esQueranto: Differentiable Structured Quantum Light for Automated Scientific Discovery

AI summary

Overview

Research area: Quantum photonics, structured light, differentiable physics simulation, and AI-driven automated experiment design.

Technical level: Advanced. The paper assumes familiarity with Fock-space quantum optics, beam optics, and gradient-based optimization, though its central argument is stated in plain terms.

Scope: This paper introduces esQueranto, a JAX-based differentiable simulator that places photon-number quantum optics and paraxial structured-light propagation inside one calculation, and demonstrates it across six quantum-optical applications.

What This Paper Is About

Automated "de-novo" experiment design uses search algorithms plus a physics simulator to propose experiments humans would not think of, so the range of discoverable experiments is limited by what the simulator can physically describe. Existing photonic simulators cover either photon-number quantum optics or classical structured wave optics, but not both in one differentiable model. The authors build esQueranto to close that gap, so that the spatial evolution of a beam (diffraction, lenses, Gouy phases, mode overlap) directly changes non-classical quantum states and their interference, and so that the whole calculation can be differentiated and optimized.

Key Contributions

1. A unified physical description. esQueranto combines Fock-basis photon-number quantum optics with Laguerre–Gaussian (LG) transverse modes propagated through the ABCD/paraxial formalism, including mode-dependent Gouy phases and the change of transverse basis that occurs when differently propagated fields interfere at a beam splitter.

2. Differentiability as a first-class feature. The framework is implemented in JAX, providing automatic differentiation, just-in-time compilation, vectorized/parallel execution, and hardware acceleration. A circuit can be compiled once with named continuous parameters, then re-evaluated and differentiated without rebuilding the experimental topology.

3. A common interface spanning sources, noise, and measurement. One workflow — space → state → circuit → simulate → measure — covers Fock, coherent, and squeezed inputs; beam splitters, phase shifts, free-space propagation and lenses; photon loss and photon-number dephasing channels; and photon-number-resolving detection with efficiency, finite resolution, conditional (heralded) states, and fidelity.

4. Six worked applications plus a composed experiment. The demonstrations span multiphoton interference between nonlinear pair sources, NOON-like state generation, high-probability heralding of a path-entangled state, lossy/dephased interferometry, Hong–Ou–Mandel interference with structured transverse profiles, and a final setup that combines squeezed-light heralding, structured propagation, and dephasing for radial-mode sorting.

Main Findings

• Spatial propagation is not a trivial add-on to Fock space. Because lenses and propagation change the spatial modes, they change the Fock basis itself; photons that start in orthogonal modes can become partially indistinguishable, directly altering quantum interference outcomes.

• Gouy phases are photon-number dependent. An n-photon occupation acquires n times the single-photon mode-dependent phase, and the paper notes this has been observed experimentally for two-photon states (Ref. [21]).

• Interference between differently propagated fields is handled in a shared basis. When two inputs with different beam parameters reach a beam splitter, both are re-expressed in a common transverse reference chosen symmetrically, so beam-width and wavefront-curvature differences modify the interference amplitudes rather than being inserted as an external visibility factor.

• Higher-order pair emission shifts multiphoton interference. In the four-two-mode-squeezer setup with a phase shift of π in path a, the low-gain destructive-interference condition for P(1,1,1,1) no longer holds at stronger squeezing; the simulated minimum shifts because of higher-order Fock components, in agreement with the analytical high-gain treatment of Ref. [30].

• NOON-like states from classical and quantum light are reproduced. A coherent state and a squeezed vacuum mixed at a balanced (50:50) beam splitter produce a joint photon-number distribution in which photons tend to emerge together from either output port.

• Automatically discovered heralding schemes can be re-evaluated. The optimized NOON₈ scheme from Ref. [24] — a four-mode input |3,3,3,1⟩ through five beam splitters, simulated in a fixed-truncation four-mode space with cutoff 10 — is expressed directly in the framework.

• Positioning relative to existing tools (Table 1). In the paper's comparison table, esQueranto is marked as natively supporting the Fock representation, photon loss, dephasing, structured-light propagation, and automatic differentiation, while not supporting the Gaussian representation, general open-system dynamics, vectorial classical wave propagation, or frequency-domain interferometry.

• Reported scope limits and missing figures. The provided text is truncated partway through Section III.3, so quantitative results for the later applications (interferometry under loss and dephasing, structured HOM interference, radial-mode sorting) are not reported in the available content. Appendix A is said to benchmark the CPU-to-GPU crossover for passive universal interferometers, but no benchmark numbers appear in the supplied text.

Methodology in Plain English

The authors built a software library rather than running a single experiment. States are stored as vectors or density operators in the Fock basis, with three truncation choices (global, fixed photon number, and per-mode cutoff) whose Hilbert-space dimensions are given explicitly in the paper. Optical elements are written as operators acting on those states, and a circuit is just an ordered list of them.

For structured light, each optical path is expanded into Laguerre–Gaussian modes indexed by azimuthal number ℓ and radial number p, and those profiles are propagated using ray-transfer (ABCD) matrices for free space and thin lenses. This lets the code track beam radius, wavefront curvature, and Gouy phase, and it lets the code convert between mode bases when two beams with different histories meet.

The computational trick is separation of static and dynamic structure: occupation maps and contraction patterns are prepared once at the NumPy level, while the numerical evolution runs in JAX so that state amplitudes and physical parameters stay inside the traced, differentiable computation. A compile step exposes named parameters (angles, phases, squeezing strengths, loss, dephasing, geometry), and users can attach any differentiable objective, such as one minus the fidelity to a target state, and obtain gradients with respect to all of them. Structural search over different circuit topologies is left to an external procedure.

Why This Matters

Automated scientific discovery is only as good as the physics its simulators can express. By putting spatial beam physics and non-classical photon statistics into one differentiable object, esQueranto widens the set of experiments a search algorithm can propose — particularly experiments whose signal is carried by a spatial change in a quantum state.

Real-world applications implied by the demonstrated capabilities:

• Quantum metrology and sensing: NOON-like states and high-probability heralding schemes for phase super-resolution, plus interferometry evaluated under realistic loss and dephasing.
• Structured-light quantum imaging and sensing: radial-mode sorting that converts mode-dependent Gouy phases into path-dependent interference, and Hong–Ou–Mandel interference between photons with tailored transverse profiles.
• Photonic quantum computing and entanglement engineering: modeling higher-order pair emission from nonlinear sources, which affects entanglement-generation success probabilities and interference fidelity.
• Computational optics and instrument design: gradient-based optimization of lens positions, waists, and propagation distances alongside quantum-optical parameters in the same objective.

Industry relevance: photonic quantum hardware, quantum sensing and metrology instruments, optical and microscopy system design, and any workflow where gradient-based optimization of a physical apparatus replaces manual parameter tuning.

Future Directions

• Extend beyond the stated model limits. The paper explicitly restricts the structured-light description to paraxial, coaxial fields at a common wavelength, with temporal, spectral, and polarization degrees of freedom outside the spatial model. Adding polarization, frequency/time structure, and vectorial propagation is the natural next step.

• Add open-system dynamics and frequency-domain interferometry. Table 1 marks esQueranto as not supporting general open-system dynamics or frequency-domain descriptions, both of which are capabilities of existing specialized tools.

• Couple the simulator to structural search. The current design fixes a topology and differentiates continuous parameters, leaving exploration of alternative structures to an external procedure; closing that loop is the route to fully automated discovery.

• Experimental validation of the composed proposals. The final example composes squeezed-light heralding, structured propagation, and dephasing for radial-mode sorting; whether such combined designs outperform dedicated setups remains an experimental question.

Target Audience

Researchers and engineers working on photonic quantum technologies, structured light, quantum metrology, and computational optics, as well as scientists building AI-driven or gradient-based tools for automated experiment design. It is also relevant to developers of quantum simulation software who want to understand where a Fock-plus-structured-light simulator fits relative to MrMustard, Piquasso, Perceval, Strawberry Fields, QuTiP, XLuminA, Chromatix, Finesse, and Differometor. Readers without a quantum optics background will find the motivation accessible, but the technical sections require comfort with Fock states, beam optics, and automatic differentiation.

Authors’ abstract

Uncovering new phenomena in nature requires experiments. For centuries, their design has been exclusively a human endeavor. Today, a new paradigm is emerging in which artificial intelligence and computational methods can design experiments themselves. Realizing this form of automated scientific discovery requires simulators that are sufficiently expressive to represent the diverse physical processes and couplings from which new experiments can be constructed. In this direction, we introduce \textsc{esQueranto}, a differentiable software that brings photon-number quantum optics and structured-light propagation into a common description, allowing the spatial evolution of light to directly influence non-classical quantum states and their interference. \textsc{esQueranto} is implemented in JAX, providing automatic differentiation and hardware-accelerated evaluation for optimization and automated search. We demonstrate the framework across a broad range of quantum-optical applications that exercise complementary aspects of its physical description. By combining quantum and structured-light physics within a single differentiable simulator, \textsc{esQueranto} takes an important step towards the dream of a foundational simulator in physics.