Previously, machine learning inferred properties only for limited families of optical quantum states. A unified model for practically important multimode non-Gaussian states without full tomography was missing. Here, the first foundation model is presented for characterizing states across a wide complexity range—determined by non-Gaussianity, number of modes, and squeezing level. Pre-trained on low-complexity states, it directly applies to more challenging ones; with minimal fine-tuning, it predicts quantum fidelity and Wigner negativity for diverse experimental states: Schrödinger cat states, systems up to ten modes, and squeezed states up to 10.4 dB. This unified platform enables efficient state certification from limited measurement data, with impact on optical quantum computing, communications, and metrology.
Usually, to understand a quantum state of light, you need to perform dozens of complex measurements—like trying all the keys on a keychain. The new AI model works differently: it's a universal key that, after a slight adjustment, opens many 'locks'—simple and the most tricky states. To do this, it only needs a rough estimate of intensity—photometry—instead of labor-intensive spectroscopy.
The model predicts how close the state is to ideal and whether it has unusual quantum properties—for example, negativity of the Wigner function, without which data teleportation is impossible. It's especially impressive that it even handles Schrödinger's cats (light oscillating in two opposite phases at once) and multi-beam squeezed states, where each beam is like a note in an orchestra, playing softer than a whisper but together producing a pure tone.
This key dramatically simplifies verification: for the largest optical quantum computer with a hundred beams, without the model, years of measurements would be needed—now it takes just minutes. This paves the way for accelerated development of quantum networks transmitting photons at the speed of light and ultra-precise sensors.
🎯 The largest optical quantum computer uses over a hundred beams of light. Without the AI model, verifying its operation would take years.