betaFeedback

WorksPhysics and Astronomy

Complex fluid flow dynamics by bright scattershot boson sampling

Abstract

Boson sampling machines have become a fruitful tool for establishing new concepts, which connect fundamental physics with computer science applications. Here, we exploit the intrinsically high-dimensional correlations generated by the measurement outcomes of a multiphoton boson sampling machine and demonstrate how a reconfigurable architecture can be programmed to autonomously model the time evolution of nonlinear dynamical systems that describe complex fluid flows. By extending on the original scattershot boson sampling (SBS) scheme to include higher photon number components – a scheme which we label bright scattershot boson sampling (BSBS) — this scheme overcomes scaling issues inherent to SBS-based schemes by utilizing all randomly generated input photon number distributions. We illustrate the capability of this scheme for two complex fluid-flow models: an eight-dimensional Lorenz-type model of two-dimensional thermally driven convection and a 16-dimensional model of global atmospheric circulation. Using a digital twin of the BSBS architecture based on multiphoton statistics, we show that the photonic approach accurately reproduces the temporal dynamics of both systems while being resilient to photon losses of up to 50 per cent. We also verified that the BSBS case outperforms a comparable run with coherent light. Moreover, its prediction horizon exceeds that of a classical reservoir-computing model with the same query complexity. These results highlight the potential of near-term photonic quantum computing for addressing relevant real-world tasks on noisy intermediate-scale devices.

Introduction

Boson sampling [1] and its extension to Gaussian input states [2, 3, 4] has emerged as a route to demonstrating quantum computational advantage with photonic systems [5, 6, 7, 8, 9]. While initially regarded as an abstract concept, albeit one with important implications connecting fundamental principles of physics and computer science, the solving of tasks with practical application has since been demonstrated, including the modeling of molecular spectra [10, 11] and clique detection in macromolecular networks [12]. Initially, these tasks were motivated by the possibility of demonstrating a practical quantum advantage, however, follow-up investigations led to the development of more efficient classical solutions [13, 14] and better understanding of where the complexity in these problems lies [15]. These findings motivated the search for investigations into alternative applications, such as Monte Carlo integration [16, 17, 18] and asset clustering [19], as well as alternative boson sampling architectures [20, 21], where it is hoped that a practical quantum advantage may be found.

An alternative avenue for realizing practical quantum advantage is in using these boson sampling machines to implement (quantum) machine learning tasks [21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32]. In particular, boson sampling-based machines have been primarily used for classification and image recognition [22, 25, 26, 28, 29]. Quantum machine learning (QML) algorithms successfully exploit many-particle quantum systems to perform classification, regression, and prediction tasks, as well as generate new synthetic data [33, 34, 35, 36]. QML is one of the most promising applications of Boson samplers because these systems provide a natively large Hilbert space. Furthermore, some system biases are expected to be compensated for by the algorithms, see e.g. a discussion in [37] in classical cases. In this mode of operation, the boson sampler can be regarded as a “quantum extreme learning machine,” or more generally a quantum reservoir computing model (RCM) [38, 39, 40].

Recently, first proof-of-principle demonstrations of reservoir computing have been realised on photonic platforms. Nerenberg et al. [41] proposed a photon-number-resolving quantum RCM and discussed several input schemes based on different polarisations. Paparele et al. [23] used squeezed states to forecast one-dimensional nonlinear time series. Selimović et al. [42] presented a reservoir computing model based on a photonic quantum memristor implemented using a tunable Mach-Zehnder interferometer. They applied this model to predict one-dimensional nonlinear time series and the dynamics of a single degree of freedom governed by the chaotic Mackey-Glass equation [43]. More recently, a boson-sampling architecture has been used to demonstrate time-series forecasting and investigate the advantage offered by indistinguishable photons [9, 44]. Building on these proof-of-principle demonstrations, the field is now exploring schemes that enable scaling in the number of input states, system size, or feature-space dimension. Further applications beyond classification are also being sought to address real-world problems, with particular interest in schemes capable of investigating higher-dimensional nonlinear systems.

In this work, we present a scheme that addresses scalability by exploiting the high dimensionality of the Fock space of bright parametric down-conversion sources as a deterministic generator of two-mode squeezed vacuum (TMSV) states. In this way, we purposely exploit higher-order photon-pair emission events to expand the available parameter space and thus efficiently harness higher-order correlation functions. The presented scheme is therefore an extension to what is commonly referred to as scattershot boson sampling, which we coin bright scattershot boson sampling (BSBS). We also introduce an encoding scheme to predict complex fluid-flow dynamics without solving the underlying nonlinear partial differential equations. For training the system the full multiport interferometer is programmed such that the scheme investigates trajectories in higher-dimensional state spaces than those considered in previous approaches [44], taking a step towards modelling fully developed turbulence on a quantum photonic platform. To this end, we map a reservoir computing model (RCM) onto a reconfigurable bright scattershot boson sampler, exploiting the intrinsically high-dimensional correlations among measurement outcomes.

We evaluate this approach using a “digital twin” of a small-scale multiphoton boson sampler, simulated on a high-performance computer. Using this simulated device, we predict the evolution of two nonlinear dynamical systems with up to 16 degrees of freedom that capture complex flow behavior. We benchmark its performance against a classical RCM and estimate the device scale beyond which simulating the photonic model and reconstructing the dynamics would exceed the computational resources available on our high-performance computer. For the more complex learning task, the photonic RCM achieves a prediction horizon twice that of its classical counterpart. These results outline a route towards practically relevant dynamical prediction using scattershot boson sampling, with an implementation within reach of existing quantum photonic technology.

Multiphoton reservoir computing model

Building blocks

Photonic reservoir computing models predict dynamics with quantum states of light [45, 46]. Figure 1 compares classical and photonic RCMs; both consist of three building blocks, an input layer, the reservoir with memory, and an output layer, the latter of which is optimized only in a training procedure. Next to the one-step open-loop scenario, RCMs can be set up in a closed-loop regime, which will be considered here, see Figure 1. In this regime, the predicted output state X^(t)∈RNdof\hat{\mathbf{X}}(t) \in\mathbb{R}^{N_{\mathrm{dof}}} is the input X(t)∈RNdof\mathbf{X}(t) \in\mathbb{R}^{N_{\mathrm{dof}}} to obtain X^(t+Δt)\hat{\mathbf{X}}(t+\Delta t) in the next iteration step of the RCM. Here, NdofN_{\mathrm{dof}} is the number of degrees of freedom.

Closed-loop reservoir computing models

Figure 1. Closed-loop reservoir computing models. The output data vector of the reservoir computing model at time t+Δtt+\Delta t is the programmable input for the next time step. Both models consist of an input layer WinW_{\mathrm{in}} (orange), a reservoir with memory (light blue) and an optimized output layer Wout∗W_{\mathrm{out}}^{*} (dark blue). a) Classical reservoir computing model (C-RCM). The reservoir is a sparse network of neurons. b) Bright scattershot boson sampling (BSBS) reservoir computing model. Spontaneous parametric down-conversion photon sources create NsrcN_{\mathrm{src}} TMSV states. One mode of each TMSV state is directly detected which heralds the number of photons in the second mode of the TMSV that enter different channels of our linear network. The input unitary UinU_{\mathrm{in}} encodes the dynamical system state X(t)\mathbf{X}(t) at time step tt (here the past reservoir output) via a self-adjoint Hamiltonian. In combination with a random unitary Urand[ϕ]U_{\mathrm{rand}}[\phi] a linear optical network Utotal=Urand[ϕ]UinU_{\mathrm{total}}=U_{\mathrm{rand}}[\phi]U_{\mathrm{in}} of phase shifters and beam splitters processes the input. The green box therefore represents the programmable interferometer, which implements UtotalU_{\mathrm{total}}. The photon number detection at network exit, which includes all measured detection events, generates a high-dimensional photon number distribution with probabilities. A subsequent reduction layer is inserted to downscale the high-dimensional GBS output vector by taking photon number correlations at the 2Nsrc+Nvac2N_{\mathrm{src}}+N_{\mathrm{vac}} modes and determine photon number moments ⟨ni⟩\langle n_i\rangle, ⟨ninj⟩\langle n_i n_j\rangle, and in some cases even ⟨ninjnk⟩\langle n_i n_j n_k\rangle. See also Table I for details. The reduced data vector r′(t+Δt)\mathbf{r}'(t+\Delta t) enters the optimized linear output layer Wout∗W_{\mathrm{out}}^{*} of the reservoir after blending with the past state r(t)\mathbf{r}(t) by leaking rate ε\varepsilon. This leads to the dynamical system prediction X^(t+Δt)\hat{\mathbf{X}}(t+\Delta t) which is the input for the next processing step.

A limiting case of reservoir computing models are extreme learning machines, algorithms that have no short-term memory and consist of a single hidden layer in the classical configuration [47]. In an extreme learning machine, the quantum reservoir state Ψ\mathbf{\Psi} at time t+Δtt+\Delta t is solely determined by the classical data input at time tt. This class includes nonlinear vector autoregression algorithm, also termed next generation reservoir computing [48, 49]. Qubit-based implementations of extreme learning machines have been discussed in [50, 51], and a small scale proof-of-principle demonstration with photonic systems in ref. [52].

Classical and quantum reservoir computing models R\mathcal{R} follow an update rule with short-term memory for the quantum reservoir state, which is given by

Ψ(t+Δt)=R[Ψ(t),X(t)].(1)\mathbf{\Psi}(t+\Delta t)=\mathcal{R}[\mathbf{\Psi}(t),\mathbf{X}(t)]. \tag*{(1)}

This enables us to model nonlinear chaotic dynamics such as turbulence, that is always characterized by finite-time correlations [30, 45, 53, 54, 55, 56, 57, 58]. Figure 1 (a) shows a classical reservoir computing model, contrasted with our quantum RCM shown in panel (b). In our model, the time protocol is a so-called feedback-driven protocol [30, 56, 58]. A measurement at time step tt determines the next step at t+Δtt+\Delta t; the state of the quantum reservoir is destroyed (since we apply a projective measurement), but information from these measurements is fed back into the quantum RCM cycle. Other time protocols are discussed in [53, 59, 60].

We design our approach to be compatible with the Paderborn Quantum Sampler, an academic Boson Sampling machine [61]. The pipeline of the data processing in the quantum photonic experiment is illustrated in Figure 1(b). The steps from photon generation, encoding of inputs in the interferometer, photon detection, reduction to correlation functions, reservoir step, and finally prediction of the dynamical quantities in the next time step are described in the following sections step by step.

Photon generation

The starting point is NsrcN_{\mathrm{src}} photon sources that generate single spectro-temporal TMSV states through spontaneous parametric down conversion (SPDC) [62], indicated as red spots in Figure 1(b). The SPDC process nondeterministically converts one or more photons of high energy into one or more pairs of photons of lower energy, known as signal and idler; measuring one mode of the downconverted fields heralds the presence of photons in the other field, which enters a programmable interferometer implementing a unitary transformation. Each state is given by

∣ψ⟩SPDC=1−λ2∑n=0∞λn∣n,n⟩(2)\left|\psi\right\rangle_{\mathrm{SPDC}}=\sqrt{1-\lambda^2}\sum_{n=0}^{\infty}\lambda^n\left|n,n\right\rangle \tag*{(2)}

which describes the creation of nn photons in each of two modes, parameterised by 0≤λ<10\leq\lambda<1. This is related to the mean photon number generated by the source nˉ=λ2/(1−λ2)\bar{n}=\lambda^2/(1-\lambda^2). It has been shown that one can achieve mean photon numbers of up to O(104)O(10^4) [62], corresponding to an upper bound λ>0.9995\lambda>0.9995. The joint photon number distribution P(n)P(n) of a certain photon number being spontaneously emitted in each mode is

P(n,λ)=(1−λ2)λ2n.P(n,\lambda)=(1-\lambda^2)\lambda^{2n}.

Encoding of inputs

The photons enter the programmable interferometer as a Fock state ∣Ψin⟩=∣n1,n2,…,nNsrc,0,0,…,0⟩\left|\Psi_{\mathrm{in}}\right\rangle=\left|n_1,n_2,\ldots,n_{N_{\mathrm{src}}},0,0,\ldots,0\right\rangle into which the data vector X\mathbf{X} is encoded. The unitary matrix UtotalU_{\mathrm{total}} of the interferometer is formed as a product of a randomly chosen unitary matrix Urand[ϕ]U_{\mathrm{rand}}[\phi], which can be seen as part of the reservoir, and an encoding unitary matrix Uin[X(t)]U_{\mathrm{in}}[\mathbf{X}(t)] as

Utotal(t)=Urand[ϕ]Uin[X(t)]∈U(Nsrc+Nvac).(3)U_{\mathrm{total}}(t)=U_{\mathrm{rand}}[\phi]U_{\mathrm{in}}[\mathbf{X}(t)]\in U(N_{\mathrm{src}}+N_{\mathrm{vac}}). \tag*{(3)}

For each time step tt, a different interferometer matrix Utotal(t)U_{\mathrm{total}}(t) is programmed into the interferometer after performing the measurement, i.e., a reconfigurable optical chip comprising a linear optics network composed of beam splitters and phase shifters where angles and splitting ratios at the phase shifters and beam splitters, respectively can be chosen to implement the given unitary matrix, see Figure 1(b). The random reservoir-matrix can be parametrized by a real vector ϕ∈R(Nsrc+Nvac)2\phi\in\mathbb{R}^{(N_{\mathrm{src}}+N_{\mathrm{vac}})^2} which is kept fixed. The inputs X(t)\mathbf{X}(t) are encoded into the unitary encoding matrix Uin(t)U_{\mathrm{in}}(t) via a Hermitian matrix HH as

Uin(t)=exp⁡(iH[X(t)])(4)U_{\mathrm{in}}(t) = \exp(iH[\mathbf{X}(t)]) \tag*{(4)}

with the dynamical variables on the off-diagonals of the Hermitian matrix HH, e.g. for four dynamical inputs X(t)=(X1(t),X2(t),X3(t),X4(t))\mathbf{X}(t) = (X_1(t), X_2(t), X_3(t), X_4(t)) in a four-dimensional unitary

H[X(t)]=14+α(0X1(t)X2(t)X3(t)X1(t)0X4(t)0X2(t)X4(t)00X3(t)000).(5)H[\mathbf{X}(t)] = \mathbb{1}_4 + \alpha \begin{pmatrix} 0 & X_1(t) & X_2(t) & X_3(t) \\ X_1(t) & 0 & X_4(t) & 0 \\ X_2(t) & X_4(t) & 0 & 0 \\ X_3(t) & 0 & 0 & 0 \end{pmatrix}. \tag*{(5)}

The parameter α\alpha can be used to adjust to the dynamic range of the dynamical variables and is kept fixed. Throughout this work, we have used α=0.01\alpha= 0.01. Even though the encoding matrix UinU_{\mathrm{in}} has a rather regular structure due to the underlying structure of the Hermitian matrix HH, the product with the random reservoir matrix Urand[ϕ]U_{\mathrm{rand}}[\phi] will result in an unstructured and fully occupied interferometer matrix UtotalU_{\mathrm{total}} with high probability.

Photon detection and sampling

The generated photons pass through the programmed interferometer, resulting in the state ∣Ψout⟩|\Psi_{\mathrm{out}}\rangle, as

∣Ψout⟩=U^total(t)∣Ψin⟩,(6)|\Psi_{\mathrm{out}}\rangle= \hat{U}_{\mathrm{total}}(t)|\Psi_{\mathrm{in}}\rangle, \tag*{(6)}

where U^total(t)\hat{U}_{\mathrm{total}}(t) denotes the many-particle operator that corresponds to the action of the interferometer with the unitary matrix Utotal(t)U_{\mathrm{total}}(t). Subsequent photon-number resolving detection results in a Fock state ∣Ψout,i⟩=∣ni,1,ni,2,…,ni,2Nsrc+Nvac⟩|\Psi_{\mathrm{out},i}\rangle= |n_{i,1}, n_{i,2}, \ldots, n_{i,2N_{\mathrm{src}}+N_{\mathrm{vac}}\rangle} with ii being sample number index. Repeating this sampling procedure NsampleN_{\mathrm{sample}} times results in a set of samples {ni,1,ni,2,…}i=1,…,Nsample\{n_{i,1}, n_{i,2}, \ldots\}_{i=1,\ldots,N_{\mathrm{sample}}} of photon counts of the 2Nsrc+Nvac2N_{\mathrm{src}}+N_{\mathrm{vac}} modes. For Nsample→∞N_{\mathrm{sample}} \to\infty, these samples represent the photon number distributions in the state ∣Ψout⟩|\Psi_{\mathrm{out}}\rangle.

Reduction to mode-correlation functions

The measured photon number probability distribution, i.e., the measurement probabilities of the states ∣n1,n2,…,n2Nsrc+Nvac⟩|n_1,n_2,\ldots,n_{2N_{\mathrm{src}}+N_{\mathrm{vac}}}\rangle for the individual interferometer configuration Utotal(t)U_{\mathrm{total}}(t) could now be used in various ways as the inputs for the reservoir memory and linear output layer. However, using the state probabilities directly introduces an optimal selection problem where states would have to be selected in the photon number space such that their probability can, on the one hand, be accurately measured with a low number of samples, and on the other hand is sufficiently dynamically dependent on the dynamical quantities X(t)\mathbf{X}(t) via the interferometer unitary Utotal(t)U_{\mathrm{total}}(t). Even though experimentally available detectors are limited to distinguishing nmax⁡n_{\max} photons at each detector, for example 3 or 4 photons in the case of superconducting nanowire single-photon detectors [63], the available photon number space grows exponentially with the number of interferometer modes. To avoid this selection problem, instead the samples are reduced to first-, second- and higher-order moments of the particle number, i.e., the expectation values

⟨n^in^j…⟩=∑a=1Nsample⟨na,1,na,2,…∣n^in^j…∣na,1,na,2,…⟩=∑a=1Nsamplena,ina,j…(7)\begin{aligned} \langle\hat{n}_i\hat{n}_j\ldots\rangle= \sum_{a=1}^{N_{\mathrm{sample}}} \langle n_{a,1}, n_{a,2}, \ldots|\hat{n}_i\hat{n}_j\ldots|n_{a,1}, n_{a,2}, \ldots\rangle\\ &= \sum_{a=1}^{N_{\mathrm{sample}}} n_{a,i}n_{a,j}\ldots \tag*{(7)} \end{aligned}

or, in other words, photon number expectation values ⟨n^i⟩\langle\hat{n}_i\rangle and multi-mode correlation functions ⟨n^in^j…⟩\langle\hat{n}_i\hat{n}_j\ldots\rangle. The maximum photon-number moment OmO_m, i.e., maximal number of terms in the products na,ina,j…n_{a,i}n_{a,j}\ldots determines the total number of expectation values NobsN_{\mathrm{obs}}, which is given by

Nobs=∑k=1Om(N~+k−1)!k!(N~−1)! with N~=2Nsrc+Nvac(8)N_{\mathrm{obs}} = \sum_{k=1}^{O_m} \frac{(\widetilde{N}+k-1)!}{k!(\widetilde{N}-1)!}\text{ with }\widetilde{N}=2N_{\mathrm{src}}+N_{\mathrm{vac}} \tag*{(8)}

for N~=2Nsrc+Nvac\widetilde{N}=2N_{\mathrm{src}}+N_{\mathrm{vac}} modes because photon number operators commute. Hence, the number of expectation values grows like O(N~Om)O(\widetilde{N}^{O_m}) for Om≪N~O_m \ll\widetilde{N} instead of O((nmax⁡+1)N~)O((n_{\max}+1)^{\widetilde{N}}) if using the state probabilities. In this work, Om=2O_m=2 and Om=3O_m=3 are examined, and Table I lists the corresponding counts for the investigated examples.

LabelNmodN_{\mathrm{mod}}NsrcN_{\mathrm{src}}NvacN_{\mathrm{vac}}OmO_mNobsN_{\mathrm{obs}}
L8a / L96a844290
L8b / L96b8443454
L8c / L96c12662189
L8d / L96d126631329
L96e168832924

Table I. TABLE I. List of runs. Each label encodes the dynamical system, either the extended Lorenz-type model of thermal convection (L8) or the Lorenz 96 atmospheric circulation model (L96). Configurations (a–e) of BSBS-RCM setup differ in the number of interferometer modes NmodN_{\mathrm{mod}}, the number of photon sources NsrcN_{\mathrm{src}}, the number of vacuum entries Nvac=Nmod−NsrcN_{\mathrm{vac}}=N_{\mathrm{mod}}-N_{\mathrm{src}}, maximum photon number moment order OmO_m, and the number of observables NobsN_{\mathrm{obs}}.

Memory and output reduction layer

After the reduction of the sampled photon number measurements to expectation values of moments of the particle number, the vector of moments (here Om=2O_m = 2),

r′(t+Δt)=(⟨n^1⟩,⟨n^2⟩,…,⟨n^N~⟩,⟨n^1n^1⟩,⟨n^1n^2⟩,…,⟨n^N~n^N~⟩)T(9)\mathbf{r}'(t+\Delta t)=\left(\langle\hat{n}_1\rangle,\langle\hat{n}_2\rangle,\ldots,\langle\hat{n}_{\widetilde{N}}\rangle,\langle\hat{n}_1\hat{n}_1\rangle,\langle\hat{n}_1\hat{n}_2\rangle,\ldots,\langle\hat{n}_{\widetilde{N}}\hat{n}_{\widetilde{N}}\rangle\right)^T \tag*{(9)}

can serve as the nonlinear activation vector r′(t+Δt)\mathbf{r}'(t+\Delta t) for the next time step. Our photonic reservoir computing model (BSBS-RCM) will be a hybrid quantum-classical algorithm since the memory is implemented classically outside the quantum photonic device in the next step of the data processing pipeline [30, 56]. The reservoir step at time t+Δtt+\Delta t is obtained by

r(t+Δt)=(1−ε)r(t)+εr′(t+Δt).(10)\mathbf{r}(t+\Delta t)=(1-\varepsilon)\mathbf{r}(t)+\varepsilon\mathbf{r}'(t+\Delta t). \tag*{(10)}

The first term stands for a linear memory and the second term for the nonlinear activation. Note that the nonlinear activation follows from the encoding via the Hamiltonian [30, 53]. The leaking rate ε\varepsilon blends between both contributions [64]. For ε=1\varepsilon=1, one obtains again an extreme learning machine. In our case, this reduced information will re-enter the BSBS-RCM, similar to refs. [56, 58]. Finally, the reservoir state r(t+Δt)\mathbf{r}(t+\Delta t) is fed into the optimized output layer Wout∗W_{\mathrm{out}}^* to give the state vector of the dynamical system at the next time step,

X^(t+Δt)=Wout∗r(t+Δt).(11)\hat{\mathbf{X}}(t+\Delta t)=W_{\mathrm{out}}^*\mathbf{r}(t+\Delta t). \tag*{(11)}

See again Figure 1 for an illustration of the whole data processing pipeline. Rather than ridge regression, we use the pseudoinverse for most cases, since the multiphoton boson sampler output is still small in comparison to the number of training steps Ntrain/Nout≫1N_{\mathrm{train}}/N_{\mathrm{out}}\gg1, see Appendix A § 1.6.2 for further details.

Estimate of number of operations

Before presenting the prediction performance, it is pertinent to estimate the resources required in classical and photonic RCMs. The classical RCM lives typically in a space inflated by two orders of magnitude; thus we set dim⁡R∼102Ndof\operatorname{dim}R\sim10^2N_{\mathrm{dof}}. The size of the photonic reservoir is estimated as follows: when using NsrcN_{\mathrm{src}} heralded Fock and NvacN_{\mathrm{vac}} additional vacuum states of photon number up to nmax⁡n_{\max} then the dimensionality becomes dim⁡R∼(nmax⁡+1)2Nsrc+Nvac\operatorname{dim}R\sim(n_{\max}+1)^{2N_{\mathrm{src}}+N_{\mathrm{vac}}}. Single-photon sources (nmax⁡=1n_{\max}=1), as mostly used, would lead to a qubit-like exponential growth of the reservoir dimension, 22Nsrc+Nvac2^{2N_{\mathrm{src}}+N_{\mathrm{vac}}}. Multiphoton systems can grow significantly faster in their dimensionality.

The loading into a classical reservoir requires O(102Ndof2)O(10^2N_{\mathrm{dof}}^2) operations. The advantage of BSBS-RCM is the potential dimensionality that can be encoded. With a Ndof×Ndof\sqrt{N_{\mathrm{dof}}}\times\sqrt{N_{\mathrm{dof}}} unitary UinU_{\mathrm{in}}, we can encode NdofN_{\mathrm{dof}} classical values. Thus we need O(Ndof)O(N_{\mathrm{dof}}) operations. The number of shots NsN_s in case of BSBS-RCM remains somewhat uncertain. However, the repetition rates of experimental BSBS-RCMs in the 10 MHz regime will cause an advantage in the photonic case, even though it is not a universal quantum computer.

On the computational complexity, QML is a paradigm for data-based problem solving and approximation; it is not a computational problem in itself and therefore does not live in a specific complexity class. We do not make a statement of complexity theoretic advantage, rather a practical advantage on performing/encoding the data-processing algorithm for RCM, using a programmable quantum sampling machine—a concept often referred to as quantum utility [65]. Nevertheless, since BSBS-RCM is based on boson sampling, we know it can be quantum complex [1].

Table II summarizes our estimates. We see that both approaches scale with Ndof2N_{\mathrm{dof}}^2 for the query complexity.

C-RCMBSBS-RCM
Input vector sizeNdofN_{\mathrm{dof}}NdofN_{\mathrm{dof}}
Reservoir dimension dim⁡R\dim_{\mathrm{R}}102Ndof10^{2}N_{\mathrm{dof}}(nmax⁡+1)2Nsrc+Nvac(n_{\max}+1)^{2N_{\mathrm{src}}+N_{\mathrm{vac}}}
Number of “neurons”102Ndof10^{2}N_{\mathrm{dof}}2Ndof2N_{\mathrm{dof}}
Input operationsNdof×102NdofN_{\mathrm{dof}}\times10^{2}N_{\mathrm{dof}}NdofN_{\mathrm{dof}}
Reservoir operations102Ndof×102Ndof10^{2}N_{\mathrm{dof}}\times10^{2}N_{\mathrm{dof}}NdofN_{\mathrm{dof}}
Output operationsNdof×102NdofN_{\mathrm{dof}}\times10^{2}N_{\mathrm{dof}}NsN_{\mathrm{s}}
Gate complexity CGC_{\mathrm{G}}—O(Ndof2)\mathcal{O}(N_{\mathrm{dof}}^{2})
Number of shots NsN_{\mathrm{s}}111/ϵ21/\epsilon^{2}
Query complexity CQ=CG×NsC_{\mathrm{Q}}=C_{\mathrm{G}}\times N_{\mathrm{s}}O(104Ndof2)\mathcal{O}(10^{4}N_{\mathrm{dof}}^{2})O(Ndof2Ns)\mathcal{O}(N_{\mathrm{dof}}^{2}N_{\mathrm{s}})

Table II. List of the estimates for the (model) operations for one integration step of the classical dynamical system. We present complexity estimates for classical reservoir computing model (C-RCM) and the BSBS-based photonic reservoir computing model (BSBS-RCM). Here NsrcN_{\mathrm{src}} is the number of photon sources, NvacN_{\mathrm{vac}} the number of vacuum modes. We denote by NsN_{\mathrm{s}} the number of shots or repetitions. The desired error threshold ϵ∼1/Ns\epsilon\sim1/\sqrt{N_{\mathrm{s}}}.

Results

Prediction of thermal convection flow

We present the results for two prediction tasks of complex, chaotic fluid flows. The first task is for the two-dimensional thermal convection flow, which is described as a dynamical system in the form of a truncated Galerkin model [66]. The number of degrees of freedom in this model is Ndof=8N_{\mathrm{dof}}=8 [30, 67]; it is thus an extended Lorenz-type model based on the classical work [68] and will be termed L8 model in the following. The second prediction task is for a minimal model of global chaotic atmospheric circulation, the Lorenz 96 model [69, 70]. This model contains all ingredients of the Navier-Stokes equations which typically describe turbulent fluid motion and thus geophysical turbulence: nonlinearity, external forcing, dissipation, spatial coupling, and chaos. The number of degrees of freedom is now Ndof=16N_{\mathrm{dof}}=16; it will be termed L96 model.

In thermal convection flows, buoyancy forces drive the complex fluid motion. The present system consists of a two-dimensional fluid layer, which is heated uniformly from below with a dimensionless rescaled temperature T=1T=1 and cooled from above with a dimensionless T=0T=0, thus giving an imposed temperature difference ΔT>0\Delta T>0. Once this thermal driving exceeds a threshold, which is expressed by the critical Rayleigh number parameter, fluid motion in the form of a pair of counter-rotating circulation rolls is initiated by a linear instability. Hot lighter fluid ascends, while cold heavier fluid descends, as shown in Figure 2(a,Figure 2c). For sufficiently high Rayleigh number, the fluid motion becomes time-dependent and finally turbulent. The L8 model is given in compact form by

Prediction of the temperature in cuts through the convection layer

Figure 2. Prediction of the temperature in cuts through the convection layer. (a) Convection flow snapshot at dimensionless time λ1t=4.38\lambda_1 t = 4.38; λ1\lambda_1 is the largest Lyapunov exponent of the dynamical system. (b) Principal sketch of the convection layer. (c) Further snapshot at λ1t=7.11\lambda_1 t = 7.11. We display the time evolution of the temperature field TT along two xx-axes at z/H=0.3z/H = 0.3 and 0.70.7, obtained from the prediction of the BSBS-RCM for the L8 model and compared to the ground truth (GT). Panels (d) and (e) show the GT of temperature TT in a space-time plot. Panels (f) and (g) show the corresponding prediction T^\hat{T} from the multiphoton boson sampler (pred). Panels (h) and (i) display the difference between ground truth and prediction, ∣T−T^∣\lvert T - \hat{T}\rvert. The color bar in the bottom row is given in logarithmic units. The example is taken for a particular quantum reservoir seed with ϵ=0.02\epsilon= 0.02, Nmod=8N_{\mathrm{mod}} = 8, Nsrc=4N_{\mathrm{src}} = 4, and Om=2O_{\mathrm{m}} = 2 and gives a prediction horizon of λ1Tp=5.9\lambda_1 T_{\mathrm{p}} = 5.9, which is indicated by the thick white dashed line in the space-time plots. The solid vertical lines in (d) and (e) indicate the times of the snapshots in (a) and (c), respectively.

dXi(τ)dτ=∑j,k=1NdofAijkXj(τ)Xk(τ)+∑j=1NdofBijXj(τ).(12)\frac{\mathrm{d}X_i(\tau)}{\mathrm{d}\tau} = \sum_{j,k=1}^{N_{\mathrm{dof}}} A_{ijk}X_j(\tau)X_k(\tau) + \sum_{j=1}^{N_{\mathrm{dof}}} B_{ij}X_j(\tau). \tag*{(12)}

This system of coupled ordinary differential equations is obtained by a Fourier mode expansion of velocity and temperature fields and subsequent truncation of the nonlinear partial differential Boussinesq equations that model a thermal convection flow. Here, AijkA_{ijk} and BijB_{ij} are constant coefficients. It contains the quadratic nonlinearity that is characteristic for complex fluid flow problems. The ground truth is obtained by a numerical integration

of (12) with a fourth-order Runge-Kutta scheme. The principal sketch of the layer geometry is shown in panel (b) of Figure 2. The full set of equations and its derivation are given in Appendix B.

Figure 2(d–i) displays the predictive capability of the BSBS-RCM for one example run. We show the space-time evolution of temperature field T(x,t)T(x,t) at two fixed vertical locations: z=0.3z = 0.3 in the bottom half and z=0.7z = 0.7 in the top half as a space-time contour plot. Both predictions are obtained from a typical BSBS-RCM realization L8a (see again Table I). The predicted temperature fields T^(x,t)\hat{T}(x,t) at z=0.3z = 0.3 and 0.70.7 closely reproduce the ground truth T(x,t)T(x,t) at first, very well capturing the spatial tial structure and temporal quasi-periodicity of the convective plumes. This includes the varying tilting of the thermal plumes—lumps of hot and cold fluid that detach from the bottom and top wall, respectively—in connection with a local shear flow motion of the circulation rolls, visible in panels (a) and (c) of Figure 2.

As a first performance metric, we take the normalized mean square error (NMSE), given by

NMSE⁡(τ)=∑t′=1τ∣X(t′)−X^(t′)∣2∑t′=1τX2(t′).(13)\operatorname{NMSE}(\tau)=\frac{\sum_{t'=1}^{\tau}\left|X(t')-\hat{X}(t')\right|^{2}}{\sum_{t'=1}^{\tau}X^{2}(t')}. \tag*{(13)}

Recall that we denote the BSBS-RCM prediction by X^\hat{X} and the ground truth by XX. In the closed-loop scenario, the predictive capability of the reservoir is characterised by the valid prediction time or prediction horizon TpT_{\mathrm{p}} rather than by an NMSE [71]. The prediction horizon therefore defines a second metric and measures how accurately the evolution of characteristic structures of the fluid flow can be predicted. To evaluate this prediction horizon, one uses the time-dependent NMSE⁡(t)\operatorname{NMSE}(t) from (13) and defines

λ1Tp=max⁡{λ1τ∣NMSE⁡(λ1t)<10% ∀t<τ}.(14)\lambda_{1}T_{\mathrm{p}}=\max\{\lambda_{1}\tau\mid\operatorname{NMSE}(\lambda_{1}t)<10\%\ \forall t<\tau\}. \tag*{(14)}

Times are always measured in units of the maximal Lyapunov exponent λ1\lambda_{1} of the dynamical system, which carries the physical dimension of an inverse time and is given for the present set of system parameters, that is a Prandtl number σ=10\sigma=10, a relative Rayleigh number r=28r=28, and a rescaled aspect ratio b=8/3b=8/3 by λ1=0.89\lambda_{1}=0.89 [72]. The largest Lyapunov exponent quantifies the chaoticity of the dynamics. The larger the number of degrees of freedom NdofN_{\mathrm{dof}}, the larger typically λ1\lambda_{1} and thus sensitivity of the dynamics with respect to small differences in initial conditions. The parameters (σ,r,b)(\sigma,r,b) are held constant in the present case study; see again Appendix B.

Figure 2 displays the time series for approximately 9 time units. It is clearly visible that BSBS-RCM captures the system trajectory for a time lag of 3 Lyapunov times with a normalized MSE ≲10−2\lesssim10^{-2}. Even when the NMSE is further increased at later times, the overall qualitative dynamics is kept intact. The frequency and amplitude in the time series remain comparable, see again the space-time contour plots of Figure 2(d–Figure 2g). The latter is essential for higher-dimensional turbulent systems with a stronger chaoticity. A step into this direction is done with the second use case of this study.

Prediction of atmospheric circulation

The second prediction task is for the so-called Lorenz 96 model—a minimal model for atmospheric circulation processes on Earth. In L96, the variables Xi(t)X_i(t) are arranged on a one-dimensional ring to represent NdofN_{\mathrm{dof}} sectors of the thin spherical-shell atmosphere layer as illustrated in Figure 3(a). The L96 model is given by the following system of nonlinear ordinary differential equations,

Prediction of global circulation patterns with photonic reservoir computing model

Figure 3. Prediction of global circulation patterns with photonic reservoir computing model. (a) Sketch of decomposition of the global atmosphere into sectors; each sector is represented by one variable XjX_j. Nonlinear coupling to the neighboring sectors generates traveling wave patterns. (b) One example case of the spatio-temporal evolution of the field. The top panel shows the ground truth, the middle panel shows the corresponding prediction by BSBS-RCM, and the bottom row displays the difference between the GT and the prediction. The color bar to the right holds for all three panels and gives the amplitude. The example is taken for a particular quantum reservoir seed at ε=0.04\varepsilon= 0.04, Nmod=12N_{\mathrm{mod}} = 12, Nsrc=6N_{\mathrm{src}} = 6, and Om=3O_{\mathrm{m}} = 3 and gives a prediction horizon of λ1Tp=2.54\lambda_1 T_{\mathrm{p}} = 2.54. (c) Reproduction of the statistical properties of L96. We show the probability density function (PDF) of XjX_j for ground truth and prediction and compare the data with a Gaussian PDF.

dXidt=(Xi+1−Xi−2)Xi−1−Xi+F,(15)\frac{\mathrm{d}X_i}{\mathrm{d}t}=(X_{i+1}-X_{i-2})X_{i-1}-X_i+F, \tag*{(15)}

with FF being a constant forcing term, F=10F=10. Index i=0,…,Ndof−1i=0,\ldots,N_{\mathrm{dof}}-1. The maximum Lyapunov exponent is now given by λ1=2.055\lambda_{1}=2.055. The first term on the right hand side stands for the nonlinear advection, the driver of complex motion. The second linear term represents dissipation and friction effects. We assume periodic boundary conditions, Xi+Ndof=XiX_{i+N_{\mathrm{dof}}}=X_i. Figure 3(b) shows a typical example of L96 dynamics in a space-time contour plot. The upper panel presents the ground truth dynamics, the middle panel the BSBS-RCM prediction. The bottom contour plot of panel (b) displays the difference between both. The prediction horizon λ1Tp\lambda_{1}T_{\mathrm{p}} of the ensemble study (which will be discussed in detail in the next section) is inserted as a dashed yellow line in all panels. The BSBS-RCM can predict more than two Lyapunov times on average before the relative error exceeds the 10% threshold.

Even beyond the prediction horizon, the statistical properties of the system are accurately reproduced by the BSBS-RCM. This is illustrated in Figure 3(c), which shows the probability density functions (PDF) of the variables XiX_i. Although an individual predicted trajectory inevitably diverges from the ground truth after a few Lyapunov times, the underlying statistical properties remain well reproduced. The PDF is taken over the full test time interval.

This observation is further supported by Figure 4, where we compare the power spectra of the ground truth and the RCM predictions for both use cases in doubly logarithmic plots. At the highest frequencies, corresponding to the shortest time scales, neither the classical nor the photonic RCM fully captures the dynamics quantitatively; spectra decay faster with increasing ff. However, qualitatively the decay is reproduced well. In contrast, the lower-frequency and more slowly varying components are reproduced very well. Capturing these large-scale statistical properties and long-time behavior is an important prerequisite for using the model as a surrogate for complex fluid flow dynamics.

Power spectra of ground truth and prediction

Figure 4. Power spectra of ground truth and prediction. (a) L8 model. (b) L96 model. The low-frequency range fits almost perfectly. Characteristically, both models, classical reservoir computing model (C-RCM) and bright scattershot boson sampling (BSBS), the photonic RCM, fail to predict high frequency signal components. Both cases are compared to the ground truth (GT).

Figure 5 shows the time-correlation function ⟨C(τ)⟩dof\langle C(\tau)\rangle_{\mathrm{dof}} averaged over all 16 degrees of freedom of L96, comparing the BSBS and classical RCM (which we term C-RCM) predictions with the ground truth. All three curves decay from unity over a similar timescale and the ground truth exhibits a weak recurrence, visible as a small secondary peak near λ1τ≈2.5\lambda_1 \tau\approx2.5 that recedes and reappears at longer lags. We summarize this behavior with the long-time integral correlation time τint=∫0τmax⁡⟨C(τ)⟩dof dτ\tau_{\mathrm{int}} = \int_{0}^{\tau_{\max}} \langle C(\tau)\rangle_{\mathrm{dof}}\,\mathrm{d}\tau, with λ1τmax⁡=8\lambda_1 \tau_{\max} = 8. BSBS reproduces the integral time closely with τint=0.21±0.07\tau_{\mathrm{int}} = 0.21 \pm0.07 compared to the ground truth value of τint=0.22\tau_{\mathrm{int}} = 0.22. The C-RCM overshoots in magnitude after the initial decay and does not recover the oscillations, yielding an integral time of τint=−0.33±0.18\tau_{\mathrm{int}} = -0.33 \pm0.18. Unlike the power spectral density, which carries the same information in the frequency domain, or the single-time probability distribution, which carries no temporal information at all, ⟨C(τ)⟩dof\langle C(\tau)\rangle_{\mathrm{dof}} probes the dynamical memory of the surrogate model: how strongly a variable’s current state constrains its own future, and whether that memory is structured (e.g. recurrent) rather than purely decaying. The close agreement of BSBS with the ground truth indicates that it reproduces not only the correct aggregate statistics, but also the temporal structure of the attractor’s internal dynamics more faithfully than C-RCM.

Time-correlations of L96 model

Figure 5. Time-correlations of L96 model. The time correlation function ⟨C(τ)⟩dof\langle C(\tau)\rangle_{\mathrm{dof}}, which is an average over the individual correlation function of the degrees of freedom, is plotted versus delay time τ\tau. Shaded bands for BSBS and C-RCM denote variance over reservoir seeds. The long-time integral correlation time is τint=0.22\tau_{\mathrm{int}} = 0.22 for the ground truth (GT, black solid line); τint=0.21±0.07\tau_{\mathrm{int}} = 0.21 \pm0.07 for BSBS (red solid line), and ∣τint∣=0.33±0.18|\tau_{\mathrm{int}}| = 0.33 \pm0.18 for C-RCM (green solid line).

Prediction capability for different reservoir seeds and photon loss rates

We showed the prediction capability of BSBS-RCM for two use-cases. In the following, we still have to investigate the robustness of these results with respect to hyperparameters of the reservoir computing models. In the photonic system, these are the size of the reservoir, which is mainly determined by the number of photon sources, NsrcN_{\mathrm{src}}, the number of interferometer modes, NmodN_{\mathrm{mod}}, the maximum order of photon number moments, OmO_{\mathrm{m}}, and the leaking rate ε\varepsilon. Table I summarizes the parametric studies that were conducted. We used Nsrc=Nmod/2N_{\mathrm{src}} = N_{\mathrm{mod}}/2 for all photonic reservoir model runs.

For each set of parameters of L8 and L96, we run an ensemble of 70 different random seeds of the unitary Urand[ϕ]U_{\mathrm{rand}}[\phi] to calculate means and standard deviations of the prediction horizon. In addition, for each parameter set and each random seed, we ran the model for 50 different leaking rates 0<ε≤10 < \varepsilon\leq1 in increments of Δε=0.02\Delta\varepsilon= 0.02. We verified that the prediction horizon λ1Tp=0\lambda_1 T_p = 0 for ε=0\varepsilon= 0. This sums up to a total of 3500 runs for each BSBS-RCM setup. Each of these runs proceeds through a training phase of 3×1043 \times10^4 time steps Δt=0.002\Delta t = 0.002; the testing phase consists of 1×1041 \times10^4 steps. All runs were executed by simulations on a CPU cluster with 192 CPU cores, applying the Python-based photonic simulation software package Strawberry Fields [73, 74] for photonic circuits in combination with The Walrus library [75]. The photon-number moments were calculated numerically exactly.

We also compare the results of BSBS-RCM with a classical RCM, which went through a comprehensive hyperparameter tuning by a systematic grid search [71, 72]. Details of this optimization are contained in Appendix C.

Figure 6 summarizes the results. Panel (a) shows the prediction horizons for L8. The solid lines represent the ensemble means, while the surrounding shaded regions indicate the corresponding standard deviation. Recall that the prediction horizon is defined as the time at which the NMSE first exceeds 10%. The best performance is obtained for small leaking rates, whereas the prediction horizon decreases with increasing ε\varepsilon. In particular, run L8a achieves the best performance for Nmod=8N_{\mathrm{mod}} = 8, Nsrc=4N_{\mathrm{src}} = 4, and Om=2O_{\mathrm{m}} = 2 at ε=0.02\varepsilon= 0.02. This corresponds to the regime in which the contribution of the linear photonic network to the reservoir update is minimal. For the classical RCM, the prediction horizon is approximately 25% larger, with the optimum slightly shifted to ε=0.06\varepsilon= 0.06. Interestingly, Pfeffer et al. [30] observed a similar dependence on the leaking rate for a qubit-based RCM, with the best prediction performance likewise obtained at small values of ε\varepsilon. It underlines the importance of the linear memory.

Ensemble analysis of photonic reservoir computing model

Figure 6. Ensemble analysis of photonic reservoir computing model. (a) Prediction horizon λ1Tp\lambda_{1}T_{p} for different leaking rates ε\varepsilon and comparison with a classical RCM of comparable size for the L8 model. We used 50 different leaking rates 0<ε≤10 < \varepsilon\le1, starting with ε=0.02\varepsilon= 0.02 in equidistant steps. Each data point is obtained from an ensemble of 70 different reservoir seeds. The solid lines connect the means, the margins indicate the variance of the statistics. The different L8 model runs are also listed in Table I. The black dashed lines in both figures stand for the classical RCM that was optimized by a comprehensive hyperparameter grid search. (b) Same as (a) for the L96 model runs. (c) Comparison of bright scattershot boson sampling BSBS with coherent light run at exactly the same parameter set for L8. (d) Same as (c) for L96. (e) Prediction performance for different photon loss rates that are indicated in the legend. Analysis is for L8. (f) Same as (e) for L96.

Furthermore, we observe that an increase in the number of photon sources, modes, and of the order of the photon number moment does not improve the prediction performance of the L8 case. The performance actually drops significantly for larger reservoir sizes. In the case of L8, the photonic RCM does not outperform a highly optimized classical RCM, the latter of which obtains a mean prediction horizon of almost 8 Lyapunov times. However, BSBS-RCM still captures the dynamics for up to 6 Lyapunov times.

Panel (b) of Figure 6 summarizes the results for L96; now the classical case is clearly outperformed by the photonic one. The mean prediction horizon drops to more than 2.5 Lyapunov times for this higher-dimensional use case, it is by about a factor of 2 better than the classical one. The best ensemble members reach more than 3 Lyapunov times. The higher dimensionality of the prediction task (Ndof=16N_{\mathrm{dof}} = 16) requires a higher-dimensional photonic reservoir and a higher maximum moment order. The best performance is now in line with an increased leaking rate of ε=0.04\varepsilon= 0.04, suggesting a slightly stronger contribution of the boson sampler to the update step. Our results confirm a well-known circumstance of reservoir computing models: the choice of the system parameters is task-specific. This also shows how we can make use of the genuine high dimensionality of the photonic system for complex learning tasks.

Next, we investigated the robustness of the BSBS-RCM performance for different rates of photon loss with our digital twin. They are present in each experiment and thus an intrinsic property of real photonic reservoir computing devices. We have evaluated different loss rates at the inputs and outputs of up to 50% (see also Appendix A). Panels (e) and (f) of Figure 6 summarize our results for model runs L8a and L96d, respectively. While there is a small impact for model L8a, we do not observe any difference for L96d. These results indicate a high degree of resilience to photon loss.

In summary, we can state that our results demonstrate clearly that an application case that is sufficiently high-dimensional outperforms the corresponding classical algorithm even with included errors – a promising route to photonic quantum advantage.

Comparison to runs with coherent light

To further validate the conclusions drawn from the previous subsection, we simulate the system using coherent light, which serves as a heuristic benchmark for a classical analog of the system. To this end, we take the best-performing BSBS-RCM configurations and simulate both use cases with coherent-light input. Figure 7 illustrates the corresponding BSBS-RCM configurations at input to obtain a consistent comparison between both cases. The average photon number in the BSBS case is ⟨n⟩=0.73\langle n\rangle= 0.73 for the modes that enter the interferometer and heralding arms. In the case of coherent light, the mean photon number of the modes that enter the interferometer has to be doubled, ⟨n′⟩=2⟨n⟩\langle n'\rangle= 2\langle n\rangle to provide a fair comparison – i.e. total optical power remains constant, although we do note this also means that double the power enters the interferometer. As in the photonic quantum cases, we use the same number of random seeds and the same set of leaking-rate values.

Comparison of bright scattershot boson sampling and coherent-light photon inputs

Figure 7. Comparison with coherent light. Sketch of the photon input for bright scattershot boson sampling (BSBS) and the experiment with coherent light for consistent comparison.

The corresponding results are shown in panels (c) and (d) of Figure 6. They clearly demonstrate that the reservoir computing model performs better with quantum light than with classical coherent light. This is in particular the case for the higher-dimensional Lorenz 96 model for the atmospheric circulation.

Discussion

The results shown above are based on an exact numerical simulation of the proposed BSBS setup. The simulation uses Gaussian states, which allow an efficient realization of the state generation, action of the interferometer, and losses. At the detection stage, photon-number moments are obtained exactly from expectation values of the Gaussian state. This step includes the intrinsic computational complexity, as it requires the exponentially scaling calculation of Hafnians with dimensions linear in the order of the photon number moments OmO_m.

Compared to the experimental setup, where both training and subsequent prediction would be executed in situ, i.e., directly on the experimental setup, a number of effects are not yet considered in our digital twin simulations: The transition from directly calculated expectation values to sampled photon number moments contributes statistical noise, i.e., shot noise, which could affect the training quality if too few samples are used. The BSBS approach has been designed, such that processing of the samples is straightforward. The reduction of the samples to expectation values can be performed directly, for example, by an FPGA close to the detectors without storing the samples explicitly. Another effect not considered in the simulations above is the saturation of detectors, where, for example, typical superconducting nanowire single-photon detectors can only detect a limited number of photons per event. This saturation behavior does not impact the viability of the proposed setup because it does not require the photon number moments to be precise. Instead, the relation between dynamical inputs X\mathbf{X} and the sampled photon number moments (⟨n^1⟩≈,⟨n^2⟩≈,…,⟨n^1n^1⟩≈,⟨n^1n^2⟩≈,…\langle\hat{n}_1\rangle_{\approx},\langle\hat{n}_2\rangle_{\approx},\ldots,\langle\hat{n}_1\hat{n}_1\rangle_{\approx},\langle\hat{n}_1\hat{n}_2\rangle_{\approx},\ldots) must be sufficiently nonlinear.

In addition to neglecting shot noise and detector saturation, an ideal programmable interferometer was assumed. Real programmable interferometers only have a limited programming fidelity, i.e., the implemented unitary only approximately resembles the desired unitary. However, the proposed scheme does not rely on an exact implementation of the unitary, but the main requirement is that the programming be repeatable with sufficient accuracy and that close unitaries result in close interferometer characteristics, i.e., a continuous mapping. This guarantees that if a unitary occurs during prediction that is similar to a unitary used during training, it will result in a similar interferometer behavior and, thus, a faithful prediction. In general, we expect the proposed approach to be resilient against systematic and repeatable programming inaccuracies of the interferometer but sensitive to time-dependent drifts during training or prediction.

In summary, we have demonstrated that a multiphoton boson sampler, a hybrid quantum-classical quantum machine learning algorithm, is capable of predicting complex fluid flow dynamics. When scaling up the problem size — which implies here to go from the thermal convection flow model to the atmospheric circulation model — we obtained a significantly better performance in comparison to a classical reservoir computing model. We showed in addition that our digital twin of the Paderborn Quantum Sampler is robust with respect to photon loss rates up to 50%. The algorithm on the basis of quantum light outperforms furthermore a comparable configuration with coherent light. These results demonstrate an advantage in using the quantum states considered here rather than coherent states, consistent with the expectation that the former exhibit more complex photon-number correlations. The implementation of the present simulations as a real experiment and a further up-scaling of the problem size are currently underway and will be reported elsewhere.

Acknowledgments

The work of SKR and JS is supported by the European Union (ERC, MesoComp, 101052786). Views and opinions expressed are however those of the author only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. The authors gratefully acknowledge the computing time made available to them on the high-performance computer Otus [76] at the NHR Center Paderborn Center for Parallel Computing (PC2). This center is jointly supported by the Federal Ministry of Research, Technology and Space and the state governments participating in the National High-Performance Computing (NHR) joint funding program (www.nhr-verein.de/en/our-partners). This work has received funding from the German Federal Ministry of Research, Technology and Space within the PhoQuant project (Grant No. 13N16103).

Appendixes

The following appendices provide technical details on the digital twin algorithm for the BSBS-based reservoir computing model (BSBS-RCM) in Appendix A, a derivation of the extended Lorenz-type model from the Navier-Stokes equations including definitions of parameters and scales in Appendix B, and details on the classical reservoir computing model (C-RCM), which was used for comparison, in Appendix C.

Implementation of digital twin

Architectures

Figure 8 shows the model used for the simulation of the programmable boson sampler, see again Figure 1 in main text. From Nsrc+NvacN_{\mathrm{src}} + N_{\mathrm{vac}} vacuum states as inputs, NsrcN_{\mathrm{src}} two-mode squeezing gates with squeezing parameters

ri=asinh⁡(⟨n⟩),(16)r_i = \operatorname{asinh}\left(\sqrt{\langle n\rangle}\right), \tag*{(16)}

with the desired average photon number ⟨n⟩\langle n\rangle per mode, generate correlated pairs of photons. For the evaluations in the work, a value of λ=0.65\lambda= 0.65 in Eq. (2) was used, corresponding to a squeezing parameter r=0.7753r = 0.7753 and average photon number ⟨n⟩=0.7316\langle n\rangle= 0.7316. For each squeezing gate, one output is heralded, and one output is fed into the interferometer. The action of the squeezing gates, loss channels, and interferometer is efficiently simulated using a Gaussian description using the mean and covariance of its quadratures on Nsrc+NmodN_{\mathrm{src}} + N_{\mathrm{mod}} modes. For the photon-number moments ⟨ninj⋯ ⟩\langle n_i n_j \cdots\rangle in Eq. (7), numerically exact expectation values are calculated from the Gaussian state. For the comparisons with classical light in Section III D, the two-mode squeezing gates are replaced by displacement gates to generate coherent light as input for the interferometer. The displacement parameter α=⟨n′⟩\alpha= \sqrt{\langle n'\rangle} is chosen such that the average photon number is ⟨n′⟩=2⟨n⟩=1.46\langle n'\rangle= 2\langle n\rangle= 1.46 so that the optical power is equal in both cases.

Figure 8. Schematic figure of the digital twin. The structure is shown for the simulation for Nsrc=2N_{\mathrm{src}} = 2 photon sources and Nsrc+Nvac=Nmod=4N_{\mathrm{src}} + N_{\mathrm{vac}} = N_{\mathrm{mod}} = 4 interferometer modes. Panel (a) of the figure depicts the programmable boson sampler in Figure 1 (BSBS setup) with the two-mode squeezing gates S2, interferometer, and loss channels. Panel (b) shows the classical analogous setup using coherent light generated by the displacement gates D.

Training of photonic reservoir computing model

The readout weights in the BSBS-RCM are obtained via the Moore–Penrose pseudo-inverse of the state matrix, an estimator that is exact but numerically reliable only when the matrix is well-conditioned. At smaller encoding scales (L96, Nmod=8/12N_{\mathrm{mod}} = 8/12), the state matrix remains well-conditioned and the pseudo-inverse trains reliably. At the largest configuration (L96, Nmod=16N_{\mathrm{mod}} = 16), the state matrix is rank-deficient by roughly two orders of magnitude, a property of the encoding itself, and the pseudo-inverse becomes numerically unstable, collapsing the trained model’s prediction capability. Writing the singular value decomposition of the state matrix as Rtrain=UΣVTR_{\mathrm{train}} = U\Sigma V^{\mathsf{T}}, this instability is removed by an SVD-based ridge estimator, Wout∗=Vdiag⁡(σiσi2+γ)UTYtrainW_{\mathrm{out}}^{*} = V\operatorname{diag}\left(\frac{\sigma_i}{\sigma_i^2+\gamma}\right)U^{\mathsf{T}}Y_{\mathrm{train}}, where σi\sigma_i are the singular values of RtrainR_{\mathrm{train}} (the diagonal entries of Σ\Sigma) and YtrainY_{\mathrm{train}} is the target, with a single regularization strength γ=10−9\gamma= 10^{-9} fixed across the unitary seeds, which is used only for this configuration. The results from smaller configurations are obtained by using the pseudo-inverse.

Derivation of the model for thermal convection flow

Details of the thermal convection flow model are provided in the following. The flow domain is A=[0,22]×[0,1]A = [0, 2\sqrt{2}] \times[0, 1] in units of the height of the layer HH as shown in Figure 2(b). The horizontal coordinate is xx, the vertical one is zz. For small temperature differences between bottom and top, heat is carried by diffusion only from the bottom to the top while the fluid is at rest. The resulting temperature profile is linear, T‾(z)=Tc−ΔTHz\overline{T}(z) = T_c - \frac{\Delta T}{H}z. This is the equilibrium configuration of the convection flow. The velocity field has two components u(x,z,t)=(ux(x,z,t),uz(x,z,t))\boldsymbol{u}(x,z,t) = (u_x(x,z,t),u_z(x,z,t)). For the following, we consider the temperature deviation θ(x,z,t)=T(x,z,t)−T‾(z)\theta(x,z,t) = T(x,z,t) - \overline{T}(z). The temperature deviation and velocity fields, θ(x,z,t)\theta(x,z,t) and u(x,z,t)\boldsymbol{u}(x,z,t), are coupled by the Boussinesq equations [77]. They are given by

∇⋅u=0,(17)\nabla\cdot\boldsymbol{u} = 0, \tag*{(17)}
∂u∂t+(u⋅∇)u=−∇p+σ∇2u+Ra⁡σθez,(18)\frac{\partial\boldsymbol{u}}{\partial t} + (\boldsymbol{u}\cdot\nabla)\boldsymbol{u} = -\nabla p + \sigma\nabla^2\boldsymbol{u} + \operatorname{Ra}\sigma\theta\boldsymbol{e}_z, \tag*{(18)}
∂θ∂t+(u⋅∇)θ=∇2θ+uz(19)\frac{\partial\theta}{\partial t} + (\boldsymbol{u}\cdot\nabla)\theta= \nabla^2\theta+ u_z \tag*{(19)}

where we use the kinematic pressure pp. The Rayleigh number Ra⁡\operatorname{Ra} and the Prandtl number σ\sigma are the two parameters that characterize the strength of the thermal driving via the temperature difference ΔT\Delta T and the ratio of momentum to temperature diffusion, respectively. The boundary conditions in xx-direction are periodic. At z=0,1z = 0,1, one takes

uz∣z=0,1=0,∂ux∂z∣z=0,1=0andθ∣z=0,1=0.(20)u_z\big|_{z=0,1}=0,\qquad\left.\frac{\partial u_x}{\partial z}\right|_{z=0,1}=0\qquad\text{and}\qquad\theta\big|_{z=0,1}=0. \tag*{(20)}

They correspond to isothermal, impermeable, free-slip walls. Incompressibility and two-dimensionality allow to reduce the velocity vector field further to a scalar stream function ζ(x,z,t)\zeta(x,z,t) by ux=−∂ζ∂zu_x = -\frac{\partial\zeta}{\partial z} and uz=∂ζ∂xu_z = \frac{\partial\zeta}{\partial x}. This ansatz satisfies (B1) automatically and the equations of motion (B2)–(B3) are now given by two scalar functions only,

∂∇2ζ∂t=∂ζ∂z∂∇2ζ∂x−∂ζ∂x∂∇2ζ∂z+σ∇4ζ+Ra⁡σ∂θ∂x,(21)\frac{\partial\nabla^2\zeta}{\partial t} = \frac{\partial\zeta}{\partial z}\frac{\partial\nabla^2\zeta}{\partial x} - \frac{\partial\zeta}{\partial x}\frac{\partial\nabla^2\zeta}{\partial z} + \sigma\nabla^4\zeta+ \operatorname{Ra}\sigma\frac{\partial\theta}{\partial x}, \tag*{(21)}
∂θ∂t=∂ζ∂x∂θ∂z−∂ζ∂z∂θ∂x+∇2θ+∂ζ∂x,(22)\frac{\partial\theta}{\partial t} = \frac{\partial\zeta}{\partial x}\frac{\partial\theta}{\partial z} - \frac{\partial\zeta}{\partial z}\frac{\partial\theta}{\partial x} + \nabla^2\theta+ \frac{\partial\zeta}{\partial x}, \tag*{(22)}

with boundary conditions in the vertical direction

ζ∣z=0,1=0,∂2ζ∂z2∣z=0,1=0andθ∣z=0,1=0.(23)\zeta\big|_{z=0,1}=0,\qquad\left.\frac{\partial^2\zeta}{\partial z^2}\right|_{z=0,1}=0\qquad\text{and}\qquad\theta\big|_{z=0,1}=0. \tag*{(23)}

The nonlinear coupled partial differential equations (PDE) (B5) and (B6) are then reduced by an expansion into trigonometric Fourier modes, which satisfy the boundary conditions for the stream function and temperature, both in the horizontal and vertical directions. The Fourier modes encode the possible spatial structure, and thus reduce the original system of nonlinear PDEs into an infinite system of nonlinear coupled ordinary differential equations. The finite dimension of the model (or the finite number of degrees of freedom) follows by a truncation of this set. This is the basic idea of a Galerkin model [66]. Here, we take four modes for the expansions of temperature deviation θ\theta and four modes for the stream function ζ\zeta, which leads to eight time-dependent amplitudes X0(τ)X_0(\tau) to X7(τ)X_7(\tau), see also refs. [30, 72].

The following truncated expansion with a total of eight Fourier modes, all of which satisfy the boundary conditions (B7), is used to arrive at the truncated Galerkin model,

ζ(x,z,t)=Cζ[X0(t)sin⁡(αx)sin⁡(βz)+X3(t)sin⁡(βz)+X4(t)cos⁡(αx)sin⁡(2βz)+X6(t)sin⁡(3βz)]\zeta(x,z,t)=C_{\zeta}[X_0(t)\sin(\alpha x)\sin(\beta z)+X_3(t)\sin(\beta z)+X_4(t)\cos(\alpha x)\sin(2\beta z)+X_6(t)\sin(3\beta z)]
θ(x,z,t)=Cθ[2X1(t)cos⁡(αx)sin⁡(βz)−X2(t)sin⁡(2βz)+X5(t)sin⁡(αx)sin⁡(2βz)+X7(t)sin⁡(4βz)]\theta(x,z,t)=C_{\theta}[\sqrt{2}X_1(t)\cos(\alpha x)\sin(\beta z)-X_2(t)\sin(2\beta z)+X_5(t)\sin(\alpha x)\sin(2\beta z)+X_7(t)\sin(4\beta z)]

with the normalization constants

Cζ=2α2+β2αβC_{\zeta}=\sqrt{2}\frac{\alpha^2+\beta^2}{\alpha\beta}
Cθ=(α2+β2)3α2βRa⁡,C_{\theta}=\frac{(\alpha^2+\beta^2)^3}{\alpha^2\beta\operatorname{Ra}},

where α\alpha and β\beta are the horizontal and vertical wavenumbers, respectively, representing the flow in the fluid layer described by the resulting Lorenz-type model.

The Lorenz-type model equations for 8 degrees of freedom X0X_0, X1X_1, ..., X7X_7, the L8 model, are given by

dX0dτ=σ(X1−X0)−c32c1X3X4+3c−52c1X4X6(24)\begin{aligned} \frac{dX_0}{d\tau} &= \sigma\left(X_1-X_0\right)-\frac{c_3}{\sqrt{2c_1}}X_3X_4+\frac{3c_{-5}}{\sqrt{2c_1}}X_4X_6 \tag*{(24)} \end{aligned}
dX1dτ=−X1+rX0+12X3X5−32X5X6−X0X2(25)\begin{aligned} \frac{dX_1}{d\tau} &= -X_1+rX_0+\frac{1}{2}X_3X_5-\frac{3}{2}X_5X_6-X_0X_2 \tag*{(25)} \end{aligned}
dX2dτ=−bX2+X0X1(26)\frac{dX_2}{d\tau}=-bX_2+X_0X_1 \tag*{(26)}
dX3dτ=−σb4X3+322X0X4(27)\frac{dX_3}{d\tau}=-\frac{\sigma b}{4}X_3+\frac{3}{2\sqrt{2}}X_0X_4 \tag*{(27)}
dX4dτ=−c4c1X4−σc12c4X5+c02c4X0X3−3c−8c4X0X6(28)\begin{aligned} \frac{dX_4}{d\tau}={}&-\frac{c_4}{c_1}X_4-\frac{\sigma c_1}{\sqrt{2c_4}}X_5+\frac{c_0}{\sqrt{2c_4}}X_0X_3\\ &-\frac{3c_{-8}}{c_4}X_0X_6 \tag*{(28)} \end{aligned}
dX5dτ=−c4c1X5−2rX4−X3X1+3X6X1−22X4X7(29)\begin{aligned} \frac{dX_5}{d\tau}={}&-\frac{c_4}{c_1}X_5-\sqrt{2}rX_4-X_3X_1+3X_6X_1\\ &-2\sqrt{2}X_4X_7 \tag*{(29)} \end{aligned}
dX6dτ=−9σb4X6−122X0X4(30)\frac{dX_6}{d\tau}=-\frac{9\sigma b}{4}X_6-\frac{1}{2\sqrt{2}}X_0X_4 \tag*{(30)}
dX7dτ=−4bX7+2X4X5.(31)\frac{dX_7}{d\tau}=-4bX_7+\sqrt{2}X_4X_5. \tag*{(31)}

Here, cn=α2+nβ2c_n=\alpha^2+n\beta^2, where n∈Zn\in\mathbb{Z}. The parameters are the Prandtl number σ=10\sigma=10, b=8/3b=8/3, and the relative Rayleigh number r=(Ra−Rac)/Rac=28r=(Ra-Ra_c)/Ra_c=28 with the critical Rayleigh number RacRa_c for the onset of convection. For these dynamical system parameters, α=π/2\alpha=\pi/\sqrt{2} and β=π\beta=\pi. The original time tt of the fluid flow equations is rescaled to τ=c1t\tau=c_1t.

Details of the classical reservoir computing model

In classical reservoir computing model, the reservoir state is updated as follows [72]

r(t+Δt)=(1−ε)r(t)+εtanh⁡[Wr(t)+WinX(t+Δt)],\mathbf{r}(t+\Delta t)=(1-\varepsilon)\mathbf{r}(t)+\varepsilon\tanh[W\mathbf{r}(t)+W_{\mathrm{in}}\mathbf{X}(t+\Delta t)],

where the reservoir matrix WW is characterised by the spectral radius ρ\rho, WinW_{\mathrm{in}} is the input weight matrix, and X\mathbf{X} is the input vector from the prediction task. The output is given by the linear system y=Woutr\mathbf{y}=W_{\mathrm{out}}\mathbf{r}, where WoutW_{\mathrm{out}} is determined by ridge regression with parameter λridge\lambda_{\mathrm{ridge}}.

The optimal hyperparameters for the classical RCM are found in a grid search. This is performed on 11 different seeds of the reservoir matrix, each with the following grid. The spectral radii of the reservoir matrix take values of ρ∈[0,1.3]\rho\in[0,1.3] increasing in steps of 0.1; the ridge regression parameters (to avoid overfitting) take values of λridge∈[10−11,10−2]\lambda_{\mathrm{ridge}}\in[10^{-11},10^{-2}] which are increased in factors of 10; the leaking rate takes values of ε∈[0.01,1]\varepsilon\in[0.01,1] with the increment refined systematically. The number of reservoir nodes, Nr=2048N_r=2048, and the density of active reservoir nodes, D=0.08D=0.08, is kept fixed. Then the maximum prediction horizon irrespective of {ρ,λridge}\{\rho,\lambda_{\mathrm{ridge}}\} is plotted with ε\varepsilon to identify the optimal leaking rate ε=0.7\varepsilon=0.7 for L8 model and ε=0.02\varepsilon=0.02 for L96 model. Subsequently, three potential pairs {ρ,λridge}\{\rho,\lambda_{\mathrm{ridge}}\}, based on their number of occurrences, are selected. The robustness of these sets of hyperparameters is tested against 100 seeds for the reservoir matrix. The results for the C-RCM presented in the main text correspond to the hyperparameter set {ρ,λridge,ε}={0.1,10−5,0.7}\{\rho,\lambda_{\mathrm{ridge}},\varepsilon\}=\{0.1,10^{-5},0.7\} for L8 model and {ρ,λridge,ε}={0.6,10−3,0.02}\{\rho,\lambda_{\mathrm{ridge}},\varepsilon\}=\{0.6,10^{-3},0.02\} for L96 model.

References

References

  1. [1]S. Aaronson and A. Arkhipov, The computational complexity of linear optics, STOC ’11: Proceedings of the forty-third annual ACM Symposium on Theory of Computing , 333 (2011).
  2. [2]A. P. Lund, A. Laing, S. Rahimi-Keshari, T. Rudolph, J. L. O’Brien, and T. C. Ralph, Boson sampling from a Gaussian state, Phys. Rev. Lett. 113, 100502 (2014).
  3. [3]C. S. Hamilton, R. Kruse, L. Sansoni, S. Barkhofen, C. Silberhorn, and I. Jex, Gaussian boson sampling, Phys. Rev. Lett. 119, 170501 (2017).
  4. [4]R. Kruse, C. S. Hamilton, L. Sansoni, S. Barkhofen, C. Silberhorn, and I. Jex, Detailed study of Gaussian boson sampling, Phys. Rev. A 100, 032326 (2019).DOI
  5. [5]H. Wang, J. Qin, X. Ding, M.-C. Chen, S. Chen, X. You, Y.-M. He, X. Jiang, L. You, Z. Wang, C. Schneider, J. J. Renema, S. Höfling, C.-Y. Lu, and J.-W. Pan, Boson sampling with 20 input photons and a 60-mode interferometer in a 10¹⁴-dimensional Hilbert space, Phys. Rev. Lett. 123, 250503 (2019).
  6. [6]H.-S. Zhong, H. Wang, Y.-H. Deng, M.-C. Chen, L.-C. Peng, Y.-H. Luo, J. Qin, D. Wu, X. Ding, Y. Hu, P. Hu, X.-Y. Wang, W.-J. Zhang, H. Li, Y. Li, X. Jiang, L. Gan, G. Yang, L. You, Z. Wang, L. Li, N.-L. Liu, C.-Y. Lu, and J.-W. Pan, Quantum computational advantage using photons, Science 370, 1460 (2020).DOI
  7. [7]L. S. Madsen, F. Laudenbach, M. F. Askarani, F. Rortais, T. Vincent, J. F. F. Bulmer, F. M. Miatto, L. Neuhaus, L. G. Helt, M. J. Collins, A. E. Lita, T. Gerrits, S. W. Nam, V. D. Vaidya, M. Menotti, I. Dhand, Z. Vernon, N. Quesada, and J. Lavoie, Quantum computational advantage with a programmable photonic processor, Nature 606, 75 (2022).DOI
  8. [8]J. Romero and G. Milburn, Photonic quantum computing (2024), arXiv:2404.03367.DOI
  9. [9]Y.-X. Fu, H.-Y. Shen, K.-M. Hu, J.-J. He, Y.-L. Nie, H. Song, B.-J. Liu, L.-S. Yang, X.-Y. Wu, P.-L. Du, Y.-Z. Zhu, Y. Xie, D.-H. Huang, Y.-F. Liu, H. Yan, J.-H. Chen, Y.-L. Yu, C.-Y. Peng, W.-H. Zhou, F. Lu, Y.-Q. Peng, C.-S. Xia, Z.-C. Wang, Z.-H. Li, L. Chen, C.-Y. Zhang, C.-Y. Jin, Y.-R. Zong, X.-Y. Xu, J.-P. Dou, X.-T. Fang, P.-Q. Zhou, H. Tang, C.-S. Wang, L. Yang, and X.-M. Jin, A chip-scale space-time multiplexed gaussian boson sampling processor beyond 10,000 photons (2026), arXiv:2609.11922.arxiv.org/abs/2609.11922
  10. [10]J. Huh, G. G. Guerreschi, B. Peropadre, J. R. McClean, and A. Aspuru-Guzik, Boson sampling for molecular vibronic spectra, Nat. Photon. 9, 615 (2015).DOI
  11. [11]H. H. Zhu, H. Sen Chen, T. Chen, Y. Li, S. B. Luo, M. F. Karim, X. S. Luo, F. Gao, Q. Li, H. Cai, L. K. Chin, L. C. Kwek, B. Nordén, X. D. Zhang, and A. Q. Liu, Large-scale photonic network with squeezed vacuum states for molecular vibronic spectroscopy, Nat. Commun. 15, 6057 (2024).DOI
  12. [12]S. Yu, Z.-P. Zhong, Y. Fang, R. B. Patel, Q.-P. Li, W. Liu, Z. Li, L. Xu, S. Sagona-Stopel, E. Mer, S. E. Thomas, Y. Meng, Z.-P. Li, Y.-Z. Yang, Z.-A. Wang, N.-J. Guo, W.-H. Zhang, G. K. Tranner, Y. Dong, Y.-T. Wang, J.-S. Tang, C.-F. Li, I. A. Walmsley, and G.-C. Guo, A universal programmable Gaussian boson sampler for drug discovery, Nat. Comput. Sci. 3, 839 (2023).DOI
  13. [13]C. Oh, Y. Lim, Y. Wong, B. Fefferman, and L. Jiang, Quantum-inspired classical algorithms for molecular vibronic spectra, Nat. Phys. 20, 225 (2024).DOI
  14. [14]Y. Lim and C. Oh, Efficient classical algorithms for linear optical circuits (2025), arXiv:2502.12882 [quant-ph].
  15. [15]J.-L. Eickmann, K.-H. Luo, M. Roiz, J. Lammers, S. Atzeni, C. Pandey, F. Lüttner, R. G. Shirazi, F. Schlue, B. Brecht, V. V. Rybkin, M. Stefszky, and C. Silberhorn, Bridging chemistry and gaussian boson sampling: a photonic hierarchy of approximations for molecular vibronic spectra, npj Quantum Inf. 12, 89 (2026).DOI
  16. [16]J. E. Andersen and S. Shan, Using Gaussian boson samplers to approximate Gaussian expectation problems (2025), arXiv:2502.19336 [quant-ph].arxiv.org/abs/2502.19336
  17. [17]J. E. Andersen and S. Shan, Estimating the percentage of GBS advantage in Gaussian expectation problems (2025), arXiv:2502.19362 [quant-ph].arxiv.org/abs/2502.19362
  18. [18]M. C. Anguita, T. Roelink, S. Marzban, W. Briels, C. Filippi, and J. Renema, Experimental demonstration of boson sampling as a hardware accelerator for Monte Carlo integration (2025), arXiv:2509.25404 [quant-ph].arxiv.org/abs/2509.25404
  19. [19]D. M. Lopena, D. Buglak, Z. Li, E. Mer, S. H. Winston, S. Yu, M. Cucuringu, D. Rajan, P. Intallura, and R. B. Patel, Gaussian boson sampling for asset clustering in statistical arbitrage portfolios (2026), arXiv:2607.19279 [quant-ph].arxiv.org/abs/2607.19279
  20. [20]N. Spagnolo, D. J. Brod, E. F. Galvão, and F. Sciarrino, Non-linear boson sampling, npj Quantum Inf. 9, 3 (2023).DOI
  21. [21]F. Hoch, E. Carluccio, G. Rodari, T. Francelanci, A. Suprano, T. Giordani, G. Carvacho, N. Spagnolo, S. Koudia, M. Proietti, C. Liorni, F. Cerocchi, R. Albiero, N. Di Giano, M. Gardina, F. Ceccarelli, G. Corrielli, U. Chabaud, R. Osselame, M. Dispenza, and F. Sciarrino, Quantum machine learning with adaptive boson sampling via post-selection, Nat. Commun. 16, 902 (2025).DOI
  22. [22]M. Rambach, A. Roy, A. Gilchrist, A. Sakurai, W. J. Munro, K. Nemoto, and A. G. White, Photonic quantum-accelerated machine learning (2026), arXiv:2512.08318 [quant-ph].arxiv.org/abs/2512.08318
  23. [23]I. Paparelle, J. Henaff, J. García-Beni, É. Gillet, D. Montesinos, G. L. Giorgi, M. C. Soriano, R. Zambrini, and V. Parigi, Experimental memory control in continuous-variable optical quantum reservoir computing, Nat. Photon. 20, 413 (2026).DOI
  24. [24]Z. Yin, I. Agresti, G. de Felice, D. Brown, A. Toumi, C. Pentangelo, S. Piacentini, A. Crespi, F. Ceccarelli, R. Oselame, B. Coecke, and P. Walther, Experimental quantum-enhanced kernel-based machine learning on a photonic processor, Nat. Photon. 19, 1020 (2025).DOI
  25. [25]A. Sakurai, M. P. Estarellas, W. J. Munro, and K. Nemoto, Quantum extreme reservoir computation utilising scale-free networks, Phys. Rev. Appl. 17, 064044 (2022).DOI
  26. [26]V. Cimini, M. M. Sohoni, F. Presutti, B. K. Malia, S.-Y. Ma, R. Yanagimoto, T. Wang, T. Onodera, L. G. Wright, and P. L. McMahon, Large-scale quantum reservoir computing using a gaussian boson sampler, npj Quantum Inf. 12, 104 (2026).arxiv.org/abs/2505.13695
  27. [27]H. Zhu, W. Luo, R. Yan, C. Ren, J. Guo, Z. Zhao, H. Ma, T. Chen, F. Gao, L. C. Kwek, H. Cai, Y. Wang, J. Yang, and A.-Q. Liu, GBS-assisted quantum unsupervised machine learning on a universal programmable integrated quantum chip, Research 8, 1006 (2025).DOI
  28. [28]S.-Q. Gong, M.-C. Chen, H.-L. Liu, H. Su, Y.-C. Gu, H.-Y. Tang, M.-H. Jia, Y.-H. Deng, H.-T. Sun, Q. Wei, H. Wang, H.-S. Zhong, X. Jiang, L. Li, N.-L. Liu, D.-L. Deng, C.-Y. Lu, and J.-W. Pan, Enhancing image recognition using Gaussian boson sampling, Phys. Rev. Lett. 137, 080603 (2026).DOI
  29. [29]M. Sharifian and A. Bayat, Hybrid boson sampling-neural network architecture for enhanced classification, npj Quantum Inf. 10.1038/s41534-026-01321-2 (2026).DOI
  30. [30]P. Pfeffer, F. Heyder, and J. Schumacher, Hybrid quantum-classical reservoir computing of thermal convection flow, Phys. Rev. Res. 4, 033176 (2022).DOI
  31. [31]J. Henaff, M. Ansquer, M. C. Soriano, R. Zambrini, N. Treps, and V. Parigi, Optical phase encoding in a pulsed approach to reservoir computing, Opt. Lett. 49, 2097 (2024).DOI
  32. [32]J. García-Beni, I. Paparelle, V. Parigi, G. L. Giorgi, M. C. Soriano, and R. Zambrini, Quantum machine learning via continuous-variable cluster states and teleportation, EPJ Quantum Technol. 12, 63 (2025).
  33. [33]J. Biamonte, P. Wittek, N. Pancotti, N. Wiebe, and S. Lloyd, Quantum machine learning, Nature 549, 195 (2017).DOI
  34. [34]M. Schuld and N. Killoran, Quantum machine learning in feature Hilbert spaces, Phys. Rev. Lett. 122, 040504 (2019).DOI
  35. [35]M. Cerezo, G. Verdon, H. Y. Huang, L. Cinzio, and P. J. Coles, Challenges and opportunities in quantum machine learning, Nat. Comput. Sci. 2, 567 (2022).DOI
  36. [36]M. S. Rudolph, N. B. Toussaint, A. Katabarwa, S. Johri, B. Peropadre, and A. Perdomo-Ortiz, Generation of high-resolution handwritten digits with an ion-trap quantum computer, Phys. Rev. X 12, 031010 (2022).arxiv.org/abs/2012.03924
  37. [37]J. Herteux and C. Räth, Breaking symmetries of the reservoir equations in echo state networks, Chaos 30, 123142 (2020).DOI
  38. [38]H. Jaeger and H. Haas, Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication, Science 304, 78 (2004).
  39. [39]M. Lukoševičius and H. Jaeger, Reservoir computing approaches to recurrent neural network training, Comp. Sci. Rev. 3, 127 (2009).DOI
  40. [40]G. Tanaka, T. Yamane, J. B. Héroux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose, Recent advances in physical reservoir computing: A review, Neural Netw. 115, 100 (2019).DOI
  41. [41]S. Nerenberg, O. Neill, G. Marcucci, and D. Faccio, Photon number-resolving quantum reservoir computing, Optica Quantum 3, 201 (2025).
  42. [42]M. Selvičić, I. Agresti, M. Siemaszko, J. Morris, B. Dakić, R. Albiero, A. Crespi, F. Ceccarelli, R. Oselame, M. Stobińska-Moretto, and P. Walther, Experimental neuromorphic computing based on quantum memristor, PRX Quantum 7, 033067 (2026).
  43. [43]M. C. Mackey and L. Glass, Oscillation and chaos in physiological control systems, Science 197, 287 (1977).pubmed.ncbi.nlm.nih.gov/267326
  44. [44]R. Di Bartolo, S. Piacentini, F. Ceccarelli, G. Corrielli, R. Osellame, V. Cimini, and F. Sciarrino, Time-series forecasting with multiphoton quantum states and integrated photonics, npj Quantum Inf. 12, 91 (2026).
  45. [45]P. Mujal, R. Martínez-Pena, J. Nokkala, J. García-Beni, G. L. Giorgi, M. C. Soriano, and R. Zambrini, Opportunities in quantum reservoir computing and extreme learning machines, Adv. Quantum Technol. 4, 2100027 (2021).DOI
  46. [46]A. Labay-Mora, J. García-Benit, G. L. Giorgi, M. C. Soriano, and R. Zambrini, Neural networks with quantum states of light, Phil. Trans. R. Soc. A 382, 20230346 (2024).
  47. [47]G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, Extreme learning machine: a new learning scheme for feedforward neural networks, in 2004 IEEE International Joint Conference on Neural Networks (IEEE Cat. No.04CH37541) (2004) pp. 985–990 vol. 2.DOI
  48. [48]D. J. Gauthier, E. Bollt, A. Griffith, and W. A. S. Bar, Next generation reservoir computing, Nat. Commun. 12, 5564 (2021).
  49. [49]L. Jaurigue and K. Lüdge, Connecting reservoir computing with statistical forecasting and deep neural networks, Nat. Commun. 13, 227 (2022).DOI
  50. [50]W. Xiong, G. Facelli, M. Sahebi, O. Agnel, T. Chotibut, S. Thanasilp, and Z. Holmes, On fundamental aspects of quantum extreme learning machines, Quantum Mach. Intell. 7, 20 (2025).DOI
  51. [51]L. Innocenti, S. Lorenzo, I. Palmisano, A. Ferraro, M. Paternostro, and G. M. Palma, Potential and limitations of quantum extreme learning machines, Commun. Phys. 6, 118 (2023).DOI
  52. [52]L. Wang, P. Sun, L.-J. Kong, Y. Sun, and X. Zhang, Quantum next-generation reservoir computing and its quantum optical implementation, Phys. Rev. A 111, 022609 (2025).DOI
  53. [53]K. Fujii and K. Nakajima, Harnessing disordered-ensemble quantum dynamics for machine learning, Phys. Rev. Appl. 8, 024030 (2017).DOI
  54. [54]J. Chen, H. I. Nurdin, and N. Yamamoto, Temporal information processing on noisy quantum computers, Phys. Rev. Appl. 14, 024065 (2020).DOI
  55. [55]R. Araiza Bravo, N. Khadijeh, X. Gao, and S. F. Yelin, Quantum reservoir computing using arrays of Rydberg atoms, PRX Quantum 3, 030325 (2022).arxiv.org/abs/2111.10956
  56. [56]P. Pfeffer, F. Heyder, and J. Schumacher, Reduced-order modeling of two-dimensional turbulent Rayleigh-Bénard flow by hybrid quantum-classical reservoir computing, Phys. Rev. Res. 5, 043242 (2023).DOI
  57. [57]O. Ahmed, F. Tennie, and L. Magri, Prediction of chaotic dynamics and extreme events: A recurrent-free quantum reservoir computing approach, Phys. Rev. Res. 6, 043082 (2024).arxiv.org/abs/2405.03390
  58. [58]K. Kobayashi, K. Fujii, and N. Yamamoto, Feedback-driven quantum reservoir computing for time-series analysis, PRX Quantum 5, 040325 (2024).DOI
  59. [59]P. Mujal, R. Martínez-Peña, G. L. Giorgi, M. C. Soriano, and R. Zambrini, Time-series quantum reservoir computing with weak and projective measurements, npj Quantum Inf. 9, 16 (2023).DOI
  60. [60]S. Čindrak, B. Donvil, K. Lüdge, and L. Jaurigue, Enhancing the performance of quantum reservoir computing and solving the time-complexity problem by artificial memory restriction, Phys. Rev. Res. 6, 013051 (2024).DOI
  61. [61]M. Stefszky, K.-H. Luo, J.-L. Eickmann, S. Atzeni, F. Lütkewitte, C. Pandey, F. Schlue, J. Lammers, M. Roiz, T. Schapeler, L. Ares, M. Yahya pour, A. Kastner, J. Martinek, M. Mittermair, C. Sevilla, M. Leyendecker, O. Kohout, D. Mitin, R. Holzwarth, J. Sperling, T. Bartley, F. Steinlechner, B. Brecht, and C. Silberhorn, Benchmarking Gaussian and non-Gaussian input states with a hybrid sampling platform (2025), arXiv:2512.08433 [quant-ph].DOI
  62. [62]K.-H. Luo, D. Kopylov, F. Lütkewitte, J.-L. Eickmann, S. Atzeni, F. Schlue, B. Brecht, T. Meier, P. Sharapova, M. Stefszky, and C. Silberhorn, An ultra-bright, highly-scalable, squeezed light source for hybrid quantum photonics (2026), arXiv:2608.18950 [quant-ph].arxiv.org/abs/2608.18950
  63. [63]T. Schapeler, N. Lamberty, T. Hummel, F. Schlue, M. Stefszky, B. Brecht, C. Silberhorn, and T. J. Bartley, Electrical trace analysis of superconducting nanowire photon-number-resolving detectors, Phys. Rev. Appl. 22, 014024 (2024).DOI
  64. [64]M. Inubushi and K. Yoshimura, Reservoir computing beyond memory-nnnlinearity trade-off, Sci. Rep. 7, 10199 (2017).DOI
  65. [65]N. Herrmann, D. Arya, M. W. Doherty, A. Mingare, J. C. Pillay, F. Preis, and S. Prestel, Quantum utility – definition and assessment of a practical quantum advantage, in 2023 IEEE International Conference on Quantum Software (QSW) (2023) pp. 162–174.DOI
  66. [66]P. Holmes, J. L. Lumley, G. Berkooz, and C. W. Rowley, Turbulence, Coherent Structures, Dynamical Systems and Symmetry (Cambridge University Press, Cambridge, UK, 2012).
  67. [67]A. Gluhovsky, C. Tong, and E. Agee, Selection of modes in convective low-order models, J. Atmos. Sci. 59, 1383 (2002).doi.org/10.1175/1520-0469(2002)059<1383:somicl>2.0.co;2
  68. [68]E. N. Lorenz, Deterministic nonperiodic flow, J. Atmos. Sci. 20, 130 (1963).
  69. [69]E. N. Lorenz, Predictability: A problem partly solved, in Proceedings of the Seminar on Predictability, Vol. 1 (European Centre for Medium-Range Weather Forecasts, Reading, UK, 1995) pp. 1–18.
  70. [70]A. Chattopadhyay, P. Hassanzadeh, and D. Subramanian, Data-driven predictions of a multiscale Lorenz 96 chaotic system using machine-learning methods: reservoir computing, artificial neural network, and long short-term memory network, Nonlin. Processes Geophys. 27, 373 (2020).
  71. [71]S. K. Rathor, M. Ziegler, and J. Schumacher, Asymmetrically connected reservoir networks learn better, Phys. Rev. E 111, 015307 (2025).DOI
  72. [72]S. K. Rathor, L. Jaurigue, M. Ziegler, and J. Schumacher, Prediction performance of random reservoirs with different topology for nonlinear dynamical systems with different number of degrees of freedom, Chaos 36, 033117 (2026).DOI
  73. [73]N. Killoran, J. Izaac, N. Quesada, V. Bergholm, M. Amy, and C. Weedbrook, Strawberry Fields: A software platform for photonic quantum computing, Quantum 3, 129 (2019), arXiv:1804.03159.DOI
  74. [74]T. R. Bromley, J. M. Arrazola, S. Jahangiri, J. Izaac, N. Quesada, A. D. Gran, M. Schuld, J. Swinarton, Z. Zabaneh, and N. Killoran, Applications of near-term photonic quantum computers: software and algorithms, Quantum Science and Technology 5, 034010 (2020).
  75. [75]B. Gupt, J. Izaac, and N. Quesada, The walrus: a library for the calculation of hafnians, hermite polynomials and gaussian boson sampling, Journal of Open Source Software 4, 1705 (2019).DOI
  76. [76]S. Ehtesabi, M. Hossain, T. Kenter, A. Krawinkel, H. Nitsche, L. Ostermann, C. Plessl, H. Riebler, S. Rohde, R. Schade, M. Schwarz, J. Simon, N. Winnwa, A. Wiens, and X. Wu, Otus supercomputer (2025), arXiv:2512.07401 [cs.DC].arxiv.org/abs/2512.07401
  77. [77]F. Chillà and J. Schumacher, New perspectives in turbulent Rayleigh-Bénard convection, Eur. Phys. J. E 35, 58 (2012).DOI

Paper details

vlnc:W_ss7hpkzk
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents