Federated Sampling of Molecular Conformers via Compositional Flows
We present a framework for federated sampling of continuous-time generative models via compositional flows (CoFlo) that enables collaborative, and privacy-preserving generation of molecular structures across distributed clients. In our setting, clients independently train local flow matching models on private data; at inference, a lightweight router---trained via federated learning---combines the pretrained client velocity fields as a posterior-weighted mixture. While the framework is general, we develop and evaluate it on molecular conformer generation problems. Our router-based inference approach outperforms the best local clients, as well as uniform velocity averaging across all heterogeneous splits on the global test set while matching local client validation set accuracy.
