Stability of training and assessment of CG model quality without extensive simulations

Investigate the stability properties of training procedures for machine-learned coarse-grained molecular dynamics force-fields and develop systematic methods to effectively assess the quality of such coarse-grained force-fields without relying on extensive molecular dynamics simulations.

Background

The authors note that validation losses (force/score matching) do not necessarily correlate with best simulation performance and report instances where model quality fluctuated or deteriorated despite decreasing validation losses, highlighting concerns about training stability.

Given that their evaluation required long CG simulations to assess free energy surfaces relative to atomistic references, the authors explicitly identify the need for reliable, non-simulation-based assessment strategies for CG force-field quality.