Title: GFFMERGE: A Framework for Efficiently Merging Graph Neural Force Fields and Beyond

Graph Neural Networks (GNNs) have transformed the landscape of atomistic simulations by delivering Neural Force Fields that rival quantum mechanical accuracy while significantly lowering computational expenses. However, adapting these foundation models to novel chemical environments typically demands costly retraining. Drawing inspiration from model merging techniques established in vision and natural language processing, we present GFFMERGE, a novel, principled framework that enables closed-form model merging within GNNs. By leveraging the linear architecture of message-passing layers, we define the merging process as a convex embedding-alignment problem, which yields an analytical solution.

Our study includes the first comprehensive benchmarking of model merging for GNNs. The results demonstrate that conventional methods, originally designed for vision and language tasks, perform poorly in force field regression contexts. In contrast, GFFMERGE restores performance levels that closely match those achieved through gold-standard joint training. Evaluated across molecular datasets (MD17, MD22), solid-state systems (LiPS20), and large-scale graph benchmarks, both GFFMERGE and its generic GNN equivalent, GNNMERGE, deliver speed improvements ranging from 5 to 27 times while facilitating the modular combination of specialized models. Notably, the closed-form solution alone surpasses all baseline methods prior to any fine-tuning and serves as a superior initialization point, enabling faster and more data-efficient convergence.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC