⚛️ quantum physics

Automated discovery of high-probability heralded schemes for path-entangled states

This paper demonstrates how AI-driven automated discovery can identify and generalize heralded linear-optical schemes that exponentially improve the generation of large path-entangled states, offering a scalable route to overcome the limitations of weak photon interactions in photonic quantum technologies.

Original authors: Marcello Armezzani, Colin P. Lualdi, Xuemei Gu, Paul G. Kwiat, Mario Krenn

Published 2026-07-29

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Marcello Armezzani, Colin P. Lualdi, Xuemei Gu, Paul G. Kwiat, Mario Krenn

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where light isn't just a wave or a particle, but a team of dancers holding hands in perfect, invisible sync. In the realm of quantum physics, this "sync" is called entanglement. When photons (particles of light) become entangled, what happens to one instantly affects the others, no matter how far apart they are. This spooky connection is the secret sauce for future technologies like unbreakable codes, super-fast computers, and sensors that can see the tiniest details of the universe.

However, there's a catch. Photons are notoriously shy; they don't like to bump into each other or interact. In the real world, you can't just tell two photons to "hold hands" and stay that way on command. Usually, scientists have to try and try again, hoping that by sheer luck, the photons will end up entangled. This is like trying to build a skyscraper by throwing bricks into the air and hoping they land in a perfect tower. To make this work, scientists use a trick called heralding. Think of it as a "success signal." You set up a complex machine with many paths. If a specific, rare pattern of light appears in a side channel (like a referee blowing a whistle), you know the main event—the perfect entangled state—has been created in the other channels. You don't have to destroy the prize to know you won; the whistle tells you it's safe to use.

The big challenge has always been: How do we get lots of photons to dance in this perfect sync? As the number of photons grows, the odds of the "whistle" blowing successfully drop to almost zero, making large-scale quantum experiments incredibly difficult and slow.

This is where the story of the paper begins. The researchers, a team of scientists and AI experts, asked a bold question: What if we let a computer figure out the dance moves? Instead of relying on human intuition to design the optical circuits (the paths the light travels), they built a fast AI simulator to search through millions of possible arrangements. They weren't just looking for any solution; they were hunting for the most efficient way to get that "success signal" for large groups of entangled photons.

What they found was a game-changer. The AI didn't just tweak existing designs; it discovered a completely new, scalable family of solutions they call the "modular comb" family. Imagine trying to organize a massive parade. Old methods were like trying to line up everyone in a single, long, wobbly line. The new "modular comb" method is like organizing the parade into smaller, perfect squads that merge together in a specific, rhythmic pattern.

The results are staggering. For a specific target state with 9 photons (called a NOON9 state), the new AI-discovered circuit is nearly 1,440% more successful than the best previous method. That's not just a little better; it's a massive leap. In fact, as the number of photons increases, the new method doesn't just get slightly better; it gets exponentially better, leaving the old methods in the dust.

The paper also shows that this new "modular comb" idea isn't a one-trick pony. It works for different types of entangled states and even for situations where you need to entangle photons across more than two paths (multi-mode states). While the AI found some other unique, one-off solutions that were even better for specific numbers, the "modular comb" family is the big prize because it provides a general rulebook for building these circuits for any size.

Crucially, the authors didn't just stop at the math. They checked if this could actually be built in a real lab. They found that the new circuits are surprisingly simple to construct, requiring fewer detectors and less complex equipment than the old, unreliable methods. They also noted that while making the specific input light packets needed is still a challenge, recent advances in light sources suggest it's within reach.

In short, this paper uses AI to turn a nearly impossible task—creating large groups of entangled photons—into a much more manageable one. By discovering a new, scalable way to "herald" these states, the researchers have provided a clear path toward the next generation of quantum technologies, proving that sometimes, the best way to understand the universe is to let a computer help you find the dance steps.

Technical Summary: Automated Discovery of High-Probability Heralded Schemes for Path-Entangled States

Problem Statement
Entangled states of light are fundamental resources for photonic quantum technologies, including quantum communication, computation, and metrology. However, their deterministic generation is constrained by the lack of strong photon-photon interactions. Consequently, linear optical schemes typically produce multiphoton entangled states probabilistically via post-selection or heralding. Heralding is preferred as it certifies state preparation without consuming the target state, using auxiliary photon detection to announce success. While foundational strategies exist for generating path-entangled states (such as NOON states), the landscape becomes increasingly complex as the target photon number ( $N$ ) increases. The challenge shifts from parameterizing universal interferometers to identifying specific input structures and heralding mechanisms that enable the exact cancellation of unwanted amplitudes with high success probability. Existing passive linear-optical families, such as the Pryde-White (PW) and Zou-Pahlke-Mathis (ZPM) schemes, face limitations in efficiency and scalability for large $N$ .

Methodology
The authors employ a large-scale, automated discovery framework utilizing machine learning techniques to design heralded linear-optical circuits.

Simulation and Optimization: A universal multiport interferometer is modeled using the JAX library. The system is optimized via stochastic gradient descent (specifically the Adam optimizer) to maximize both the heralding success probability ( $p_{succ}$ ) and the fidelity ( $F$ ) with respect to a target state.
Differentiable Search: The discrete problem of selecting detector outcomes is embedded into a continuous optimization problem. A bank of possible detector patterns is assigned trainable real logits, converted to soft acceptance weights via a sigmoid function. This allows the simultaneous optimization of optical parameters (beam splitter angles, phases) and heralding conditions.
Generalization: The search is conducted over various input product states and target states. Once high-performing circuits are discovered, the authors analyze their underlying mechanisms to identify generalizable physical principles and scalable families.

Key Contributions and Results

Discovery of the Modular Comb Family:
The automated search identified a new, exact, and scalable family of circuits for generating NOON states of the form $|N, 0\rangle + e^{i\phi}|0, N\rangle$ .
- Mechanism: The scheme decomposes the total photon number $N$ into packets ( $N = \sum r_\ell m_\ell$ ). These packets pass through a multiport (Fourier-type transformation) to generate a "modular comb" of photon-number sectors. Unwanted intermediate sectors are eliminated using single-photon Fock filters (ancillary photons mixed on beam splitters), leaving only the desired edge terms $|N, 0\rangle$ and $|0, N\rangle$ .
- Performance: This family generalizes previous known schemes (PW and ZPM) as limiting cases. However, for non-trivial parameter choices, it offers exponential improvements over the PW scheme and super-exponential improvements over the ZPM scheme as $N$ increases.
- Example (NOON9): For $N=9$ , the new scheme achieves a success probability of $\approx 2.88\%$ , representing a $\sim 1440\%$ increase over the best existing scheme. It also reduces the number of required detectors and avoids reliance on unreliable "vacuum heralding" (detecting no photons), instead using a more robust $|1, 1\rangle$ heralding pattern.
Extension to Multi-Mode NOON States:
The modular comb mechanism was successfully generalized to multi-mode NOON states (simultaneous estimation of multiple parameters).
- The family improves upon the best-known linear-optical scheme by Zhang and Chan.
- Asymptotic analysis shows the new family provides a scaling advantage of $N^{1/2} 10^{cN}$ over the Zhang-Chan scheme, where $c > 0$ .
Discovery of Non-Modular Superior Schemes:
The automated search revealed specific, high-performance solutions that do not fit the modular comb family structure.
- NOON8: A compact circuit for $N=8$ was found that outperforms the ZPM construction by $\sim 494\%$ in success probability while requiring fewer detectors and lower input photon numbers (maximum input occupation of 3 vs. 4).
- $m, m'$ States: The strategy was extended to discover passive linear-optical schemes for $m, m'$ path-entangled states (robust to loss), outperforming previously known nonlinear families.
Experimental Feasibility Analysis:
The authors analyzed the practical implementation of the NOON8 and NOON9 circuits.
- Detectors: The schemes are robust to detector inefficiencies. The replacement of full vacuum heralding with $|1, 1\rangle$ patterns (e.g., in NOON9) maintains higher fidelity and probability as detector efficiency decreases.
- Sources: While the NOON9 scheme requires input Fock states $|3\rangle$ , which are challenging to generate, theoretical studies suggest "repeated addition" schemes could achieve success probabilities in the tens of percent range. The NOON8 scheme is advantageous as it requires only $|1\rangle$ and $|3\rangle$ inputs, avoiding the $|4\rangle$ requirement of previous schemes.
- Circuitry: The circuits require a small number of modes and beam splitters (five for each scheme), making them compatible with current free-space bulk optics or low-loss integrated photonic platforms.

Significance and Claims
The paper claims that automated discovery can transcend mere optimization to reveal new physical understanding. By generalizing the specific solutions found by the AI, the authors extracted a "modular comb" principle that explains how to construct scalable, high-probability heralding schemes.

Theoretical Advance: The work provides a new exact family for heralding NOON states that scales exponentially and super-exponentially better than the state of the art.
Experimental Route: The compact nature of the proposed circuits and their compatibility with existing detector and source technologies suggest that generating large path-entangled states (e.g., NOON8 and NOON9) is attainable in the near future.
AI as a Scientific Tool: The study demonstrates that AI can act as an agent of understanding, uncovering transferable physical mechanisms (like the modular comb) that generalize across different target states (two-mode, multi-mode, and $m, m'$ states), rather than simply reproducing known intuition.

Technical Summary: Automated Discovery of High-Probability Heralded Schemes for Path-Entangled States

More like this