main Methods:

• Utilized a Δ-ML approach to enhance a B3LYP-based MB-PIP potential.
• Developed a new PBE0+MBD MB-PIP potential trained on Pair Natural Orbital Local Coupled Cluster (PNO-LCCSD(T)-F12) energies.
• Calculated and compared potential energy minima and vibrational power spectra for various alkane chain lengths.

Main Results:

• The new Δ-corrected potentials accurately predict the energy differences between linear and hairpin minima for alkanes C₁₂H₂₈ to C₂₈H₅₈.
• The developed potentials show improved accuracy compared to benchmark PNO-LCCSD(T)-F12 energies.
• Vibrational power spectra for C₁₄H₃₀ were computed using both original and Δ-ML corrected potentials.

Conclusions:

• The developed MB-PIP potentials represent the most accurate models for linear alkanes to date.
• These potentials can be reliably used for simulating and studying the properties of linear alkanes.
• The Δ-ML approach offers a promising strategy for enhancing the accuracy of molecular potentials.