MF-Net: New AI model decodes hidden dynamics in complex systems

A new paper from researcher Xingji Cui introduces Mechanical Field Networks (MF-Net), a recurrent dynamical model designed to uncover hidden interaction structures in multivariate systems. Unlike existing approaches that either impose fixed dynamics or leave relations implicit, MF-Net represents all variables in a shared field state and updates that state through a learned "mechanical" transition rule. Each variable carries a field component, and these components evolve jointly based on learned relations that shape state-dependent flows, field responses, and motion tendencies. The key innovation is that the relation structure is part of the rollout process itself: the same learned quantities drive both forecasting and structural readout, making the model inherently interpretable without sacrificing predictive power.

On the challenging 40-dimensional Lorenz-96 chaotic testbed, MF-Net achieved an eight-step R² of 0.798 ± 0.018. More impressively, its learned relation matrix recovered the local coupling structure with near-perfect accuracy: a local/nonlocal strength ratio of 19.80 ± 1.00 and Precision@K of 1.000 ± 0.000. The model also performed competitively on known-law interaction systems, real neural recordings, and ecological time series, demonstrating both short- and medium-horizon forecasting ability while retaining inspectable structural readout. This work provides a practical framework for researchers studying complex dynamical systems where the underlying interactions are unknown but can be inferred from trajectory data, with potential applications in neuroscience, climate science, and systems biology.