betaFeedback

WorksPhysics and Astronomy

Deterministic nonlinear domain engineering algorithm with variable sub-coherence-length domains

Abstract

We present a new and improved deterministic algorithm that calculates the poling domains for the implementation of arbitrary phase-matching functions (PMF) in quasi-phase-matched nonlinear crystals for parametric down-conversion photon sources. In contrast to previous work, the algorithm presented here makes full use of the limits of the fabrication process, such as the minimum feature size and resolution of the lithographic process underpinning the poling procedure. This minimizes the tracking error to the target electric field and improves the accuracy in reaching the desired PMF. We give a tutorial on how to calculate the target function for arbitrary PMFs and run the algorithm. We also demonstrate the procedure on a relevant application in quantum sensing, where we use a quadratic phase on the PMF to implement a broadband photon spectrum while maintaining high nonlinearity across the crystal. The designed crystal is experimentally tested, and the result matches expectations.

Introduction

The development of versatile, bright and practical sources for quantum states of light is of great interest for photonic quantum technologies [1, 2]. Single-photons or entangled photon pairs, to name types of such quantum states, are used in technologies such as quantum communication [3, 4], cryptography [5], metrology [6, 7, 8], or computing [9].

Popular sources for these states include, for example, quantum dots [10], colour centres in diamond [11], or nonlinear optics such as spontaneous parametric down-conversion (SPDC) [12] or spontaneous four-wave mixing [13]. SPDC in particular is very versatile and has been successfully used to achieve high spectral purities [14, 15, 16, 17] and to generate entangled states in various degrees of freedom, such as polarization [18, 19], energy-time [20, 21], frequency-bins [22, 23], pulse modes [24, 25] or combinations thereof [26, 27, 28, 29, 30]. Widely used materials for this process are non-inversion symmetric crystals such as potassium titanyl phosphate (KTP) or lithium niobate. By alternating the orientation of the crystals during the growth process via electric poling, the nonlinear susceptibility of the material can be controlled and quasi-phase-matching can be achieved via constructive interference of the down-converted fields from subsequent domains [31]. The most basic poling structures achieving quasi-phase-matching consist of a fixed domain width that equals the coherence length of the nonlinear process. These simple structures have a fixed period, a 50 % duty cycle, and produce a (often undesired) sinc-shape phase-matching function (PMF). The sinc-shape arises from the Fourier transform of the constant effective nonlinearity within the nonlinear medium. Different PMFs can be realized by varying the duty cycle and period to alter the effective nonlinearity profile along the medium’s propagation axis. Such domain engineering is a flexible procedure to create photonic quantum states with engineered spectral and temporal properties [15, 32]. State-of-the-art algorithms use domains of predetermined width and vary their poling orientation [33, 34]. Analytical approaches that obtain a poling pattern directly from the desired nonlinear profile have also been proposed and tested [35, 36, 37]. Applications of these algorithms range from quantum communication [38, 39], over the generation of tailored pulse-mode entanglement useful for quantum information [25], to the generation of spectrally pure states [16, 17, 40] or states with tailored spectral correlations [41].

Here, we present an improved algorithm that sets the width and orientation of each domain individually, utilizing the full manufacturing potential of nonlinear crystals and, thereby, achieving smaller deviations between the target and generated down-converted fields compared to previous algorithms. We provide a practical guide on how to apply it for a large variety of custom requirements. As a specific application example, we demonstrate step-by-step the design of a degenerate SPDC-pair source with Gaussian spectra of broad bandwidth and high brightness, and experimentally validate it.

Section I introduces nonlinear domain engineering and presents a new algorithm, which deterministically finds poling structures with increased performance. In section II, this algorithm is used to produce a novel design for a nonlinear crystal used as a bright and broadband photon pair source for quantum sensing. The performance of this crystal is experimentally verified in section II B. Furthermore, a detailed step-by-step tutorial is presented in section III, guiding researchers through the domain engineering process. Section IV discusses further improvements to the algorithm, which help in tracking the PMF over large frequency ranges.

Domain engineering

Domain engineering finds poling structures of domains inside nonlinear media. Typical poling structures consist of a fixed domain width that equals the coherence length of the nonlinear process, thus enabling quasi-phase-matching.

Quasi-phase-matching

Quasi-phase-matching enables nonlinear optical processes for combinations of crystal materials and wavelengths, which are inaccessible to conventional methods such as birefringent phase-matching [31]. For three-wave-mixing processes (such as SPDC), the phase-mismatch is Δk=kp−ks−ki\Delta k = k_p - k_s - k_i, where kjk_j, with j∈{p,s,i}j \in\{\mathrm{p},\mathrm{s},\mathrm{i}\}, is the wavenumber of the pump, signal and idler photons, respectively and kj=2πnjλjk_j = \frac{2\pi n_j}{\lambda_j}, with λj\lambda_j (njn_j) denoting the wavelength (refractive index) of the respective wave. The coherence length lcl_c of the nonlinear process is given by

lc=πΔk.(1)l_c = \frac{\pi}{\Delta k}. \tag*{(1)}

It defines the length after which the converted waves (signal or idler) begin to interfere destructively with themselves. By flipping the lattice orientation within the nonlinear crystal every coherence length, an effective π\pi-phase-shift is introduced. This allows for constructive interference of the nonlinear process and yields an effective nonlinearity profile g(z)=χeff(2)(z)/χbulk(2)g(z) = \chi_{\mathrm{eff}}^{(2)}(z)/\chi_{\mathrm{bulk}}^{(2)}, where χeff(2)\chi_{\mathrm{eff}}^{(2)} is the effective nonlinear coefficient of second order and χbulk(2)\chi_{\mathrm{bulk}}^{(2)} is the nonlinear coefficient of the bulk material [42]. The drawback of quasi-phase-matching is that it has a maximum achievable effective nonlinearity of 2π<1\frac{2}{\pi} < 1. However, quasi-phase-matching allows for much greater flexibility than other phase-matching methods. By carefully designing the width and sign of each nonlinear domain, a process known as domain engineering, the tailoring of the biphoton spectral amplitude to specific applications becomes possible. This is, for example, the reason why heralded single-photon sources with domain engineered nonlinear crystals exhibit record purities [16, 17]. The PMF

ϕ(Δk)=∫−∞+∞g(z)eiΔkz dz(2)\phi(\Delta k) = \int_{-\infty}^{+\infty} g(z)e^{i\Delta k z}\,dz \tag*{(2)}

is the Fourier transform of the effective nonlinearity profile g(z)g(z), which alternates its sign between domains. Consequently, varying the effective nonlinearity along the crystal coordinate zz, enables tuning of the nonlinearity profile and PMF. Finding an optimal poling for a desired PMF is not trivial. Since χeff(2)\chi_{\mathrm{eff}}^{(2)} is constant within each domain, the desired nonlinearity profile can only be approximated by a series of step functions with finite widths. Moreover, practical considerations such as the minimum feature size and resolution of the lithographic process used in the manufacturing of the poling structure play a crucial role. Deterministic algorithms find working poling structures with relative ease, and the theoretical background is well documented in [33, 34].

Domain engineering algorithms

First algorithms used simulated annealing [43, 44] and other analytical methods to find poling structures of nonlinear crystals [15, 35, 36, 45], with further improvements found by gradient descent approaches [46]. Deterministic tracking algorithms have since been demonstrated, which dramatically reduce the computational effort compared to their predecessors [33, 34]. These greedy optimization algorithms set the poling structure from the beginning to the end of the nonlinear crystal by tracking the complex electric field amplitude of the generated (signal or idler) waves and try to match it to the desired field. In principle, the desired field amplitude profile

Atarget(z)=∫0zgtarget(z~) dz~.(3)A_{\mathrm{target}}(z) = \int_{0}^{z} g_{\mathrm{target}}(\tilde{z})\,d\tilde{z}. \tag*{(3)}

is calculated from an integral over the desired nonlinearity profile gtarget(z)g_{\mathrm{target}}(z), which is, in turn, given by the inverse Fourier transform of the desired PMF [33]

gtarget(z)∝∫−∞+∞ϕtarget(Δk)e−izΔk dΔk.(4)g_{\mathrm{target}}(z) \propto\int_{-\infty}^{+\infty} \phi_{\mathrm{target}}(\Delta k)e^{-iz\Delta k}\,d\Delta k. \tag*{(4)}

Note that gtarget(z)g_{\mathrm{target}}(z) is continuous and complex-valued, with no additional constraints. This approach is much more efficient than calculating the PMF in every step of an annealing algorithm. For example, a field amplitude following an error function is desired, when the PMF should be a Gaussian curve. The resulting nonlinearity profile will be Gaussian as well.

Most domain engineering designs have looked at a PMF with a flat phase, as this is necessary for high-purity state generation [47]. However, here we demonstrate that if purity is not relevant, carefully engineering this phase further increases flexibility and can lead to drastic improvements, such as higher brightness (see section II A).

Previous domain engineering algorithms, such as in [33] and [34], mainly use different configurations of domains of equal width. Graffitti et al. [34] additionally used stochastic annealing to modify the individual domain widths slightly and therefore further improve the result. However, these algorithms impose an unnecessary restriction, since modern photolithography can produce arbitrary domain widths in any configuration [48, 49]. Therefore, the optimization process is ultimately only limited by the minimal feature size lmin⁡l_{\min}, usually smaller than the coherence length lcl_{\mathrm{c}}, and the lithographic resolution lresl_{\mathrm{res}}, which is usually orders of magnitude smaller than lmin⁡l_{\min}.

A new algorithm utilizing variable domain widths

As discussed above, the target field within the crystal AtargetA_{\mathrm{target}} is determined from the desired PMF. The field generated by the poling profile for a given phase mismatch Δk\Delta k is denoted as A(z,Δk)A(z,\Delta k). At each step of the algorithm, where an additional domain of width ll is added, the new field is updated via

Anew(z+l,Δk)=A(z,Δk)+∫zz+lg(z~)eiΔkz~ dz~.(5)A_{\mathrm{new}}(z+l,\Delta k)=A(z,\Delta k)+\int_{z}^{z+l}g(\tilde{z})e^{i\Delta k\tilde{z}}\,\mathrm{d}\tilde{z}. \tag*{(5)}

This field is compared to the target field by evaluating the cost

ϵ=∣Anew−Atarget∣.(6)\epsilon=\lvert A_{\mathrm{new}}-A_{\mathrm{target}}\rvert. \tag*{(6)}

For this the algorithm uses the following procedure: it starts by setting the first domain with a width equal to the minimum feature size of the manufacturing process lmin⁡l_{\min} with positive poling.

  • Generate a candidate for swapped poling (Figure 1 left) by introducing a new domain of the width lmin⁡l_{\min} and an opposing poling to the previous domain.

    • Calculate new field using equation (5) and cost via equation (6).

Working principle of the domain engineering algorithm

Figure 1. Working principle of the domain engineering algorithm. As an example, one domain with the width of lil_i has already been computed. The white arrows indicate the poling direction. In the next step of the algorithm, either a new domain with opposite poling and the minimal domain width lmin⁡l_{\min} is added (left), or the current domain is extended by the domain resolution lresl_{\mathrm{res}} (right). For both these cases, the difference between the target field (red) and the computed realized field (orange) is calculated using equation (5), and the option with lower deviation (ϵj\epsilon_j from equation (6)) is chosen. In this example, the new-domain-option (left) is closer to the target and is therefore chosen.

  • Generate a candidate for remain poling (Figure 1 right) by extending the current domain by the photolithographic resolution lresl_{\mathrm{res}}.

    • Calculate new field using equation (5) and cost via equation (6).

  • Find the candidate with lower cost and adapt the domain structure accordingly.

  • Repeat until the end of the crystal is reached.

A first successful application of this algorithm was used to create hyperentanglement in polarization and frequency through joint spectrum shaping [30].

Algorithm comparison

To benchmark the proposed algorithm, we compare its performance against the method of Graffitti et al. [34] (without annealing) for real-valued Gaussian phase matching. We simulate the SPDC process in a 30 mm-long KTP crystal and assume a pump wavelength of 791 nm producing degenerate down-converted photons at 1582 nm. This wavelength combination allows for group velocity matching, and therefore high spectral purity [50, 51]. Since the fixed domain width, ww, in the method of Graffitti et al. must be selected manually, we consider two choices: (i) w=lc=23w = l_c = 23 µm, referred to as the coherence-length setting; and (ii) w=lmin⁡w = l_{\min}, referred to as the minimum-domain-width setting. In Figure 2(a), the computed field amplitude distance (mean absolute error) from the target field is shown for different minimum feature sizes. Although for large and small minimum feature sizes, the resulting

Figure 2. Comparison of algorithms for Gaussian phase-matching in 30 mm-long KTP for group velocity matched degenerate photon pairs at 1582 nm: For the algorithm by Graffitti et al. [34] two settings need to be considered: the coherence-length setting: w=lcw = l_c, and the minimum-domain-width setting: w=lmin⁡w = l_{\min}. For the algorithm presented in this work, the parameters do not need to be adjusted. (a) Distance between target and computed field as a function of minimum feature size. (b) Spectral purity from Schmidt decomposition [52] as a function of minimum feature size.

structure’s performance is practically identical to that of [34] (matching the coherence-length setting for lc≈lmin⁡l_c \approx l_{\min} and the minimum-domain-width setting for lc≫lmin⁡l_c \gg l_{\min}), we find improved performance for moderate minimum feature sizes on the order of lc/lmin⁡≈4l_c/l_{\min} \approx4. In this regime, we also find an increase in spectral purity as shown in Figure 2(b). Generally, for this work, the purity is more stable over a wide range of manufacturing parameters. The underlying reason for this is that with the added flexibility, the algorithm is able to follow the target field more closely, as further explained in appendix A. In conclusion, by allowing for more flexibility, the algorithm presented here maintains better performance across a wider range of coherence lengths without parameter tuning and removes the need to choose the right setting as in Graffitti et al..

An application of domain engineering for quantum sensing

The example we use to demonstrate the design process and this new algorithm centers on a quantum light source for a novel optical coherence tomography (OCT) approach using time-energy entangled photon pairs and Franson interference [53, 54]. We aim for degenerate photons centered at 809.4 nm and a Gaussian spectrum with a width of 45 nm (full width at half maximum (FWHM)). In conventional OCT, this leads to an axial resolution of 6.4 μ\mum [55]. OCT using the Hong-Ou-Mandel effect [56] (also termed Quantum-OCT) has already been proposed [57] and demonstrated [58], providing a twofold resolution enhancement compared to its classical counterpart. Record axial resolution has been achieved by using a chirped quasi-phase-matched lithium tantalate device, where the poling period was linearly increased along the crystal [59]. This allows multiple wavelengths to phase-match at different positions within the crystal, thereby increasing the spectral bandwidth. Although a Quantum-OCT resolution of 0.54 μ\mum could be reached, the interference pattern exhibits strong undesired sidelobes, stemming from the non-Gaussian spectral shape, which in turn arises from the chirped quasi-phase-matching used. Most thus far explored sources for broadband photons have used such simple, strictly linear chirping of the poling period, even though there is great potential in incorporating truly aperiodic structures [32]. It would also be possible to obtain a large spectral bandwidth and, therefore, a high OCT resolution, using a very thin nonlinear crystal, which naturally allows for a larger phase mismatch [60]. However, such a small length of only a few hundred micrometers to a few millimeters leads to a very small interaction region between pump and down-converted fields, resulting in low total photon flux. By simply extending the length of the crystal, this flux increases, but the bandwidth also reduces, necessitating a trade-off between the two observables. In contrast, we present here a different approach that avoids this trade-off.

Improved quasi-phase-matching

Since the envisioned OCT-scheme does not require interference between independent photons, the source does not need to produce spectrally pure states, and we can imprint a quadratic phase on the PMF

ϕtarget(Δk)=exp⁡(−Δk22σϕ2)⋅exp⁡(i a Δk2),(7)\phi_{\mathrm{target}}(\Delta k)=\exp\left(-\frac{\Delta k^{2}}{2\sigma_{\phi}^{2}}\right)\cdot\exp\left(i\,a\,\Delta k^{2}\right), \tag*{(7)}

where σϕ\sigma_{\phi} is the width that eventually determines the spectral bandwidth of the down-converted photons (see equation (1) and equation (B2)), with k=2πn/λk=2\pi n/\lambda, and aa is the quadratic phase coefficient, used to increase the width of the nonlinearity profile, thus increasing brightness. This is ultimately limited by the available crystal length. From this desired PMF, the target nonlinearity profile can be computed (equation (4)) and, in turn, using equation (3), the target field can be calculated for each position within the crystal. Analytical expressions are presented in appendix B, and they are also visualized in Figure 3. From equation (7), the quadratic phase coefficient is

Figure 3. (a) Analytically computed nonlinearity profile gtarget(z)g_{\mathrm{target}}(z) inside the crystal of length L=30L = 30 mm, with σχ=L/4\sigma_{\chi} = L/4 and σϕ\sigma_{\phi} corresponding to a bandwidth of 45 nm. (b) The resulting target field Atarget(z)A_{\mathrm{target}}(z). Analytical expressions are presented in appendix B.

a=σχ2σϕ2−12σϕ2,(8)a=\frac{\sqrt{\sigma_{\chi}^{2}\sigma_{\phi}^{2}-1}}{2\sigma_{\phi}^{2}}, \tag*{(8)}

with σχ\sigma_{\chi} being the width (standard deviation) of the absolute value of the nonlinearity profile within the crystal. As suggested by Tambasco et al. [33], a reasonable value is σχ=L/4\sigma_{\chi}=L/4, with LL denoting the length of the crystal, which is a compromise between accuracy to the desired shape and brightness. A larger σχ\sigma_{\chi} would increase brightness at the cost of inducing oscillations in the PMF as a result of the truncation of the nonlinearity profile at the crystal boundaries.

Note that, even though the PMF has a quadratic phase, which equates to a linear chirp, this design is fundamentally distinct from linearly increasing the poling period along the crystal, which is called chirped quasi-phase-matching [61, 62, 63]. The main difference is that we obtain a highly nontrivial poling pattern (see appendix C), since this approach utilizes more degrees of freedom, increasing the flexibility. This allows the generation of customizable spectral shapes. With this design, substantial increases in spectral bandwidth lead to only minor reductions in overall brightness, underscoring that this undesirable trade-off has been largely mitigated, as shown in appendix D in more detail.

In our application example, we choose collinear type-II SPDC in KTP and deploy the algorithm described in section 1C to find a domain structure matching the target field. The resulting electric field matches the target within the thickness of the lines used to visualize them in Figure 3(b). We numerically verify the PMF resulting from this domain structure for the entire two-dimensional joint spectral state and show the result in Figure 4(a), together with a visualization of the amplitude of a narrowband pump laser. The product of these two functions gives rise to the joint spectral amplitude (JSA) of this SPDC output state.

Figure 4. (a) Simulated PMF and pump amplitude. This two-dimensional visualization of the PMF shows the amplitude of a given wavelength pair created by the crystal. The blue line labeled ”pump” represents all wavelength pairs that adhere to conservation of energy for a narrowband pump at 404.7 nm. (b) Experimentally measured signal and idler spectra at a crystal temperature of 70 °C, in comparison to a simulated spectrum using a plane wave propagated through the crystal. To reduce high-frequency noise, a moving average with a window of 2 nm (20 data points) was used.

Experimental results

The crystal length is constrained by the lithography fabrication process, and is therefore fixed at a current achievable maximum of 30 mm. The crystal was commercially fabricated by SLF Quantum for a signal-idler wavelength-degeneracy temperature of 50 °C. It is placed inside a temperature-controlled oven and put into an optical setup to test its performance. The setup consists of a fiber-coupled continuous wave pump laser at λp=404.7\lambda_p = 404.7 nm, which is polarization filtered using a polarizing beam splitter (PBS) to ensure only the relevant horizontally polarized light reaches the crystal. The pump is focused into the center of the crystal to achieve a waist radius of ≈15\approx15 µm, and the signal, idler, and pump modes are collimated again after the crystal. A dichroic mirror and a hard-coated longpass filter are used to filter out the remaining pump light.

The orthogonally polarized down-converted photons are deterministically split using another PBS before being coupled into separate single-mode fibers for detection. Using superconducting nanowire single-photon detectors (quantum efficiency 88(3) %), we were able to measure single photon count rates of up to Rsignal≈Ridler≈200R_{\mathrm{signal}} \approx R_{\mathrm{idler}} \approx200 kHz per mW of pump power and a pair rate of Rpair≈40R_{\mathrm{pair}} \approx40 kHz per mW, resulting in a raw, uncorrected heralding efficiency of η=RpairRsignalRidler≈20\eta= \frac{R_{\mathrm{pair}}}{\sqrt{R_{\mathrm{signal}}R_{\mathrm{idler}}} } \approx20 %.

The individual spectral intensities were measured using a grating spectrometer, and the results are shown in Figure 4(b) for the measured degeneracy temperature of ≈70\approx70 °C. A comparison to a simulated spectrum (which is approximately identical for signal and idler), for which plane waves are numerically propagated through the crystal, shows good agreement. The largest deviation between simulation and measurement can be seen for the signal photon’s spectrum between 780 nm and 790 nm. A mirror image of this deviation is present in the idler spectrum at about 840 nm, too. This suggests some minor imperfections in the PMF itself, probably arising from inaccuracies in the fabricated nonlinearity profile. Note that there is a small discrepancy between the desired spectral bandwidth of 45 nm and a fit to the result of the simulation with 43.5 nm, stemming from discretization errors when approximating the smooth target nonlinearity with a real and finite binary step function. This highlights the need to verify the generated poling structure and to possibly compensate accordingly. After manufacturing the crystal, we found a slightly modified design approach addressing this issue, and we present simulations in section IV A.

Even the here implemented source, as designed with the basic version of our algorithm, roughly produces the same pair rate as a similar setup containing a 10 mm long periodically poled KTP crystal with a poling period of 9.9 µm. However, we achieve over 50 times larger spectral bandwidth.

Step-by-step guide

Here we provide a tutorial on how to approach the domain engineering process in several steps.

  1. Define desired properties of the photon pair:

An important consideration is whether the state needs to be spectrally pure. This is the case if independent photons from such states interfere with one another [52]. If no such interference is desired, one can relax this purity constraint.

  1. Note that, as evident from equation (3), domain engineering can only alter the PMF along the gradient of

Δk(λs,λi)=2π(np(λp)λp−ns(λs)λs−ni(λi)λi)(9)\Delta k(\lambda_s,\lambda_i) = 2\pi\left(\frac{n_p(\lambda_p)}{\lambda_p} - \frac{n_s(\lambda_s)}{\lambda_s} - \frac{n_i(\lambda_i)}{\lambda_i}\right) \tag*{(9)}

with λp−1=λs−1+λi−1\lambda_p^{-1} = \lambda_s^{-1} + \lambda_i^{-1}. Precise knowledge of the Sellmeier coefficients for the given material is therefore necessary [64, 65, 66]. By examining this Δk\Delta k-landscape in the wavelength region of interest, one needs to consider whether domain engineered quasi-phase-matching provides an advantage over birefringent phase-matching. Furthermore, since the various types of SPDC require different methods for separating signal and idler photons, a suitable configuration needs to be selected.

  1. Find the PMF in Δk\Delta k-space:

Using equation (1) and equation (14), one can express the desired central wavelength and bandwidth in terms of coherence lengths of the nonlinear process. This allows a definition of a PMF, such as the chirped Gaussian of equation (7) chosen here as an example.

Small discrepancies between the desired and fabricated poling structure can change the down-converted photon properties. Most notably, the temperature at which signal and idler wavelengths are degenerate is hard to predetermine precisely. For this reason, it is advisable to set this degeneracy temperature several tens of degrees Celsius above room temperature. This allows easy temperature control using resistive heaters in an oven.

  1. Fourier transform the desired PMF into an effective nonlinearity profile gtarget(z)g_{\mathrm{target}}(z) and normalize it so that its maximum (here located in the center of the crystal) coincides with the maximum achievable by periodic poling for the central wavelengths, i.e., gtarget(L/2)=2/πg_{\mathrm{target}}(L/2) = 2/\pi [33].

  1. The target field Atarget(z)A_{\mathrm{target}}(z) can now be calculated analytically using equation (3). Note that it is sufficient to perform these calculations for only one Δk\Delta k, significantly reducing computational costs.

  1. Specify manufacturing parameters (crystal length, minimal domain width, and domain resolution) and run the tracking algorithm. Export the start and endpoints, as well as the poling orientation, for all the domains.

  1. Validate the domain structure using a numerical simulation that propagates the entire two-dimensional joint field (as a function of wavelengths λs\lambda_s, λi\lambda_i; for narrow-band pumps the variation of a single wavelength is sufficient)

through the crystal using equation (5). Note that this task is much more computationally expensive than running the algorithm itself. However, it is highly parallelizable [67]. Note that, depending on the requirements, a simple plane wave propagation is sufficient, or, in the case of tight focusing, the spatial mode profiles of the pump, signal, and idler fields need to be considered as well [68, 69]. The JSA obtained in this way can now be analysed. For example, one can estimate the spectral purity via Schmidt decomposition [52] or project the JSA onto signal and idler axes, respectively, to obtain the marginal spectral distributions. These can be compared to the desired spectra from step 1.

Further improvements: spectral shape optimization

Here we propose additional improvements to the algorithm. These modifications were only found after the crystal was manufactured and tested, which is why no experimental verification was performed, and we show here only simulation results. As discussed above, the greedy optimization of domain engineering algorithms tracks the target field by matching its complex value at a single phase mismatch Δk0\Delta k_0. While this is sufficient to approximate the target PMF at a specific phase mismatch, ϕtarget(Δk0)\phi_{\mathrm{target}}(\Delta k_0), it does not guarantee that the PMF is approximated well for different values of Δk\Delta k. Considering Δk=Δk0+q\Delta k = \Delta k_0 + q, the deviation between the realized and ideal target PMF, δϕ(Δk)=ϕ(Δk)−ϕtarget(Δk)\delta\phi(\Delta k) = \phi(\Delta k) - \phi_{\mathrm{target}}(\Delta k), at detuning qq is (see derivation in appendix E § 1.13.1)

δϕ(q)=δA(L)eiqL−iq∫0LδA(z)eiqz dz+ϵL(q),(10)\delta\phi(q) = \delta A(L)e^{iqL} - iq \int_{0}^{L} \delta A(z)e^{iqz}\,\mathrm{d}z + \epsilon_L(q), \tag*{(10)}

where δA(z)=A(z)−Atarget(z)\delta A(z) = A(z)-A_{\mathrm{target}}(z) is the tracking error at each position zz and ϵL\epsilon_L is the difference between the target PMF evaluated over the finite crystal length and the ideal target PMF. This difference is inherent to any realized poling structure, and its magnitude depends on σχ/L\sigma_\chi/L; i.e., the more gradually the envelope of the nonlinearity approaches zero, the lower ϵL\epsilon_L. This shows that, unless δA(z)=0 ∀z∈[0,L]\delta A(z) = 0\ \forall z \in[0,L], which is generally not possible with only a binary poling, the tracking error is weighted by qq, potentially introducing distortions at large enough qq.

Demodulation

To give a first demonstration of this effect, we analyze again the design in section II with a different approach. So far, the algorithm has tried to minimize the distance between the realized field

A(z)=∫0zg(z~)eiΔkz~ dz~(11)A(z) = \int_{0}^{z} g(\tilde{z})e^{i\Delta k\tilde{z}}\,\mathrm{d}\tilde{z} \tag*{(11)}

and Atarget(z)A_{\mathrm{target}}(z), as defined in equation (3). Here, to improve upon this, we decompose the target nonlinearity profile into its amplitude and phase gtarget(z)=∣gtarget(z)∣⋅eiθ(z)g_{\mathrm{target}}(z) = \lvert g_{\mathrm{target}}(z)\rvert\cdot e^{i\theta(z)}, and we define new functions

B(z)=∫0zg(z~)ei[Δkz~−θ(z~)] dz~andBtarget(z)=∫0z∣gtarget(z~)∣ dz~,(12)B(z) = \int_{0}^{z} g(\tilde{z})e^{i[\Delta k\tilde{z}-\theta(\tilde{z})]}\,\mathrm{d}\tilde{z} \quad\text{and}\quad B_{\mathrm{target}}(z) = \int_{0}^{z} \lvert g_{\mathrm{target}}(\tilde{z})\rvert\,\mathrm{d}\tilde{z}, \tag*{(12)}

which corresponds to a change of reference frame. This transformation acts on the integrand of the field B(z)B(z), such that the algorithm generally minimizes a different quantity while leaving the target PMF unchanged.

In contrast to the complex oscillating target field Atarget(z)A_{\mathrm{target}}(z) of Figure 3(b), the new target Btarget(z)B_{\mathrm{target}}(z) is a smooth real function, as shown in Figure 5(a). Thus, minimizing the distance ∣B−Btarget∣\lvert B-B_{\mathrm{target}}\rvert and therefore δϕ\delta\phi, analogously to the procedure outlined in section IC, mitigates the effect of the inevitable discretization errors on the realized PMF at Δk≠Δk0\Delta k \ne\Delta k_0. Therefore, the target PMF can be better approximated over the entire range of frequencies (see details in appendix E § 1.13.2).

Figure 5. (a) Real and imaginary part of the normalized demodulated target field Btarget(z)B_{\mathrm{target}}(z). (b) Comparison of the computed signal intensities from the original design (default) and the demodulation approach to the target function. The inset shows the difference in intensities to the target. To reduce high-frequency noise, a moving average with a window of 2 nm (20 data points) was used.

This is demonstrated in Figure 5(b), which shows the computed signal spectrum at signal-idler wavelength degeneracy for both approaches. The demodulated case more closely approximates the desired function by reducing the spectral distortions away from the phase-matched wavelength. The Gaussian fit to the simulated spectrum of the demodulated approach yields a FWHM of 45 nm, in precise agreement with the initially desired value. Note that the simulated degeneracy temperature shifts slightly between the two approaches, from 50° C in the original to 65° C in the demodulated case. This modification improves only the spectral shape of the down-converted photons, leaving the overall flux approximately unchanged. Furthermore, for PMFs with a flat phase, i.e. θ(z)=const.\theta(z) = \mathrm{const.} (as used for the comparisons in section ID), this approach defaults back to the original algorithm.

Structure Refinement

Motivated by the observations of equation (10) and the result of the previous section, we introduce here a new and efficient refinement algorithm that deterministically and locally optimizes the initial poling structure produced by the greedy optimizer. This simultaneously reduces the complex PMF error across a set of Δk\Delta k values while respecting the fabrication constraints. The algorithm varies the position of each domain wall resulting from the initial optimisation in sequence from z=0z = 0 to z=Lz = L, evaluating the effect of positive and negative displacements, applied in integer steps of the minimal resolution lresl_{\mathrm{res}}, on the complex PMF error,

J=∑Δkξ(Δk)∣ϕ(Δk)−ϕtarget(Δk)∣2,(13)J = \sum_{\Delta k} \xi(\Delta k) \left\lvert\phi(\Delta k) - \phi_{\mathrm{target}}(\Delta k) \right\rvert^2 , \tag*{(13)}

where ξ(Δk)\xi(\Delta k) is a weighting function that prioritizes specific Δk\Delta k values (see also appendix E3). Only displacements that preserve a minimum width w≥lminw \ge l_{\mathrm{min}} of neighbouring domains are considered, and the one giving the largest reduction of JJ is accepted when the reduction exceeds a minimum threshold. Otherwise, the wall is left unchanged. In addition to these displacements, the algorithm also considers the removal of a pair of adjacent walls: three domains of alternating sign collapse into a single domain with same total width, provided this reduces JJ. No walls are inserted, so the number of domains can only decrease, and the total crystal length is preserved. This whole procedure is performed iteratively, until no reduction of JJ above the minimum threshold is found. The refinement algorithm is also described step-by-step in appendix E4.

This new refinement procedure has been applied in simulation to the designs presented above (default and demodulated) and both cases lead to an improvement of the final PMF (see more details and comparisons in appendix E5). The result of the refinement procedure remains directly manufacturable, as every displacement is an integer multiple of lresl_{\mathrm{res}} and lminl_{\mathrm{min}} is enforced, so no further discretisation step is needed. The algorithm is inexpensive (see appendix E6), accepts only reductions of JJ starting from an input structure, does not involve the randomness of simulated annealing [34], provides the same output at every run, and converges in tens to hundreds of iterations. As noted before, it does not provide a global optimizer and must be thought of as a cheap optimization of tracked structures obtained with domain engineering algorithms that can result in significant improvements.

Conclusion and outlook

The deterministic domain engineering algorithm presented here utilizes the full manufacturing potential of nonlinear crystals and, thereby, achieves smaller deviations between the target and generated down-converted fields compared to previous algorithms. We provide a detailed practical guide on how to use this algorithm, including a step-by-step tutorial and the source code written in Python. This guide is aimed at researchers requiring nonlinear optical sources, and it helps in all phases of the design process, from choosing sensible desired photon properties to final verification. An example of such a design process is given for a 30 mm-long KTP crystal used as a bright and broadband photon pair source for quantum sensing. By understanding the experimental requirements, we create a design that utilizes the maximum crystal length currently manufacturable, leading to strong nonlinear interaction and therefore a high pair rate of 40 kHz per mW for degenerate photon pairs centered at 809 nm, with quasi-Gaussian spectral shape and a FWHM-bandwidth of 43.5 nm. The commercially fabricated domain structure has been experimentally tested and shows good agreement with the expected results.

Further improvements to the algorithm are discussed and show potential for the realization of complex yet precise PMFs. We hope that this new algorithm, in tandem with this guide, further encourages the use of domain engineering in quantum light sources. This can allow a great variety of technologies to benefit from this mostly untapped potential.

Source code/data availability

An implementation of the presented algorithm, written in Python, is available at: https://git.uibk.ac.at/c7041192/apyriodic-poling-public

The data and code that support the findings of this study are openly available in reference [70].

Funding

This research was funded in part by the Austrian Science Fund (FWF), projects 10.55776/F71 and 10.55776/PIN6357923 and the MEANTIME 1669 Prototype Project funded through the 1669 Förderverein.

References

Algorithm's enhanced flexibility

When compared to Graffitti et al. [34], our work achieves better performance because the variable domain widths add flexibility (see § 1.1 D). This manifests in a closer match between the generated field to the target field, as illustrated in Figure 6. For the coherence-length setting of [34], i.e., w=lcw = l_c (Figure 6(a)), the algorithm fails to closely follow the target field when its slope is not at a maximum. For the used example of Gaussian phase-matching, this maximum in slope is only reached at the center of the crystal, while everywhere else it is smaller. Since our algorithm allows for domains with arbitrary widths, these smaller slopes can be followed more closely, as shown in Figure 6(b). When considering small domain widths close to the resolution limit (w=lmin⁡w = l_{\min}), the algorithm of [34] fails to produce a

Figure 6. Visual explanation of the enhanced performance in comparison to the algorithm by Graffitti et al. [34]. (a-b) A large discretization when the domain width equals the coherence length leads to artifacts. (c-d) Small domain widths (w=lmin⁡w = l_{\min}) fail to produce strong effective nonlinearity for coherence lengths that are not divisible by the domain widths. For both regimes, our ability to set each domain width individually leads to a smaller discrepancy from the target field. For [34], one needs to set one domain width for the entire crystal, leading to less flexibility.

strong effective nonlinearity for coherence lengths which are not divisible by lmin⁡l_{\min}, see Figure 6(c). The reason for this is that the strongest nonlinear conversion takes place if the domain width equals the coherence length. However, lcl_c will generally not be divisible by lmin⁡l_{\min}, such that after one coherence length there will be one domain that causes partially destructive interference, leading to an overall decrease in field amplitude. Again, our work reduces this discrepancy, as shown in Figure 6(d). Therefore, the distance to the target field can be decreased without parameter tuning, using the additional flexibility of the new algorithm.

Analytical forms of expressions

Note that the PMF is a function of Δk\Delta k and the computations are conveniently expressed in terms of the central coherence length lcl_c and

δlc=∣πΔk(λs−,λi−)−πΔk(λs+,λi+)∣,(14)\delta l_c = \left| \frac{\pi}{\Delta k(\lambda_{s-},\lambda_{i-})} - \frac{\pi}{\Delta k(\lambda_{s+},\lambda_{i+})} \right|, \tag*{(14)}

the bandwidth of the nonlinear process, with λs±=λs±Δλ/2\lambda_{s\pm} = \lambda_s \pm\Delta\lambda/2 and 1/λi±=1/λs+1/λi−1/λs±1/\lambda_{i\pm} = 1/\lambda_s + 1/\lambda_i - 1/\lambda_{s\pm} for a spectral bandwidth of Δλ\Delta\lambda and central wavelengths λs,i\lambda_{s,i}. Δk\Delta k is defined in equation (9). We therefore use

σϕ=π22ln⁡(2)⋅δlclc2−δlc2/4.(15)\sigma_\phi= \frac{\pi}{2\sqrt{2\ln(2)}} \cdot\frac{\delta l_c}{l_c^2 - \delta l_c^2/4}. \tag*{(15)}

The Gaussian PMF with a quadratic phase (example given above) is

ϕtarget(Δk)=exp⁡(−Δk22σϕ2)⋅exp⁡(iaΔk2),(16)\phi_{\mathrm{target}}(\Delta k) = \exp\left(-\frac{\Delta k^2}{2\sigma_\phi^2}\right) \cdot\exp\left(i a \Delta k^2\right), \tag*{(16)}

where the quadratic phase broadens the nonlinearity profile, increasing photon pair flux. Via a Fourier transform, we find the effective nonlinearity profile

gtarget(z)=2πexp⁡(−iπ2δlc2z24π2aδlc2+iln⁡(2)(δlc2−4lc2)),(17)g_{\mathrm{target}}(z) = \frac{2}{\pi}\exp\left(-\frac{i\pi^2\delta l_c^2 z^2}{4\pi^2 a\delta l_c^2 + i\ln(2)\left(\delta l_c^2 - 4l_c^2\right)}\right), \tag*{(17)}

which is shown in Figure 3(a). This leads to a target field of the form

Atarget(z)=ln⁡(2)(δlc2−4lc2)2−4iπ2aδlc2π3/2δlc[erf⁡(πδlcL2ln⁡(2)(δlc2−4lc2)2−4iπ2aδlc2)+erf⁡(πδlc(z−L2)ln⁡(2)(δlc2−4lc2)2−4iπ2aδlc2)].(18)A_{\mathrm{target}}(z) = \frac{\sqrt{\ln(2)\left(\delta l_c^2 - 4l_c^2\right)^2 - 4i\pi^2 a\delta l_c^2}}{\pi^{3/2}\delta l_c} \left[ \operatorname{erf}\left(\frac{\pi\delta l_c L}{2\sqrt{\ln(2)\left(\delta l_c^2 - 4l_c^2\right)^2 - 4i\pi^2 a\delta l_c^2}}\right) + \operatorname{erf}\left(\frac{\pi\delta l_c\left(z-\frac{L}{2}\right)}{\sqrt{\ln(2)\left(\delta l_c^2 - 4l_c^2\right)^2 - 4i\pi^2 a\delta l_c^2}}\right) \right]. \tag*{(18)}

which is shown in Figure 3(b).

Domain structure analysis

For our requirements mentioned in section II, the coherence length for the central wavelengths correspond to lc=4.95 μml_c = 4.95\,\mathrm{\mu m}, while the bandwidth corresponds to δlc=0.5 μm\delta l_c = 0.5\,\mathrm{\mu m}. We use a=134 625 μm2a = 134\,625\,\mathrm{\mu m}^2 for L=30 mmL = 30\,\mathrm{mm}. The minimal domain size was chosen to be 4 μm4\,\mathrm{\mu m} and the resolution 125 nm125\,\mathrm{nm}. We obtain a structure with 3481 domains. The smallest domain size we obtain is 4.375 μm4.375\,\mathrm{\mu m}. The different widths of the domains produced by the algorithm are shown in Figure 7(a). Note that one domain with a width of 400 μm400\,\mathrm{\mu m} at the beginning, and one of 500 μm500\,\mathrm{\mu m} at the end, are omitted for better visibility of the remaining domains. We see clear accumulations of domain widths at ≈5 μm\approx5\,\mathrm{\mu m} towards the center of the crystal, and accumulations of multiples of this value more towards the edges. This is expected, as it corresponds to the coherence length of this nonlinear process for the central wavelengths. A histogram of the domain widths around the coherence length is shown in Figure 7(b), where the bin size was set to the domain resolution. We observe that, although the largest bin of domains is the one with a width of exactly 5 μm5\,\mathrm{\mu m}, a significant portion of domains are longer or shorter. In fact, 65%65\% of these domains have a width differing from 5 μm5\,\mathrm{\mu m}. This shows that, when allowing for an arbitrary width for each domain individually, one achieves closer agreement between the target and the realized down-converted field.

Figure 7. (a) Visualization of the domain widths in the used design. For illustrative purposes, the width of each negatively poled domain is sign-inverted. (b) Histogram of positive domain widths close to the coherence length. The bin size is set to the domain resolution of 125 nm

Design parameter variations

To further discuss the design introduced in section II A, we simulate a variation in the desired spectral bandwidth and calculate the resulting total photon pair flux, i.e., the integral over the entire joint spectral intensity, in the plane-wave limit, i.e., in the absence of focusing, for a 30 mm long crystal. We find that the flux only varies slowly with bandwidth: even a large increase in bandwidth (from 43.5 nm to 81 nm) only reduces the flux by 6.7%. Likewise, a reduction in bandwidth to 22 nm leads to only a small increase in flux of 2.1%. A variation of the crystal length (for a fixed spectral bandwidth) shows a linear dependence of the total photon pair flux.

Details Improvements

Target Deviation

We derive here equation (10). Let us consider the realized field in equation (11), and the corresponding PMF,

ϕ(Δk)=∫0Lg(z)eiΔkz dz.(19)\phi(\Delta k) = \int_{0}^{L} g(z)e^{i\Delta k z}\,dz. \tag*{(19)}

Introducing the detuning q=Δk−Δk0q = \Delta k - \Delta k_0, the PMF can be written as

ϕ(q)=∫0Lg(z)eiΔk0zeiqz dz=∫0LdA(z)dzeiqz dz,(20)\phi(q) = \int_{0}^{L} g(z)e^{i\Delta k_0 z}e^{iqz}\,dz = \int_{0}^{L} \frac{dA(z)}{dz}e^{iqz}\,dz, \tag*{(20)}

where the second equality comes from the definition of the field in equation (11). Integration by parts and the fact that A(0)=0A(0) = 0 yield

ϕ(q)=A(L)eiqL−iq∫0LA(z)eiqz dz,(21)\phi(q) = A(L)e^{iqL} - iq\int_{0}^{L} A(z)e^{iqz}\,dz, \tag*{(21)}

which reduces to the expected expression for q=0q = 0, i.e., ϕ(q=0)=A(L)=ϕ(Δk0)\phi(q = 0) = A(L) = \phi(\Delta k_0). Since we never used any specific property of g(z)g(z), the same analysis holds true for the target PMF evaluated over the finite crystal length

ϕtarget,L(q)=∫0Lgtarget(z)eiqz+iΔk0z dz\phi_{\mathrm{target},L}(q) = \int_{0}^{L} g_{\mathrm{target}}(z)e^{iqz+i\Delta k_0z}\,dz,

and the deviation δϕL(q)=ϕ(q)−ϕtarget,L(q)\delta\phi_L(q) = \phi(q) - \phi_{\mathrm{target},L}(q) between the PMFs at qq results in

δϕL(q)=δA(L)eiqL−iq∫0LδA(z)eiqz dz.(22)\delta\phi_L(q) = \delta A(L)e^{iqL} - iq\int_{0}^{L} \delta A(z)e^{iqz}\,dz. \tag*{(22)}

The deviation to the ideal target PMF can be obtained by adding ϕtarget,L(q)\phi_{\mathrm{target},L}(q) and subtracting ϕtarget(q)\phi_{\mathrm{target}}(q), leading to equation (10) with ϵL(q)=ϕtarget,L(q)−ϕtarget(q)\epsilon_L(q) = \phi_{\mathrm{target},L}(q) - \phi_{\mathrm{target}}(q).

Details Demodulation

The PMF defined in equation (7) has a chirp, which leads to a phase in the nonlinearity profile. This phase is reproduced through a local poling period that varies along the crystal, causing different down-converted frequencies to be generated at different positions within the crystal. In fact, since the algorithm ensures g(z)eiΔk0z≈gtarget(z)=∣gtarget(z)∣eiθ(z)g(z)e^{i\Delta k_0z}\approx g_{\mathrm{target}}(z)=|g_{\mathrm{target}}(z)|e^{i\theta(z)}, the binary poling must contain a spatial component g(z)∝e−i[Δk0z−θ(z)]g(z)\propto e^{-i[\Delta k_0z-\theta(z)]} with local phase mismatch ddz[Δk0z−θ(z)]=Δk0−θ′(z)\frac{d}{dz}[\Delta k_0z-\theta(z)]=\Delta k_0-\theta'(z), which equals the phase mismatch Δk=Δk0+q\Delta k=\Delta k_0+q when θ′(z)=−q\theta'(z)=-q. The down-converted frequency related to Δk\Delta k is, thus, generated only in a region of the crystal where this condition is approximately fulfilled, i.e., θ′(z)≈−q\theta'(z)\approx-q. In the default frame, the deviation of the PMF is given by equation (22), which depends on qq, leading to larger errors at large qq. In the demodulated frame, however, the same integration by parts yields

δϕL(q)=[B(L)−Btarget(L)]ei[θ(L)+qL]−i∫0L(θ′(z)+q)[B(z)−Btarget(z)]ei[θ(z)+qz] dz,(23)\delta\phi_L(q)=[B(L)-B_{\mathrm{target}}(L)]e^{i[\theta(L)+qL]}-i\int_0^L(\theta'(z)+q)[B(z)-B_{\mathrm{target}}(z)]e^{i[\theta(z)+qz]}\,dz, \tag*{(23)}

where the integral becomes small since θ′(z)≈−q\theta'(z)\approx-q, and the distance to the target PMF,

∣δϕL(q)∣≈∣B(L)−Btarget(L)∣,(24)|\delta\phi_L(q)|\approx|B(L)-B_{\mathrm{target}}(L)|, \tag*{(24)}

is minimized, for any qq, as ∣B−Btarget∣|B-B_{\mathrm{target}}| is minimized.

Refinement: Weighting Function

In practice, the weighting function ξ(Δk)\xi(\Delta k) in equation (13) is defined by considering the composition of two windows of Δk\Delta k values: a narrower “design” window centered around Δk0\Delta k_0 with uniform weights ξ=1\xi=1 (gray area in Figure 8), and a wider window with smaller weights ξ=0.1\xi=0.1 used to suppress the background at larger ∣Δk−Δk0∣|\Delta k-\Delta k_0| values. The narrower window is set in order to contain a certain fraction of the total target power, Ptarget=∫Δk∣ϕtarget(q)∣2 dqP_{\mathrm{target}}=\int_{\Delta k}|\phi_{\mathrm{target}}(q)|^2\,dq, for example 99 % of PtargetP_{\mathrm{target}}, whereas the broader window can be defined in terms of the narrower one, for example, using twice its range. Within these windows, a grid of NqN_q phase mismatch values is sampled uniformly, using at least 4 samples for every 2π/L2\pi/L interval of qq.

Figure 8. Refinement of the designs. The panels compare ϕ(q)\phi(q) of the tracked and refined structures against ϕtarget(q)\phi_{\mathrm{target}}(q), as a function of the detuning qq, for (a) the OCT default design and (b) a design targeting four Gaussians of equal height in a 5 mm crystal. Upper panels: PMF amplitude. Lower panels: complex deviation ∣δϕ∣=∣ϕ−ϕtarget∣|\delta\phi| = |\phi- \phi_{\mathrm{target}}|. All amplitudes are normalized to the maximum of the target. The grey shaded area corresponds to uniform weights ξ=1\xi= 1, the remaining plotted range to uniform weights ξ=0.1\xi= 0.1 (details in E 3).

Refinement: step-by-step procedure

  1. Start from the tracked structure and define the narrow and wide phase mismatch windows, as well as the value of NqN_q; examples: narrow window containing 99 % of the target power, wide window twice the narrow window, Nq=1201N_q=1201.

  2. Run over the walls and, for each wall, try displacements ±k⋅lres\pm k\cdot l_{\mathrm{res}}, with k∈{1,2,3,5,8,13,21}k\in\{1,2,3,5,8,13,21\}; skip displacements that violate the lmin⁡l_{\min} constraint.

  3. Accept the best displacement if it reduces JJ by more than the threshold.

  4. Evaluate the effect of deleting adjacent wall pairs.

  5. Repeat until a full iteration leaves the structure unchanged.

  6. Check the result on a grid of twice as many frequencies.

Refinement: Simulation results

We examine here the refinement of the two designs using 30 mm crystals (OCT default, as discussed in section II and OCT demodulated, as discussed in section IV A) and two designs using 5 mm crystals (Single-Gaussian PMF and a PMF with a comb of four equidistant Gaussians). Each design is refined against the ideal target PMF (see Fig. E1). Besides the objective JJ of equation (13), normalized to the weighted target power, JrelJ_{\mathrm{rel}}, we report the results in terms of the infidelity

1−F=1−∣∑Δkϕtarget∗(Δk)ϕ(Δk)∣2∑Δk∣ϕtarget(Δk)∣2∑Δk∣ϕ(Δk)∣2,(25)1-F=1-\frac{\left|\sum_{\Delta k}\phi_{\mathrm{target}}^*(\Delta k)\phi(\Delta k)\right|^2}{\sum_{\Delta k}|\phi_{\mathrm{target}}(\Delta k)|^2\sum_{\Delta k}|\phi(\Delta k)|^2}, \tag*{(25)}

which is insensitive to the overall scale of ϕ\phi, and the power ratio

PPtarget=∑Δk∣ϕ∣2∑Δk∣ϕtarget∣2,(26)\frac{P}{P_{\mathrm{target}}}=\frac{\sum_{\Delta k}|\phi|^2}{\sum_{\Delta k}|\phi_{\mathrm{target}}|^2}, \tag*{(26)}
DesignJrelJ_{\mathrm{rel}}1−F1-FP/PtargetP/P_{\mathrm{target}}DomainsIterationsTime (s)
OCT, defaulttracked2.1×10−22.1 \times10^{-2}1.9×10−21.9 \times10^{-2}0.9793481
refined4.5×10−34.5 \times10^{-3}2.4×10−32.4 \times10^{-3}0.94334693625
OCT, demodulatedtracked5.4×10−35.4 \times10^{-3}3.6×10−33.6 \times10^{-3}1.0033619
refined1.7×10−31.7 \times10^{-3}4.9×10−44.9 \times10^{-4}0.97836194130
Gaussiantracked3.3×10−53.3 \times10^{-5}3.1×10−53.1 \times10^{-5}1.000804
refined5.2×10−65.2 \times10^{-6}3.5×10−63.5 \times10^{-6}1.000804191
Four Gaussianstracked3.3×10−33.3 \times10^{-3}3.0×10−33.0 \times10^{-3}1.0041017
refined9.8×10−59.8 \times10^{-5}4.6×10−54.6 \times10^{-5}0.9991017261

Table 1. Refinement of the designs: OCT default (section II) and demodulated (section IV A) using (L=30 mmL = 30\,\mathrm{mm}, lmin⁡=4 μml_{\min} = 4\,\mu\mathrm{m} and lres=125 nml_{\mathrm{res}} = 125\,\mathrm{nm}). All values are evaluated on a validation grid with twice as many sampling points as used by the optimizer. The single Gaussian and the four-Gaussian PMFs are designed with L=5 mmL = 5\,\mathrm{mm}, envelope σχ=L/6.04\sigma_{\chi} = L/6.04 (see reference [17]), and lc=23 μml_c = 23\,\mu\mathrm{m}.

which give the complementary scale value, i.e., the brightness. Both metrics are evaluated inside the design window, and all the obtained results are summarized in Table I.

For the OCT design, shown in Figure 8(a), refinement reduces JrelJ_{\mathrm{rel}} by a factor of about 5 and the infidelity within the design window by a factor of about 8. The deviation from the target PMF is lower on average over the whole range of detunings qq (see lower insets). The number of domains is almost unchanged. Starting from the demodulated tracking gives the lowest final error (Jrel=1.7×10−3J_{\mathrm{rel}} = 1.7 \times10^{-3} against 4.5×10−34.5 \times10^{-3} for the default tracking). The lower error comes with a reduction of the power in the design window of about 2.4% to 3.6%. Lower spectral error therefore does not imply unchanged brightness, and the weights ξ(Δk)\xi(\Delta k) or widths of the detuning windows balance the two, potentially necessitating some tuning according to the application.

A target with several spectral features benefits more from the use of refinement. We consider here a comb of four Gaussians of equal height, obtained from a Gaussian envelope of width σχ=L/6.04\sigma_{\chi} = L/6.04 modulated by cos⁡(aΔz)+cos⁡(3aΔz)\cos(a\Delta z) + \cos(3a\Delta z), with Δz=z−L/2\Delta z = z - L/2 and a=3/σχa = 3/\sigma_{\chi} (normalized to max⁡∣gtarget∣=2/π\max|g_{\mathrm{target}}| = 2/\pi). The PMF consists of four Gaussians centred at q=±aq = \pm a and q=±3aq = \pm3a, as shown in Figure 8(b).

For L=5 mmL = 5\,\mathrm{mm} and lc=23 μml_c = 23\,\mu\mathrm{m}, the tracked structure reproduces the peaks with heights decreasing for increasing qq, with heights of 1.06, 1.02, 0.96 and 0.93 times the target.

The outer peaks deviate the most, consistently with the expected qq-weighting of the tracking error in equation (10). Refinement levels all four peaks within 0.5% from the target, reducing JrelJ_{\mathrm{rel}} by a factor 34 and the infidelity by a factor of 65. The power in the design window remains within 0.2% of the target power, the number of domains is unchanged, and the refinement converges in 26 iterations, taking less than 2 s using a 13th Gen Intel(R) Core(TM) i5-13500 (2.50 GHz), with (Numba) precompilation excluded. The large tracking error of the comb therefore stems from its multiple spectral features, and it is this error that the refinement optimization helps removing.

Refinement: Efficiency

In the refinement procedure, moving a wall jj of the NN internal walls of a distance dd, from z→z+dz \to z+d, changes the PMF at every Δk\Delta k of

δjϕ(Δk)=2s∫zz+deiΔkzˉ dzˉ,(27)\delta_j\phi(\Delta k) = 2s \int_z^{z+d} e^{i\Delta k\bar{z}} \,\mathrm{d}\bar{z}, \tag*{(27)}

with s=±1s = \pm1 the sign of the domain on the left of zz, and where the coefficient 2 comes from the fact that within dd the sign of the discrete nonlinearity profile changes from −s-s to +s+s, i.e., Δg=2s\Delta g = 2s. By substituting zˉ=z+u\bar{z} = z + u,

δjϕ(Δk)=2seiΔkz∫0deiΔku du=2seiΔkz⋅I(d,Δk),(28)\delta_j\phi(\Delta k) = 2s e^{i\Delta k z} \int_0^d e^{i\Delta k u} \,\mathrm{d}u = 2s e^{i\Delta k z} \cdot I(d,\Delta k), \tag*{(28)}

the contribution can be separated into two components, as shown in the last equality, which correspond to a phase term depending only on the current wall position zz and an integral I(d,Δk)I(d,\Delta k) depending only on the displacement dd, for each phase mismatch value. Given a set of MM allowed displacements, for example, d=±k⋅lresd = \pm k \cdot l_{\mathrm{res}}, with k∈{1,2,3,5,8,13,21}k \in\{1, 2, 3, 5, 8, 13, 21\}, the algorithm computes I(d,Δk)I(d,\Delta k) only once for all combinations of dd values and NqN_q sampled phase mismatches (M×NqM \times N_q). During the evaluation of each wall’s displacement, the phase term is computed once per wall per iteration, and used for all candidate displacements. Considering all NN internal walls, the number of operations is proportional to M×Nq×NM \times N_q \times N per iteration, with the operations essentially consisting of multiplications and additions.

References

  1. [1]J. L. O’Brien, A. Furusawa, and J. Vučković, Nature Photonics 3, 687 (2009).
  2. [2]F. Flamini, N. Spagnolo, and F. Sciarrino, Reports on Progress in Physics 82, 016001 (2018).
  3. [3]C. H. Bennett, G. Brassard, C. Crépeau, R. Jozsa, A. Peres, and W. K. Wootters, Phys. Rev. Lett. 70, 1895 (1993).
  4. [4]C. Couteau, S. Barz, T. Durt, T. Gerrits, J. Huwer, R. Prevedel, J. Rarity, A. Shields, and G. Weihs, Nature Reviews Physics 5, 326 (2023).
  5. [5]N. Gisin, G. Ribordy, W. Tittel, and H. Zbinden, Rev. Mod. Phys. 74, 145 (2002).
  6. [6]P. Kok, Contemporary Physics 57, 526 (2016).
  7. [7]A. Lyons, G. C. Knee, E. Bolduc, T. Roger, J. Leach, E. M. Gauger, and D. Faccio, Science Advances 4, eaap9416 (2018).
  8. [8]C. Couteau, S. Barz, T. Durt, T. Gerrits, J. Huwer, R. Prevedel, J. Rarity, A. Shields, and G. Weihs, Nature Reviews Physics 5, 354 (2023).
  9. [9]E. Knill, R. Laflamme, and G. J. Milburn, Nature 409, 46 (2001).
  10. [10]P. Senellart, G. Solomon, and A. White, Nature Nanotechnology 12, 1026 (2017).
  11. [11]J.-F. Roch, V. Jacques, E. Wu, F. Grosshans, F. Treussart, A. Aspect, and P. Grangier, in 2006 Digest of the LEOS Summer Topical Meetings (2006) pp. 14–14.
  12. [12]D. C. Burnham and D. L. Weinberg, Phys. Rev. Lett. 25, 84 (1970).
  13. [13]J. Chen, X. Li, and P. Kumar, PHYSICAL REVIEW A 72, 10.1103/PhysRevA.72.033801 (2005).
  14. [14]P. J. Mosley, J. S. Lundeen, B. J. Smith, P. Wasylczyk, A. B. U’Ren, C. Silberhorn, and I. A. Walmsley, Phys. Rev. Lett. 100, 133601 (2008).
  15. [15]A. M. Brańczyk, A. Fedrizzi, T. M. Stace, T. C. Ralph, and A. G. White, Opt. Express 19, 55 (2011).
  16. [16]A. Pickston, F. Graffitti, P. Barrow, C. L. Morrison, J. Ho, A. M. Brańczyk, and A. Fedrizzi, Opt. Express 29, 6991 (2021).
  17. [17]T. Faleo, C. L. Morrison, R. Beignon, F. Graffitti, V. Remesh, S. Frick, A. Fedrizzi, G. Weihs, and R. Keil, Applied Physics Letters 128, 204004 (2026).
  18. [18]P. G. Kwiat, K. Mattle, H. Weinfurter, A. Zeilinger, A. V. Sergienko, and Y. Shih, Phys. Rev. Lett. 75, 4337 (1995).
  19. [19]A. Fedrizzi, T. Herbst, A. Poppe, T. Jennewein, and A. Zeilinger, Opt. Express 15, 5377 (2007).
  20. [20]A. K. Ekert, J. G. Rarity, P. R. Tapster, and G. Massimo Palma, Phys. Rev. Lett. 69, 192 (1992).
  21. [21]J. Brendel, N. Gisin, W. Tittel, and H. Zbinden, Phys. Rev. Lett. 82, 2594 (1999).
  22. [22]F. Kaneda, H. Suzuki, R. Shimizu, and K. Edamatsu, Opt. Express 27, 1416 (2019).
  23. [23]C. L. Morrison, F. Graffitti, P. Barrow, A. Pickston, J. Ho, and A. Fedrizzi, API Photonics 7, 066102 (2022).
  24. [24]V. Ansari, J. M. Donohue, B. Brecht, and C. Silberhorn, Optica 5, 534 (2018).
  25. [25]F. Graffitti, P. Barrow, A. Pickston, A. M. Brańczyk, and A. Fedrizzi, Phys. Rev. Lett. 124, 053603 (2020).
  26. [26]C. Reimer, S. Sciara, P. Roztocki, M. Islam, L. R. Cortés, Y. Zhang, B. Fischer, S. Loranger, R. Kashyap, A. Cino, S. T. Chu, B. E. Little, D. J. Moss, L. C. Caspani, W. J. Munro, J. Azaña, M. Kues, and R. Morandotti, Nature Physics 15, 148 (2019).
  27. [27]P. Many, J. A. Jaramillo-Villegas, M. S. Alshaykh, J. M. Lukens, O. D. Odele, A. J. Moore, D. E. Leaird, M. Qi, and A. M. Weiner, npj Quantum Information 5, 59 (2019).
  28. [28]F. Chiriano, J. Ho, C. L. Morrison, J. W. Webb, A. Pickston, F. Graffitti, and A. Fedrizzi, Optics Express 31, 35131 (2023).
  29. [29]H.-H. Lu, M. Alshowkan, K. V. Myilswamy, A. M. Weiner, J. M. Lukens, and N. A. Peters, Opt. Lett. 48, 6031 (2023).
  30. [30]T. Faleo, F. Steinhauser, G. Weihs, S. Frick, and R. Keil, Optica Quantum 4, 333 (2026).
  31. [31]D. S. Hum and M. M. Fejer, Comptes Rendus Physique 8, 180 (2007).
  32. [32]T. F. Weiss and A. Peruzzo, Applied Physics Reviews 12, 011318 (2025).
  33. [33]J.-L. Tambasco, A. Boes, L. G. Helt, M. J. Steel, and A. Mitchell, Optics Express 24, 19616 (2016).
  34. [34]F. Graffitti, D. Kundys, D. T. Reid, A. M. Brańczyk, and A. Fedrizzi, Quantum Science and Technology 2, 10.1088/2058-9565/aa78d4 (2017).
  35. [35]R. Shiloh and A. Arie, Opt. Lett. 37, 3591 (2012).
  36. [36]A. Leshem, R. Shiloh, and A. Arie, Opt. Lett. 39, 5370 (2014).
  37. [37]M. V. Chekhova, S. G. Germanskiy, D. B. Horoshko, G. K. Kitaeva, M. I. Kolobov, G. Leuchs, C. R. Phillips, and P. A. Prudkovskii, Opt. Lett. 43, 575 (2018).
  38. [38]J. Ho, G. Moreno, S. Brito, F. Graffitti, C. L. Morrison, R. Nery, A. Pickston, M. Proietti, R. Rabelo, A. Fedrizzi, and R. Chaves, npj Quantum Information 8, 13 (2022).
  39. [39]F. Graffitti, A. Pickston, P. Barrow, M. Proietti, D. Kundys, D. Rosset, M. Ringbauer, and A. Fedrizzi, Phys. Rev. Lett. 124, 010503 (2020).
  40. [40]F. Graffitti, P. Barrow, M. Proietti, D. Kundys, and A. Fedrizzi, Optica 5, 514 (2018).
  41. [41]A. Shukhin, I. Hurvitz, S. Trajtenberg-Mills, A. Arie, and H. Eisenberg, Opt. Express 32, 10158 (2024).
  42. [42]S. Frick, R. Keil, V. Remesh, and G. Weihs, Single-photon sources for multi-photon applications, in Photonic Quantum Technologies (Wiley, 2023) pp. 53–84.DOI
  43. [43]D. T. Reid, Journal of Optics A: Pure and Applied Optics 5, S97 (2003).
  44. [44]A. Dosseva, L. Cincio, and A. M. Brańczyk, Phys. Rev. A 93, 013801 (2016).
  45. [45]P. B. Dixon, J. H. Shapiro, and F. N. C. Wong, Opt. Express 21, 5879 (2013).
  46. [46]C. Cui, R. Arian, S. Guha, N. Peyghambarian, Q. Zhuang, and Z. Zhang, Phys. Rev. Appl. 12, 034059 (2019).
  47. [47]F. Graffitti, J. Kelly-Massicotte, A. Fedrizzi, and A. M. Brańczyk, Physical Review A 98, 053811 (2018).
  48. [48]L. E. Myers, R. C. Eckardt, M. M. Fejer, R. L. Byer, and W. Bosenberg, Opt. Lett. 21, 591 (1996).
  49. [49]V. Y. Shur, A. R. Akhmatkhanov, and I. S. Baturin, Applied Physics Reviews 2, 040604 (2015).
  50. [50]T. E. Keller and M. H. Rubin, Phys. Rev. A 56, 1534 (1997).
  51. [51]W. P. Grice, A. B. U’Ren, and I. A. Walmsley, Phys. Rev. A 64, 063815 (2001).
  52. [52]K. Zielnicki, K. Garay-Palmett, D. Cruz-Delgado, H. Cruz-Ramirez, M. F. O’Boyle, B. Fang, V. O. Lorenz, A. B. U’Ren, and P. G. Kwiat, Journal of Modern Optics 65, 1141 (2018).
  53. [53]J. D. Franson, Phys. Rev. Lett. 62, 2205 (1989).
  54. [54]M. Schlosser, L. Matheis, B. Habrich, V. Habrich, G. Weihs, S. Frick, and R. Keil (2026), manuscript in preparation.
  55. [55]D. Huang, E. A. Swanson, C. P. Lin, J. S. Schuman, W. G. Stinson, W. Chang, M. R. Hee, T. Flotte, K. Gregory, C. A. Puliafito, and J. G. Fujimoto, Science 254, 1178 (1991).
  56. [56]C. K. Hong, Z. Y. Ou, and L. Mandel, Phys. Rev. Lett. 59, 2044 (1987).
  57. [57]A. F. Abouraddy, M. B. Nasr, B. E. A. Saleh, A. V. Sergienko, and M. C. Teich, Phys. Rev. A 65, 053817 (2002).
  58. [58]M. B. Nasr, B. E. A. Saleh, A. V. Sergienko, and M. C. Teich, Phys. Rev. Lett. 91, 083601 (2003).
  59. [59]M. Okano, H. H. Lim, R. Okamoto, N. Nishizawa, S. Kurimura, and S. Takeuchi, Scientific Reports 5, 18042 (2015).
  60. [60]K. G. Katamadze, N. A. Borshchevskaya, I. V. Dyanonov, A. V. Paterova, and S. P. Kulik, Phys. Rev. A 92, 023812 (2015).
  61. [61]P. Maine, D. Strickland, P. Bado, M. Pessot, and G. Mourou, IEEE Journal of Quantum Electronics 24, 398 (1988).
  62. [62]S. Carrasco, J. P. Torres, L. Torner, A. Sergienko, B. E. A. Saleh, and M. C. Teich, Opt. Lett. 29, 2429 (2004).
  63. [63]M. Charbonneau-Lefort, B. Afeyan, and M. M. Fejer, J. Opt. Soc. Am. B 25, 463 (2008).
  64. [64]K. Fradkin, A. Arie, A. Skliar, and G. Rosenman, Applied Physics Letters 74, 914 (1999).
  65. [65]F. König and F. N. C. Wong, Applied Physics Letters 84, 1644 (2004).
  66. [66]S. Emanueli and A. Arie, Appl. Opt. 42, 6661 (2003).
  67. [67]H. Masoudi and J. Arnold, Quantum Electronics, IEEE Journal of 31, 2107 (1996).
  68. [68]P. Koelenderski, W. Wasilewski, and K. Banaszek, Phys. Rev. A 80, 013811 (2009).
  69. [69]R. S. Bennink, Phys. Rev. A 81, 053805 (2010).
  70. [70]M. Schlosser, T. Faleo, G. Weihs, R. Keil, and S. Frick, Supporting data and code for the manuscript ”deterministic nonlinear domain engineering algorithm with variable sub-coherence-length domains” (2026).

Paper details

vlnc:W_6g3mdy3q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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