Zhixin Lu Scientist, Allen Institute
How do evolving systems remember, infer, generate, and learn?
I connect dynamics for AI with AI for dynamics to understand and build intelligence. Physical AI and learning are central interests in this program.
Invited talk · October 2Discovering Stochastic Dynamics from Snapshots · BPPB Seminar Recording forthcoming
The research program
I study how evolving states represent, retain, and transform information, and how observations reveal the systems that produce them.
How do the geometry and evolution of computational states determine what a system can remember, infer, and generate? I connect dynamical theory with the capabilities of recurrent networks, attention, and adaptive systems.
Principles of intelligent computationWhat can observations reveal about the processes that produced them? I develop methods to infer hidden states and dynamical laws from partial measurements and population snapshots, including what the data leave undetermined.
From observations to dynamical lawsFeatured research October 2026
Can a system generate chaos backward in time?
A question about the limits of stable generation.
Replaying a trajectory backward is easy. Building a system that stably generates that reversed motion is a different problem.
For a class of chaotic dynamics including the classical Lorenz attractor, I prove an obstruction to exact, stable reversed generation by finite-dimensional autonomous systems with C2 vector fields, compact attracting sets, and continuous readouts. Adding latent variables does not remove the obstruction within this class.
My work brings together nonlinear dynamics, stochastic processes, machine learning, and neuroscience. I earned my Ph.D. with Edward Ott at the University of Maryland and was a postdoctoral researcher with Danielle Bassett at the University of Pennsylvania.
My research has explored prediction of spatiotemporal chaos, reservoir observers for hidden-state inference, learned attractors, and source separation using simulated water dynamics. These studies motivate a broader ambition: to understand the dynamical principles that make intelligent computation possible.
Ideas & explanations
Proof intuition, dynamical principles, and interactive explanations.
New · DyNoSeD & BPPB seminar
How a stationary cloud constrains stochastic dynamics, why different laws can produce the same snapshots, and what the data actually identify. With an interactive example.
New · October 2026
The geometric idea behind my recent theorem: how connectedness and entropy obstruct exact, stable generation of reversed chaos.
Part I · Recurrent models
How a driven network can form an internal representation of another dynamical system.
Part II · Attention
A geometric view of self-attention, with an interactive landscape to explore.