Modeling the behavior of concentrated aqueous HNO3 using machine learning interatomic potentials

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Investigating the behavior of nitric acid (HNO3) across different concentrations is crucial for understanding its chemical properties.
• Existing computational models often struggle to accurately capture the complex structural and thermodynamic properties of nitric acid, especially concerning its dissociation.
• The development of accurate and efficient interatomic potentials is essential for molecular dynamics simulations.

Purpose of the Study:

• To develop and validate multi-defect machine learning interatomic potentials (MLIPs) for nitric acid using the DeepMD-kit.
• To investigate the structural and thermodynamic properties of nitric acid over a wide range of concentrations using molecular dynamics (MD) simulations.
• To compare the performance of custom DeepMD MLIPs against foundational models like MACE-MP0 and MACE-OFF23.

Main Methods:

• Training two multi-defect MLIPs using BLYP-D2 and PBE-D3 density functional theories with DeepMD-kit.
• Performing MD simulations to investigate structural and thermodynamic properties, including the degree of dissociation (α) and pKa.
• Benchmarking custom DeepMD MLIPs against MACE-MP0 and MACE-OFF23, and comparing with classical force fields (FFs).

Main Results:

• The developed MLIPs accurately predict the degree of dissociation (α) and pKa of nitric acid, showing good agreement with experimental data.
• HNO3 exhibits weaker acid behavior at higher concentrations, with a standard-state pKa matching experimental values.
• Custom DeepMD MLIPs provide more compact solvation shells, reproduce density-concentration trends, and are significantly faster than MACE-MP0, outperforming foundational models in capturing subtle structural features.

Conclusions:

• Bespoke, reactive MLIPs are necessary for accurately simulating chemical reactivity and properties like dissociation, going beyond universal MLIPs.
• The developed DeepMD MLIPs offer a significant advancement for studying concentrated nitric acid systems, providing accurate predictions and improved computational efficiency.
• While classical FFs are efficient for density, they lack the chemical reactivity crucial for predicting dissociation and acidity, highlighting the value of reactive MLIPs.