The Cosmic Web on a Neural Leash express
In a dark room, music plays. Powder is scattered on the floor — fine, almost invisible. Low frequencies make it gather into intricate patterns: filaments, clumps, voids. In much the same way, gravity over billions of years weaves galaxies into the cosmic web — except here, attraction does the work instead of sound waves. To figure out what 'melody' it all started with, you need to rewind time: from the finished pattern back to the initial handful of powder. That is the task of reconstructing the initial conditions of the Universe.
Ordinary computer simulations are too heavy for such a journey backward. They honestly compute every gravitational nudge, but a single iteration can take hours. Approximate methods are faster, yet they cut corners — losing the very lace-like dynamics where halos are born. The authors propose a compromise: a cellular automaton with a neural network brain. The model LNCA (Lagrangian Neural Cellular Automaton) tracks not a fixed density grid but the matter particles themselves, moving along with them. It learns only to correct a rough but physically meaningful guess — the Zel'dovich approximation. On top of this foundation, the model adds careful residual displacements. This preserves accuracy on large scales, while training fills in the fine details.
The secret to this economy lies in locality. Each automaton cell updates its state by looking only at its nearest neighbors, and it does so step by step. The result is not just a final snapshot but the entire history of motion, like a frame-by-frame recording of a dance of dust particles. The rules of this dance rotate and shift with the matter: the architecture is rotationally and translationally equivariant. The final accuracy is impressive. Up to scales of about 12 megaparsecs (roughly 40 million light-years), the model errs by no more than a few percent. This provides a reliable pendulum for reversing time — from the modern web back to the nearly uniform beginning.
Original abstract
as written by the authors arXiv:2607.27320
Field-level inference of cosmological initial conditions from galaxy surveys requires a forward model that is simultaneously accurate in the non-linear regime, computationally efficient, and fully differentiable. Traditional N-body simulations are accurate but computationally prohibitive for iterative inference, while approximate solvers like Lagrangian Perturbation Theory (LPT) fail to capture the knotty halo-forming dynamics of the cosmic web at late times. We introduce the \textit{Lagrangian Neural Cellular Automaton} (LNCA), a hybrid deep learning framework that can be applied to emulate structure formation as a local, iterative dynamical process on a comoving lattice. Unlike standard Eulerian Convolutional Neural Networks (CNNs) which map fixed density fields, the LNCA operates in the Lagrangian frame, advecting the computational graph itself to follow the flow of mass. By training the network to learn only the \textit{residual} displacement corrections to the Zeldovich approximation, we achieve high-fidelity emulation of the non-linear physics while guaranteeing accuracy at large scales. We further constrain our model to produce complete trajectories, not just final states, by adopting an equivariant cellular automaton architecture, which recurrently iterates on its internal states to yield a dynamic history. The resulting model is strictly local, translationally and rotationally equivariant, and naturally supports continuous time integration, making it a reliable differentiable forward model for reconstructing the initial conditions of the universe from lightcone data. Our trained model supports percent-level precision in the power and cross spectra well into the non-linear regime ($$k \lesssim 0.5 \, h \text{Mpc}^{-1}$$), while requiring $$\sim10^4$$ times fewer learned parameters than comparable models which take the form of an interpretable internal dynamic rule set.
The Zel'dovich approximation, which the model takes as its foundation, was devised in 1970 to describe the growth of perturbations in an expanding universe. It has proven so successful that it still serves as the 'zeroth-order approximation' in many cosmological calculations.
In the novel 'The End of Eternity,' Isaac Asimov describes an organization that corrects history by calculating the consequences of the slightest changes. LNCA is a step toward such a 'reality computer,' capable of reconstructing initial conditions from the current pattern of galaxies.