Search papers, labs, and topics across Lattice.
This paper introduces FUSE, a novel approach to Simulation-Based Inference (SBI) that utilizes a dual-track architecture to effectively model multimodal data while maintaining the unique characteristics of different inputs. By implementing a Feynman-Kac steered sampling strategy, FUSE enhances the quality of generated samples during inference, resulting in improved posterior estimates that closely align with ground-truth MCMC benchmarks. The method demonstrates significant advantages in resolving complex parameter degeneracies in a real-world exoplanet orbital estimation task, showcasing its potential for accelerating scientific discoveries across various domains.
FUSE resolves complex parameter degeneracies in exoplanet orbital estimation, outperforming state-of-the-art methods and closely matching ground-truth MCMC results.
Simulation-Based Inference (SBI) is critical for scientific discovery, with generative models offering a promising path toward efficient inference. However, existing methods struggle with effective multimodal modeling. They often rely on brute-force fusion strategies that ignore the structural disparities between parameters and observations, thus limiting estimation fidelity. In this work, we introduce FUSE (Feynman-Kac steered mUlti-modal flow matching for efficient Simulation-based posterior Estimation). Unlike prior work, FUSE employs a dual-track architecture that preserves the distinct features of multimodal inputs while facilitating dynamic interaction. Additionally, we propose an FK-steered sampling strategy that leverages intermediate observation likelihoods to guide the generative trajectories, effectively improving the sample quality during inference. Our approach outperforms state-of-the-art baselines on standard SBI benchmarks, producing posteriors that closely match ground-truth MCMC. Furthermore, in a real-world exoplanet orbital estimation task, FUSE successfully resolves complex parameter degeneracies that challenge existing methods, highlighting its potential to accelerate complex scientific discoveries in astrophysics and beyond.
