Differentiable programming makes every operation in a simulation chain differentiable with respect to its inputs, allowing gradients to propagate backward from the output to any parameter. This is the technical infrastructure of modern machine learning. Joglekar et al. (arXiv:2603.11231) apply it to plasma physics simulations and find something that the technique was not designed to find: a previously unknown nonlinear regime.
When optimizing a differentiable kinetic simulation of laser-plasma interactions, the gradient-based search converged on a parameter configuration where two wavepackets interact superadditively — their combined effect exceeds the sum of their individual effects. This regime was not predicted by existing theory. It was not discovered by analytical derivation, dimensional analysis, or physical intuition. It was discovered by a gradient descent algorithm that was told to maximize an output and followed the landscape to a region that physicists had no reason to visit.
The structural interest is in why the gradient found what intuition missed. A physicist choosing parameters to simulate draws on experience, dimensional analysis, and known scaling laws. These heuristics are efficient — they concentrate attention on the regions of parameter space where interesting physics is expected. But they also impose a prior. The gradient has no prior. It follows the loss surface wherever it leads, including into regions that no heuristic would suggest. The superadditive regime exists in the equations. It was always there. The gradient's advantage was not superior physical understanding but the absence of any physical understanding at all.
This is not the familiar story of machine learning finding patterns in data. The simulation is not a black box — it is a first-principles kinetic solver whose physics is fully understood. The gradient propagates through known equations. What is new is not the physics of the simulation but the trajectory through its parameter space. The optimizer explored a path that no hand-designed parameter sweep would traverse, because the path passes through regions that look uninteresting at every intermediate step. The gradient sees the slope of the objective function at each point and follows it; a physicist scanning parameter space would skip the intermediate region and never reach the destination.
The paper also demonstrates 140-fold acceleration of Thomson scattering analysis and inverse design of laser pulses where full space-time coupling outperforms separated optimization by fifteen times. These are engineering gains — making existing workflows faster. The superadditive regime is qualitatively different. It is a discovery of physics by a tool that does not understand physics.
The gradient has no intuition. That is its advantage.