GFFMERGE: Efficient Merging of Graph Neural Force Fields and Beyond

2026-06-02

AI Summary

Existing graph neural network (GNN)-based force field models require costly retraining when transferred to new chemical systems and lack efficient training-free fusion strategies. This work proposes the first closed-form model fusion framework tailored for GNNs, which formulates model combination as a convex embedding alignment problem by exploiting the linear structure of message-passing layers, thereby enabling highly efficient, fine-tuning-free integration. The approach overcomes the limitations of conventional fusion methods from vision and language domains when applied to force field regression tasks and facilitates modular composition of specialized models. Experiments demonstrate that the proposed framework achieves 5–27× acceleration across molecular, solid-state, and large-scale graph datasets, with fused performance approaching that of the joint-training gold standard and significantly outperforming all training-free baselines.

Abstract

Graph Neural Networks (GNNs) have revolutionized Neural Force Fields for atomistic simulations, achieving near-quantum accuracy at reduced cost, yet adapting these models to new chemical systems requires expensive retraining of foundation models. Inspired by model merging in vision and language processing, we introduce GFFMERGE, the first principled framework for closed-form model merging in GNNs. We exploit the linear structure of message-passing layers and formulate merging as a convex embedding-alignment problem with an analytical solution. Through the first systematic benchmarking of model merging for GNNs, we show that existing methods designed for vision and language catastrophically fail on force field regression, while GFFMERGE recovers performance approaching gold standard joint training. Across molecular (MD17, MD22), solid-state (LiPS20), and large-scale graph benchmarks, GFFMERGE and GNNMERGE (its generic GNN counterpart) achieve 5-27$\times$ speedups while enabling modular composition of specialized models. Remarkably, our closed-form solution alone outperforms all baseline methods before fine-tuning and provides superior initialization for faster, data-efficient convergence.

Problem

• Graph Neural Networks
• Neural Force Fields
• Model Merging
• Atomistic Simulations
• Chemical Systems

Innovation

• Graph Neural Networks
• Model Merging
• Neural Force Fields
• Closed-form Solution
• Message Passing