AM Seminar: Engineering the Earth’s Climate

Presenter: Dr. Pulkit Dubey, Postdoc, UC Santa Cruz

Description: Neural climate emulators such as NeuralGCM and LUCIE offer efficient, differentiable alternatives to General Circulation Models (GCMs), producing climate predictions at a fraction of the cost. While work to date has focused largely on predictive accuracy, we leverage differentiability to study control of long-horizon climatological targets. Classical GCMs approach this via adjoint-based optimization. Backpropagation through time (BPTT) is its neural-network analog and inherits the same chaotic gradient explosion at long rollouts. We combine BPTT-based sensitivities with receding-horizon optimization to mitigate the chaotic divergence and enable meaningful control over climatological targets. We illustrate with two candidate climate-cooling strategies and close by sketching reinforcement-learning extensions.

About the speaker: Pulkit Dubey is a postdoc in the Department of Applied Mathematics at UC Santa Cruz. He earned his PhD at the University of New Hampshire on the simulation and modeling of turbulent flows, where he developed hybrid solvers for 2D turbulence. He joined UCSC in September 2025, where he works on control strategies for neural climate emulators, enabling long-horizon control over statistical targets in chaotic dynamical systems.

This seminar is hosted by Professor Nilah Ioannidis.