Thermodynamic Consistency in Coarse-Grained Molecular Dynamics Simulations

Abstract

Coarse-grained (CG) molecular dynamics models trade atomistic detail for computational reach, but the
structure-based mapping procedures typically used to build them reproduce target radial distribution functions while
leaving thermodynamic properties, pressure, compressibility, excess entropy, unconstrained and often substantially
in error. We systematically quantified this thermodynamic inconsistency for 16 CG models of small-molecule liquids
and liquid mixtures, built by iterative Boltzmann inversion (IBI) from atomistic reference simulations, and tested
three consistency-correction strategies: pressure-corrected IBI, a multistate iterative approach fitting simultaneously
across three thermodynamic state points, and an explicit excess-entropy-matching correction. Uncorrected
structure-only CG models reproduced target radial distribution functions with a mean structural error (integrated
RDF deviation) of 0.031, but showed a mean pressure deviation from the atomistic reference of 340 ± 95 bar and a
mean excess entropy deviation of 18.4%, confirming that structural fidelity alone does not guarantee
thermodynamic fidelity. Pressure-corrected IBI reduced mean pressure deviation to 42 ± 21 bar (an 88% reduction)
at the cost of a 2.1-fold increase in structural error relative to the uncorrected models, while the multistate approach
achieved a more balanced outcome, reducing pressure deviation to 68 ± 24 bar and excess entropy deviation to
6.7%, with only a 1.3-fold increase in structural error. The explicit excess-entropy-matching correction achieved the
lowest excess entropy deviation (3.9%) but transferred least well to state points outside the fitting range, with
transferability tests at temperatures 20 K above the reference condition showing a 2.6-fold larger increase in
pressure error than the multistate models. Across all 16 systems, pressure deviation and excess entropy deviation
of the uncorrected models were strongly correlated (Pearson r = 0.79, p < 0.001), indicating that pressure
inconsistency is a useful, easily computed proxy for the harder-to-estimate excess entropy error. These results
provide a quantitative basis for selecting among available consistency-correction strategies according to whether
accurate structure, accurate thermodynamics, or transferability across state points is the priority for a given CG
modelling application.