The Gradient Discovery

A plasma simulation computes the evolution of charged particles and electromagnetic fields. The simulation is a forward model — given initial conditions and parameters, it produces the plasma's behavior. The physicist reads the output, adjusts the parameters, runs again. The loop is manual. The discovery is human.

The authors make the simulation differentiable. Every operation — particle pushing, field solving, collision handling — is written so that gradients can propagate backward through the computation. The simulation becomes not just a forward model but an optimization landscape. Specify an objective — maximize a signal, minimize a loss, find the conditions that produce a target behavior — and the gradient points the way.

The optimization discovers something the physicists did not expect. In kinetic simulations of wavepacket interactions, the gradient-driven search finds a regime where two wavepackets interact superadditively — their combined effect exceeds the sum of their individual effects. The regime was not predicted by existing theory. It was not sought by the optimization. The objective was specified without reference to superadditivity. The gradient, following the loss landscape of the differentiable simulation, navigated to a region of parameter space that the physicists had not explored.

The discovery is a byproduct of optimization. The gradient does not understand plasma physics. It computes derivatives. But the derivatives of a faithful physical simulation encode the physics, and following them through parameter space can reach configurations that intuition-guided exploration misses.

The simulation did not generate the phenomenon. The phenomenon was always there, in the equations, waiting in a corner of parameter space. The gradient found the path.