GFFMERGE: Efficient Merging of Graph Neural Force Fields and Beyond

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.

Simulating the behaviour of molecules and materials at the atomic level is crucial for scientific discovery, but traditional methods are slow and computationally expensive. Machine learning has significantly accelerated this process using powerful Graph Neural Network-based force field models, which also reduce the required compute. However, adapting these models to new chemicals remains costly, as they typically must be retrained from scratch. To address this, we developed GFFMERGE, a method that combines multiple existing force field models into a single unified model — without costly retraining. The merged model performs on par with the individual models it is built from. While similar model merging techniques work well for language and vision tasks, they fail entirely for the energy and force prediction tasks required in molecular simulation. GFFMERGE combines models up to 27 times faster than traditional training, while maintaining the same accuracy. This allows researchers to cheaply assemble specialised AI models, dramatically accelerating the discovery of new molecules and materials.