AI RESEARCH
Hypothesis-driven construction of mesoscopic dynamics
arXiv CS.LG
•
ArXi:2605.16211v1 Announce Type: new Traditional scientific modeling typically begins with fixed, instance-wise effective equations and then carries out equation-specific analysis and computation, a procedure that becomes exceptionally challenging in complex applications such as multiscale systems. We propose an alternative paradigm by learning mesoscopic dynamics within a mathematically constrained hypothesis class. Building upon a generalized Onsager principle, we