Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

GFFMERGE: Efficient Merging of Graph Neural Force Fields and Beyond

About

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.

Parth Verma, Parv P. Singh, Vipul Garg, Ishita Thakre, N. M. Anoop Krishnan, Sayan Ranu• 2026

Related benchmarks

TaskDatasetResultRank
Node ClassificationCora
Accuracy93.35
609
Node ClassificationPubmed
Accuracy79.13
501
Node ClassificationAmazon Photo
Accuracy94.81
327
Node ClassificationarXiv
Accuracy77.47
325
Node ClassificationREDDIT
Accuracy96.8
268
Node ClassificationCiteseer
Mean Accuracy82.6
238
Link PredictionPubMed (test)
AUC96.37
120
Link PredictionCora (test)
AUC0.9445
117
Node ClassificationwikiCS
Accuracy (WikiCS)86.79
101
Node ClassificationAmzComp
Accuracy94.11
39
Showing 10 of 114 rows
...

Other info

Follow for update