The lattice Boltzmann machine

Abstract

The lattice Boltzmann (LB) method is widely regarded as a very flexible and efficient solver for a broad class of transport phenomena in fluids, soft matter and related fields. In this paper, we posit that such flexibility is a direct result of its generative nature as a rule-driven information-processing machine. Its update, a relaxation toward a local target followed by a shift along the lattice, belongs to the same class of iterated relaxation dynamics as residual networks and transformers, with the difference that in LB the operators are fixed by conservation and symmetry rather than learned from data. We show that three properties of this update are sufficient to promote the emergence of closed macroscopic laws from the underlying iteration process, the physical content of the target deciding the specific law in question. A machine whose target is drawn at random generates the transport law predicted from its moments. A network given the same three properties on a graph, where there is no lattice, is trained on a traffic process and generates a conservation law of its own. The mechanism has a precise limit, where the damping of the non-conserved modes fails, and on the lattice the macroscopic law gives no warning of it. In this reading, LB is a lattice Boltzmann machine whose informing rules are supplied by physics rather than by training. Its three properties then state what a new ingredient should preserve when it enters an LB model, whether it comes from physics or from data.