Physics simulators can support the discovery of quantum experiments by predicting the states generated by experimental configurations. When these simulators are computationally expensive, repeated simulator calls can limit the search for experiments that generate a desired quantum state. Here, we develop a physics-inspired polynomial neural surrogate for PyTheus, a graph-based quantum-optics simulator, to predict quantum states and use it to design quantum experiments. Its polynomial activations are motivated by the relation between graph perfect matchings and the resulting state amplitudes. We train separate surrogate models for four-, six-, and eight-photon systems and show that they achieve higher prediction accuracy with fewer trainable parameters than standard multilayer perceptrons. We then use the trained surrogates for inverse design of GHZ, W, and linear-cluster states. For the larger systems, the surrogates also enable faster inverse design than direct optimization with PyTheus. These results suggest that incorporating the underlying physics into neural surrogates can provide an efficient approach to quantum experiment design.
AI Summary
Can a neural surrogate accelerate repeated inverse design of photonic experiments while remaining reusable across target states? This paper learns PyTheus’s mapping from weighted, edge-coloured graphs to complete unnormalised state amplitudes. Since amplitudes are sums of perfect-matching contributions of degree $n/2$ in the graph weights, it uses a single-hidden-layer polynomial neural network with activation $z^{n/2}$, trained separately for four, six, and eight photons.
Each model uses 20 million simulator-generated examples, split 80/10/10% for training, validation, and testing. Test mean absolute amplitude errors are approximately $9.16\times10^{-7}$, $3.93\times10^{-6}$, and $5.93\times10^{-5}$, respectively. Against the tested three-hidden-layer GELU networks, the polynomial models have lower validation errors and fewer parameters; at eight photons, the comparison is 5.54 million versus 131 million parameters. Averaged absolute Jacobian sensitivities also closely match analytical calculations. Inverse design is tested on GHZ, W, and linear-cluster targets with 1,000 searches per target and size. Independent PyTheus evaluation gives fidelity at least 0.9999 for all 9,000 resulting graphs. Eight-photon mean optimization times fall from roughly 1,524–10,350 seconds with CPU PyTheus to 6–15 seconds with GPU surrogates, excluding dataset generation and training.
The stated limitations include real-valued weights, two internal modes, separate models for each size, growing hidden-layer requirements, and exponentially many output amplitudes. The architecture encodes polynomial degree but not the exact perfect-matching monomials. Runtime gains concern repeated post-training design, not total end-to-end cost.
Little sign of AI-generated textIncrementalSmall numerics onlyMissing comparison
Error found: Several uncertainties in the inverse-design runtime table exceed their nonnegative sample means, which is impossible for the standard errors of the mean claimed in the caption.
The graph-to-state derivation incorrectly restricts the n-photon contribution to expansion order n/2 despite including annihilation operators in the exponential, so the stated generating equation does not directly imply the perfect-matching polynomial.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
Physics simulators can support the discovery of quantum experiments by predicting the states generated by experimental configurations. When these simulators are computationally expensive, repeated simulator calls can limit the search for experiments that generate a desired quantum state. Here, we develop a physics-inspired polynomial neural surrogate for PyTheus, a graph-based quantum-optics simulator, to predict quantum states and use it to design quantum experiments. Its polynomial activations are motivated by the relation between graph perfect matchings and the resulting state amplitudes. We train separate surrogate models for four-, six-, and eight-photon systems and show that they achieve higher prediction accuracy with fewer trainable parameters than standard multilayer perceptrons. We then use the trained surrogates for inverse design of GHZ, W, and linear-cluster states. For the larger systems, the surrogates also enable faster inverse design than direct optimization with PyTheus. These results suggest that incorporating the underlying physics into neural surrogates can provide an efficient approach to quantum experiment design.
Can a neural surrogate accelerate repeated inverse design of photonic experiments while remaining reusable across target states? This paper learns PyTheus’s mapping from weighted, edge-coloured graphs to complete unnormalised state amplitudes. Since amplitudes are sums of perfect-matching contributions of degree n/2 in the graph weights, it uses a single-hidden-layer polynomial neural network with activation zn/2, trained separately for four, six, and eight photons.
Each model uses 20 million simulator-generated examples, split 80/10/10% for training, validation, and testing. Test mean absolute amplitude errors are approximately 9.16×10−7, 3.93×10−6, and 5.93×10−5, respectively. Against the tested three-hidden-layer GELU networks, the polynomial models have lower validation errors and fewer parameters; at eight photons, the comparison is 5.54 million versus 131 million parameters. Averaged absolute Jacobian sensitivities also closely match analytical calculations. Inverse design is tested on GHZ, W, and linear-cluster targets with 1,000 searches per target and size. Independent PyTheus evaluation gives fidelity at least 0.9999 for all 9,000 resulting graphs. Eight-photon mean optimization times fall from roughly 1,524–10,350 seconds with CPU PyTheus to 6–15 seconds with GPU surrogates, excluding dataset generation and training.
The stated limitations include real-valued weights, two internal modes, separate models for each size, growing hidden-layer requirements, and exponentially many output amplitudes. The architecture encodes polynomial degree but not the exact perfect-matching monomials. Runtime gains concern repeated post-training design, not total end-to-end cost.
Photonic technologies are main players in the second quantum revolution, providing better sensors, secure communications, and quantum-enhanced computation. Such endeavors require generating specific quantum states or efficiently performing quantum tasks. The design of the corresponding optical experiments, historically powered by human creativity, is being slowly automated by novel algorithms. Unfortunately, these tools are often restricted to very specific use cases and are difficult to generalize, which limits their practical implementation. To overcome these challenges, we developed PyTheus, a highly efficient, open-source digital discovery framework based on a graph-based representation of optical setups, which includes a wide range of modern experimental devices. PyTheus produces interpretable designs to solve complex experimental problems, like entanglement generation and distribution. Aiming for the simplest solutions, our software provides deeper understanding to human researchers, which can generalize the findings. Therefore, PyTheus can accelerate the development of quantum optics and related technologies.
Photons are the physical system of choice for performing experimental tests of the foundations of quantum mechanics. Furthermore, photonic quantum technology is a main player in the second quantum revolution, promising the development of better sensors, secure communications, and quantum-enhanced computation. These endeavors require generating specific quantum states or efficiently performing quantum tasks. The design of the corresponding optical experiments was historically powered by human creativity but is recently being automated with advanced computer algorithms and artificial intelligence. While several computer-designed experiments have been experimentally realized, this approach has not yet been widely adopted by the broader photonic quantum optics community. The main roadblocks consist of most systems being closed-source, inefficient, or targeted to very specific use-cases that are difficult to generalize. Here, we overcome these problems with a highly-efficient, open-source digital discovery framework PyTheus, which can employ a wide range of experimental devices from modern quantum labs to solve various tasks. This includes the discovery of highly entangled quantum states, quantum measurement schemes, quantum communication protocols, multi-particle quantum gates, as well as the optimization of continuous and discrete properties of quantum experiments or quantum states. PyTheus produces interpretable designs for complex experimental problems which human researchers can often readily conceptualize. PyTheus is an example of a powerful framework that can lead to scientific discoveries – one of the core goals of artificial intelligence in science. We hope it will help accelerate the development of quantum optics and provide new ideas in quantum hardware and technology.
Preparing quantum many-body states on classical or quantum devices is a very challenging task that requires accounting for exponentially large Hilbert spaces. Although this complexity can be managed with exponential ansätze (such as in the coupled-cluster method), these approaches are often tailored to specific systems, which limits their universality. Recent work has shown that the contracted Schrödinger equation enables the construction of universal, formally exact exponential ansätze for quantum many-body physics. However, while the ansatz is capable of resolving arbitrary quantum systems, it still requires a full calculation of its parameters whenever the underlying Hamiltonian changes, even slightly. Here, inspired by recent progress in operator learning, we develop a surrogate neural network solver that generates the exponential ansatz parameters using the Hamiltonian parameters as inputs, eliminating the need for repetitive computations. We illustrate the effectiveness of this approach by training neural networks of several quantum many-body systems, including the Fermi-Hubbard model.
Designing photonic circuits that prepare graph states with high fidelity and success probability is a central challenge in linear optical quantum computing. Existing approaches rely on hand-crafted designs or fusion-based assemblies. In the absence of multiplexing/boosting, both post-selected ballistic circuits and sequential fusion exhibit exponentially decreasing single-shot yields - a fundamental limitation that makes optimizing individual resource state generators particularly important, as these serve as building blocks in larger FBQC architectures. We present a general-purpose optimization framework for automated photonic circuit discovery using a novel polynomial-based simulation approach, enabling efficient strong simulation and gradient-based optimization. Our framework employs a two-pass optimization procedure: the first pass identifies a unitary transformation that prepares the desired state with perfect fidelity and maximal success probability, and the second pass implements a novel sparsification algorithm that reduces this solution to a compact photonic circuit with minimal beamsplitter count while preserving performance. This sparsification procedure often reveals underlying mathematical structure, producing highly simplified circuits with rational reflection coefficients. We demonstrate our approach by discovering optimized circuits for 3-, 4-, and 5-qubit graph states across multiple equivalence classes. For 4-qubit states, our circuits achieve success probabilities of 2.053×10−3 to 7.813×10−3, outperforming the fusion baseline by up to 4.7×. For 5-qubit states, we achieve 5.926×10−5 to 1.157×10−3, demonstrating up to 7.5× improvement. These results include the first known state preparation circuits for certain 5-qubit graph states.
Quantum physics-informed neural networks (QPINNs) have recently emerged as a promising framework for the solution of partial differential equations (PDEs), with several studies reporting improved convergence and accuracy relative to classical physics-informed neural networks (PINNs) at reduced training cost. Motivated by these advances, we introduce geometric quantum physics-informed neural networks (GQPINNs), a symmetry-aware extension of QPINNs in which the geometric structure of the underlying PDE is incorporated directly into the quantum-circuit ansatz. Building on the framework of geometric quantum machine learning, we construct parametrized circuits that encode finite-group and compact Lie-group symmetries as inductive biases through problem-specific equivariant generator sets . Using a twirling-based construction, we derive symmetry-preserving gates that ensure that the model predictions respect the symmetries of the governing equation whenever the boundary and initial data are symmetry compatible. We benchmark GQPINNs against standard QPINNs and symmetry-adapted classical PINN baselines under matched training protocols across a representative set of linear and nonlinear PDEs. Across these benchmarks, GQPINNs achieve improved solution accuracy, as quantified by lower mean absolute error, while requiring substantially fewer trainable parameters. These results identify symmetry-aware quantum-circuit design as an effective route toward improved efficiency and generalization in quantum PDE solvers and provide a systematic framework for incorporating geometric inductive biases into quantum-enhanced scientific machine learning.
Integrated quantum photonics requires chip-scale biphoton sources, but phase matching, nonlinear coupling, loss, and spectral purity are coupled in high-index anisotropic waveguides. We present NanoPhotoNet-MPM, a physics-informed neural surrogate for modal spontaneous parametric down-conversion (SPDC) in monoclinic NbOCl2 waveguides (d_22≈100 pm V^(-1)). The framework couples a DeepONet transverse eigensolver, regularised by the full-vectorial Helmholtz equations, with a quantum coupled-wave-equation propagator trained on 70 audited finite-difference eigenmode geometries. The trained model predicts a degenerate phase-matched crossing at w=283.13 nm and h=500.00 nm with an effective-index MAE of (4.16±0.75)×〖10〗^(-3), outperforming Gaussian process regression and plain multi-layer perceptron baselines by one to two orders of magnitude while guaranteeing wave-equation consistency. Joint spectral amplitude analysis at the degenerate design yields a heralded purity of P=0.122 (K=8.20), limited by symmetry-enforced group-velocity mismatch. A multi-mode genetic-algorithm search across Type-0, Type-I, and Type-II SPDC configurations identifies a Type-II design (m_p=4, m_s=0, m_i=1) at w=592.0 nm and h=270.0 nm that reaches P≈0.350 at L=0.5 mm with 95.9% idler survival under realistic propagation loss. The resulting framework is a generalizable surrogate-assisted design and active-learning guidance method for anisotropic van der Waals quantum photonic platforms.