A Quantum-Optics Simulator Can Now Tune the Experiment It Models
esQueranto puts photon-number physics and structured light in one differentiable JAX program, so researchers can optimize optical parameters instead of only sweeping them.
Most optical simulators answer a forward question: given a source, interferometer and detector, what output should appear? esQueranto adds a second capability. Because its continuous parameters remain differentiable, the same program can calculate how a change in those parameters would move an objective up or down.[1][2]
One model for two kinds of light
The framework combines photon-number, or Fock-space, operations with spatial propagation of structured optical modes. That matters when the internal quantum state and the shape of the light field cannot be treated as independent bookkeeping problems.[1][2]
The public code includes objects for Fock states, beam splitters, loss channels, measurements and heralding, alongside thin lenses and free-space propagation. Its JAX foundation supplies automatic differentiation and just-in-time compilation.[2]
Optimization replaces part of the sweep
The preprint demonstrates six simulated workflows: gravity-sensitive interferometry, heralded radial-mode sorting, structured Hong-Ou-Mandel interference, N00N-state heralding, high-N00N interference and four-crystal interference. In each case, differentiability turns selected optical settings into variables that can be optimized against a target.[1][2]
That does not mean the program invents an unrestricted experiment. A researcher still chooses the topology, components and objective; the software tunes continuous parameters within that declared model. The distinction is important because it separates a useful design engine from claims of autonomous discovery.[1][2]
The laboratory test is still ahead
All reported examples are computational. The work does not show an optimized design transferred to physical optics, nor does it establish an advantage over every existing simulator. Model mismatch, calibration error and fabrication tolerances can all weaken a numerically clean optimum.[1][2]
Independent users can now test those limits because the implementation is public. The strongest validation would connect the differentiable model to calibrated hardware and measure how much iteration time it saves compared with conventional simulation-and-sweep workflows.[2]