A critical benchmark of universal machine learning force fields and ReaxFF for iron-based oxygen carriers in chemical looping combustion: Defining applicability boundaries from static properties to dynamic CH 4 conversion

Abstract

Machine learning force fields (MLFFs) offer an attractive balance of accuracy and efficiency for energy conversion processes. However, applications to chemical looping combustion (CLC) are hindered by transferability and reliability of MLFFs to high-temperature, multiphase oxidation reactions over complex oxide interfaces of oxygen carriers (OCs). Herein, we establish a critical benchmark framework comparing typical universal MLFFs (DPA3, M3GNet, MACE-omat, NEP89, SevenNet-0) and reactive force field (ReaxFF) against density functional theory (DFT) calculations for iron-based OCs in CH4 conversion during CLC. The evaluation covered lattice parameters, local coordination, O2/CH4 adsorption over OCs, CH4 conversion, and CO2/CO/H2O product evolution. Quantitatively, NEP89 reproduced DFT lattice parameters with an error of 0.24% and the bulk modulus with an error of 3.1%. MACE-omat accurately described O2 chemisorption and weak CH4 physisorption, with an O2 adsorption-energy deviation of only 0.056 eV at the DFT equilibrium distance and an identical CH4 adsorption energy to DFT, both −0.091 eV. Universal MLFFs predicted markedly different CH4 consumption and product distributions, whereas ReaxFF was more consistent with known CLC mechanisms. This study defines the applicability boundaries of universal MLFFs for iron-based OCs and provides a methodological basis for developing CLC reaction-specific force fields.