The explosive growth of cloud computing services, AI/ML training on GPU clusters, and web-based data storage has created urgent demand for increased network capacity, especially for intra-datacenter (intra-DC) connectivity [1]. Space-division multiplexing (SDM) is an emerging candidate to address this capacity delivery challenge, offering a cost-effective path toward scalable intra-DC infrastructure [2]. Among SDM approaches, mode-division multiplexing (MDM) links leverage few-mode fiber (FMF) transmission, enabling the simultaneous transmission of independent signals through distinct spatial modes of the fiber [3]. The key challenge in intra-DC MDM links is achieving efficient multiplexing and demultiplexing of spatial modes while minimizing modal crosstalk and group delay variations. Although multiple-input, multiple-output (MIMO) digital signal processing (DSP) is being developed to mitigate modal crosstalk impairments, it introduces considerable complexity (requiring coherent reception and additional power, processing delay time, and cost) making it incommensurate with intra-DC requirements [4]. In long-reach MDM networks (inter-DC, metro, national, and intercontinental scales), MIMO processing becomes necessary due to potential mixing between mode groups within the fiber channel itself. For short-reach communication links requiring capacity multipliers, e.g., intra-DC connectivity, MDM technology can be employed provided it is available with MIMO-free implementation as well as in compact form factor in support of higher data link densities. Hence there is a need for developing compact spatial mode multiplexers that map sources (transmitters/receivers) to fiber spatial modes with low crosstalk, and means for seamlessly integrating them at the micro-scale into pluggable modules or co-packaged optics next to electronic switching hardware.

Photonic lantern (PL) mode multiplexers are based on an adiabatic spatial transition between a discrete array of single-mode (SM) waveguides and a multi-mode optical waveguide, having matching mode and waveguide counts [5]. They enable efficient and nearly lossless conversion between multi-mode and single-mode domains, as necessary for mode-division multiplexing (MDM) optical communication systems [6–9]. PL multiplexers can be designed to be mode-preserving, which places additional requirements to distinguish the SM sources from one another, breaking the modal degeneracy throughout the lantern’s transition. This is achieved by designing the input SM waveguides to be dissimilar, resulting in distinct propagation constants that are mapped via momentum-matching to coupling to distinct spatial modes [10].

PLs can be manufactured using various methods. These include tapering SM fibers placed in a lower-index capillary to form a few-mode fiber (FMF) [11–14], waveguide inscription in glass via direct laser writing [15,16], integration within photonic circuits [17,18], and, more recently, 3D nano-printing, which enables the fabrication of micro-scale PL devices [19–21]. The 3D printing approach results in polymer-based high-index-contrast waveguides with air cladding, offering precise control over their geometry and dimensions at the sub-micrometer scale. Unlike traditional PLs, which are typically several millimeters to centimeters long and rely on low-index-contrast waveguides to achieve adiabatic transitions, 3D-printed PLs are highly compact, with lengths reduced to a few hundred micrometers. This advancement and the ability to print on various types of surfaces and light sources significantly enhance integration capabilities and pave the way for next-generation spatial multiplexing solutions and their implementation within demanding intra-DC applications.

Another promising approach for compact spatial multiplexers is multi-plane light conversion (MPLC). MPLC utilizes a sequence of phase masks to transform an array of single-mode beams into a desired modal basis, enabling highly complex spatial transformations [22–24]. Traditionally, MPLC implementations rely on bulky free-space optical setups. Recently, metasurface (MS)-based MPLC has been demonstrated [25,26], achieving miniaturization of classical MPLC devices by at least two orders of magnitude. However, these devices suffer from high optical losses (exceeding 10 dB) and significant wavelength sensitivity.

Table 1 provides a comparison of different mode multiplexers, including all-fiber PLs, glass-inscribed PLs, air-clad fiber-based PLs, and MS-MPLC. The 3D-printed PL presented in this work exhibits IL and XT performance on par with existing solutions. Notably, its physical size is about two orders of magnitude smaller, offering a major advantage for direct integration with various platforms. This compactness eliminates the need for fiber interconnects, which can increase system footprint and introduce differential delays.Table 1.

Comparison of Different Scale Three-Spatial-Mode Multiplexer Types

In this work we design, fabricate, and characterize a mode-group-selective PL using 3D nanoprinting technology, with a total length of only 300 μm, where the transformation from sources to modes occurs over 81 μm of its length. This compact PL includes waveguides for interfacing with three SM sources and an expansion taper for matching the output beam to the accepting fiber for ease of butt coupling. An additional 150 μm long taper was printed on the facet of a three-mode fiber to complete the mode matching between the PL and the fiber.

A mode group multiplexed communication experiment using intensity modulation/direct detection (IM/DD) was conducted using two spatial channels operating at the same optical wavelength, demonstrating the ability to recover the information without the use of MIMO processing.

The capability to implement MIMO-free communication using 3D-printed photonic lanterns opens new possibilities for short-reach fiber communication systems. By enabling MDM in a highly compact form, these devices significantly reduce the footprint of multiplexers. Moreover, their compatibility with 3D nanoprinting allows direct fabrication onto photonic integrated circuit (PIC) transceivers [27] or dense, micron-scale VCSEL arrays [28].

We utilize three SM cores sourced from a seven-core fiber as the single-mode inputs to our PL, with cores having an MFD of 6 μm and 35 μm pitch. The three cores are interfaced to a three-mode step-index fiber having low NA and core diameter of 22 μm. The 3D-printed waveguides exhibit a high refractive index contrast between polymer core $(n_{\mathrm{core}}=1.53)$ and air cladding ($n_{\mathrm{clad}}=1$); therefore the single-mode and the three-mode waveguide diameters are 1 μm and 1.6 μm, respectively, at vacuum wavelength ${\lambda }_0=1.55\mathrm{\mu m}$ ($V=2.35$ and 3.75). The PL contains three parts [Fig. 1(a)]: waveguides to convert between the source core modes and positions to multiplexer inputs, three-mode-selective multiplexer (Mux), and a 50 μm long expanding taper, from the Mux’s output waveguide (1.6 μm) to 10 μm diameter.

To match between the PL and the three-mode fiber we print an ancillary 150 μm long taper on the fiber, starting with a 10 μm diameter (PL’s output) and ending with 28 μm diameter. From simulations, a 28 μm diameter circular polymer-air waveguide, is optimal for coupling to our three-mode low-NA fiber with simulated coupling loss $<-0.22\mathrm{dB}$ and no crosstalk between the two mode groups. We choose to print the 150 μm long taper on the fiber to mitigate mechanical instability in the PL, as a taper mass at the PL end can lead to bending. In principle, system robustness could be improved—and the need for a printed taper eliminated—by using a smaller-core 3-MF or a tapered 3-MF [31] (which was not available).

Each segment of the PL system has been designed separately. In order to design a mode-selective Mux, the input SM waveguides of the PL are arranged in a right angled isosceles triangle [Fig. 1(a)], with the two waveguides at the acute angles having the same diameter $D_G^{(2)}$ and destined to excite the two modes of the ${LP}_{11}$ mode group and the waveguide at the right angle corner of diameter $D_G^{(1)}$ for fundamental mode ${LP}_{01}$ excitation, where $D_G^{(2)}<D_G^{(1)}$. Due to the high-index-contrast waveguides, mode matching from the SM sources to the mode size of the Mux’s inputs is required. Tapering down the waveguides from the source (8.4 μm diameter) to an SM diameter (1 μm) and then expanding them to a three-mode diameter (1.6 μm) will require a long structure with very small diameter waveguides and in practice the fabricated structure would be too fragile and mechanically unstable. Instead, the source modes are tapered down to the multiplexer input diameters $D_G^{(1)}$ and $D_G^{(2)}$, and continue reducing gradually to the three-mode diameter at the Mux’s output. The challenge here is preventing the excitation of higher-order modes. The mode matching taper [“Three sources interface” in Fig. 1(a)] is optimized to output the fundamental mode with high purity (96%). Next, the Mux is optimized assuming SM inputs. Designing of the Mux requires time consuming finite-difference time-domain (FDTD) simulations. To calculate the insertion loss (IL), crosstalk (XT), and mode dependent loss (MDL) a $6\times 6$ coupling matrix needs to be calculated, and requires performing six FDTD simulations (dimension 6 originates from three spatial modes and two polarizations). By calculating the singular values of the matrix $({\lambda }_i)$ we can extract the IL and MDL by $$\mathrm{IL}=10·{\log }_{10}\frac{1}{N}\sum _{i=1}^N{|{\lambda }_i|}^2,$$(1)$$\mathrm{MDL}=10·{\log }_{10}\frac{{|{\lambda }_{\min }|}^2}{{|{\lambda }_{\max }|}^2}.$$(2)

The $\mathrm{X}\mathrm{T}$ is defined as the sum of block diagonal elements in the $6\times 6$ power coupling matrix $A$, divided by the sum of all other elements: $$\mathrm{XT}=10·{\log }_{10}\left(\frac{S_{\text{block}}}{S_{\mathrm{off}-\text{block}}}\right),$$(3)where $$S_{\text{block}}=\sum _{i=1}^2\sum _{j=1}^2{A}_{\mathit{ij}}+\sum _{i=3}^6\sum _{j=3}^6{A}_{\mathit{ij}}$$(4)is the sum of the elements in the selected block diagonal regions (which accounts for energy remaining in the mode groups), and $S_{\text{total}}$ is the sum of all elements in the matrix. The sum of the remaining elements is $$S_{\mathrm{off}-\text{block}}=S_{\text{total}}-S_{\text{block}}.$$(5)

Due to the Mux symmetry, we devised an objective function for minimizing both XT and IL and reduced the number of simulations to two only per optimization cycle. Sampling at a random input source polarization mode from group 1 ($G^{(1)}$), we launch the sampled mode through the ${\mathit{In}}_1$ input waveguide [Fig. 1(a)] and calculate the coupling coefficients of the Mux’s output field with its six eigen modes. Let $\mathrm{P}{\mathrm{G}}^{(1)}$ be the sum of the overlap coefficients of $G^{(1)}$ (total output power transmitted to mode group 1) and let $P_1$ be the sum of all six coefficients (total output power). Similarly, $\mathrm{P}{\mathrm{G}}^{(2)}$ and $P_2$ are generated by launching an input sampled mode source from one of the four group 2 ($G^{(2)}$) input waveguides ($i{n}_2$ or $i{n}_3$) and polarization. The objective function is defined by $$\widehat{F}=\frac{1}{1+\lambda }·\frac{2\mathrm{P}{\mathrm{G}}^{(1)}+4\mathrm{P}{\mathrm{G}}^{(2)}}{6}+\frac{\lambda }{1+\lambda }·\frac{2P_1+4P_2}{6}\le 1.$$(6)

The first term of $\widehat{F}$ targets crosstalk reduction and the second term maximizes the total power transmission. For small $\lambda$ ($0\le \lambda \le 1$), the optimization will be biased towards crosstalk reduction and efficiency may be low. After a few experiments we chose $\lambda =0.4$. $D_G^{(2)}$, $D_G^{(1)}$, Mux length (L), and 20 additional path defining parameters are optimized with a genetic algorithm (GA) to maximize our objective function $\widehat{F}$. The optimization flow is described in Refs. [19,20].

We performed the GA optimization procedure, improving the objective metric $\widehat{F}$ at every generation [Fig. 1(b)] resulting in $D_G^{(1)}=2.2\mathrm{\mu m}$ and $D_G^{(2)}=2.1\mathrm{\mu m}$, and an effective index difference of $6.6\times {10}^{-3}$ between the input waveguides to distinguish the mode groups. The optimization process progressed through 13 generations, with each generation consisting of 80 unique designs. Since each design underwent two simulations, this amounted to a total of $2\times 80\times 13$ FDTD simulations. Given that each simulation lasted approximately 2 min, the entire optimization procedure required around 70 h. The best design for the Mux achieved $\mathrm{I}\mathrm{L}=-0.22\mathrm{dB}$ and $\mathrm{X}\mathrm{T}=-23\mathrm{dB}$. After including the source interface waveguides and the output tapers, the complete PL system simulation achieves $\mathrm{I}\mathrm{L}=-0.6\mathrm{dB}$ and $\mathrm{X}\mathrm{T}<-21\mathrm{dB}$. Figure 1(c) presents the simulated output fields of the designed PL+taper system for the three input waveguide excitations. The resulting output fields align well with the design objectives: excitation of the ${In}_1$ waveguide produces an ${LP}_{01}$ mode, while excitation of ${In}_2$ and ${In}_3$ yields fields of the ${LP}_{11}$ mode group. Figure 1(d) displays the IL and XT of the simulated system across the 1.5–$1.6\mathrm{\mu m}$ wavelength band, showing uniform IL around $-0.6\mathrm{dB}$ and XT ranging from $-20.5$ to $-21.8\mathrm{dB}$.

The PL design is defined in 3D-CAD and imported to the Nanoscribe Photonics Professional GT printer. A prototype PL was initially fabricated on a glass substrate [Fig. 2(a)]. Multiple devices were fabricated to optimize the printing parameters (laser power, writing speed, hatching, and slicing distance). Then the device was fabricated directly on a seven-core fiber (using only three of its cores) with 7-SMF fanout (made by Chiral-Photonics [32]). The ancillary 150 μm long taper to match the PL to three-mode fiber [Fig. 2(b)] was printed on the fiber facet. Figure 2(c) shows a microscope image of the mutually aligned PL and taper in the optical setup. For a detailed description of the process, see Section 5.

We first conducted an optical power transmission measurement, as illustrated in Fig. 3(a). Utilizing an optical switch and the SM fiber fan-in, we individually excited the PL’s inputs. The optical power exiting the output taper was measured using a $1{\mathrm{cm}}^2$ InGaAs power sensor. Subsequently, we butt-coupled the PL to the optical fiber using $x,y,z$, and tip/tilt stages. Manual alignment was performed to maximize the fiber output power with care to prevent collision. Figure 3(b) illustrates the transmission power measurement from each PL input, both with and without coupling to the optical fiber across a wavelength range of 1530–1590 nm. The most efficient curve corresponds to the ${LP}_{01}$ input, exhibiting an average loss of $-1.4\mathrm{dB}$ without fiber coupling and $-2.6\mathrm{dB}$ with fiber coupling. For inputs ${In}_2$ and ${In}_3$, corresponding to the ${LP}_{11}$ mode group, average losses of $-1.6\mathrm{dB}$ and $-1.8\mathrm{dB}$ were measured without fiber, and $-2.9\mathrm{dB}$ and $-3.2\mathrm{dB}$ with fiber, respectively. A transmission fiber coupling loss of less than 2 dB was observed. The contributing sources to this degradation include taper losses, coupling losses between the taper and 3-MF, and misalignment, tilt, and finite air-gap between the PL and taper.

We then introduce a polarization scrambler to the input to assess PDL. By randomly scrambling the polarization state, we excited various polarization states across the Poincare sphere. Power levels were logged, for each wavelength, and PDL was computed using the formula $10·{\log }_{10}(\frac{P_{\min }}{P_{\max }})$, where $P_{\min }$ and $P_{\max }$ represent the minimum and maximum recorded powers, respectively. This evaluation was performed for each of the PL’s inputs, excluding the fiber system. The results in Fig. 3(c) consistently reveal PDL values between 0.2 and 0.3 dB across the entire wavelength spectrum.

To better understand the PL system’s performance, an analysis of its output field and transfer matrix measurement is necessary. We capture the system output fields, $E_x$ and $E_y$ (denoting the two orthogonal polarization states), using off-axis digital holography [33], for each input mode excitation. Figure 4(a) presents the complex output fields reconstructed using a digital holography technique at $\lambda =1.55\mathrm{\mu m}$ from all six input modes (three spatial inputs with two polarization states). Inputs $I{n}_1^{x/y}$ excite the ${LP}_{01}$ mode group, and as anticipated, both complex field and intensity profiles resemble a Gaussian mode. Similarly, $I{n}_{2/3}^{x/y}$ exhibit mode profiles corresponding to the $L{P}_{11}^{a/b}$ mode group. Through calculated modal decomposition with the three target modes (six modes considering polarization), the coupling matrix was determined. This method emulates an ideal demultiplexing scenario, while enabling the assessment of the system’s modal quality and evaluation of IL and XT. We conducted measurements of the coupling matrix across a wavelength range of 1520–1600 nm. Figure 4(b) shows the power coupling matrix over the measured wavelength range. Since the PL device is mode group selective and due to inter-mode-group mixing within the three-mode fiber, we expect a block diagonal matrix with size $2\times 2$ for the fundamental spatial mode and its two polarizations, and of size $4\times 4$ for the LP₁₁ spatial mode and its two-fold spatial degeneracy and polarization states.

For better visualization of the mode group selectivity, a mode group averaged matrix is shown in Fig. 4(c), which clearly shows high values of power along the diagonal elements and low power in the XT elements over the entire wavelength range.

From the complex coupling matrix measured by modal decomposition between the measured fields and the digital modal basis, we computed the XT for each wavelength and the IL of all modes and each mode group separately using SVD. The results are presented in Fig. 4(d), where the IL remains stable across the wavelength range, with an average IL of $-1.05\mathrm{dB}$ (minimum value of $-1.17\mathrm{dB}$). The insertion loss for mode group 1 remains below $-1\mathrm{dB}$ across most of the bandwidth, facilitating polarization diversity coherent detection. Additionally, the low insertion loss of mode group 2 (approximately $-1\mathrm{dB}$) could enable a mode-group-based MDM system [34], further enhancing data capacity. Due to the normalization of the overlap integrals, the IL represents the fraction of transmitted power directed toward the system’s targeted three modes. In this measurement, the IL represents the fraction of power remaining in the target modes after propagation through the system. The total transmitted power and optical loss of the device were measured in the preceding section. The average XT is $-17.65\mathrm{dB}$ and is consistently less than $-16\mathrm{dB}$. The device was optimized for operation at a wavelength of 1.55 μm, with the XT at this wavelength recorded as $-18.6\mathrm{dB}$. It is worth noting that the simulated XT value for the PL device was $-21\mathrm{dB}$. The observed 2.4 dB discrepancy could be attributed to fabrication errors and misalignment of the PL with the taper.

Next, we coupled into and out of a three-mode fiber with two PLs, one serving as a Mux and one as Demux. We optimized the butt-coupling at both ends of the fiber utilizing three photo-detectors to record the output power of each demultiplexed PL channel [Fig. 5(a)]. During the active alignment process, we continuously monitored both crosstalk and total transmitted power, aiming to optimize these parameters using two motorized $x,y,z$ and tip/tilt stages. After alignment, the system remained stable for several minutes; beyond that, fine-tuning of the stages was required to maintain optimal performance on account of creep. Figure 5(b) depicts the power transmission matrix at best alignment (using a wavelength of 1543 nm). Each of the matrix elements represents the fraction of transmitted power measured. XT from the ${LP}_{01}$ (${In}_1$) channel to ${LP}_{11}$ (${In}_2$ and ${In}_3$) was $-11.7\mathrm{dB}$, $-8.5\mathrm{dB}$ from ${LP}_{11}$ (${In}_2$) to ${LP}_{01}$, and $-9.4\mathrm{dB}$ from ${LP}_{11}$ (${In}_3$) to ${LP}_{01}$. These values are reasonable, considering that XT measurement of a Mux-PL coupled to a fiber in the previous section indicated $-18\mathrm{dB}$.

Finally, we conducted a mode group multiplexed IM/DD communication experiment, transmitting two on-off keying (OOK) data channels on the ${LP}_{01}$ and ${LP}_{11}$ modes through a 1 m long three-mode fiber. The system comprises a PL Mux/Demux pair, each of which is butt-coupled to 3-MF aided by the printed taper.

The experimental setup is illustrated in Fig. 6(a). Two channel transmitters were employed, each consisting of an external cavity laser (ECL) set to the same wavelength ($\lambda =1543\mathrm{nm}$) and a Mach-Zehnder modulator (MZM). A pattern generator operating at 12.5 Gb/s provided OOK data and complementary signal outputs, which were each sent to the single-sided MZMs. The modulated signals were amplified using an erbium-doped fiber amplifier (EDFA) to boost transmission power and 10 ns decorrelation delay was applied. Due to variations between the modulators and lasers in the two channels, the EDFA gains were adjusted to equalize the output power (25 mW) of each transmitter channel. At the receiver, the ${LP}_{01}$ channel was detected directly at ${\mathrm{Out}}_1$ using a photo-detector (PD). The two ${LP}_{11}$ channels ${\mathrm{Out}}_2$ and ${\mathrm{Out}}_3$ were combined at orthogonal polarizations and measured with another PD.

The combination of the two ${LP}_{11}$ channels utilized two polarization controllers (PCs) and three polarization beam splitters/combiners (PBS/Cs). The PCs were optimized to minimize the dropped output at one polarization state of the first PBS, ensuring that the two ${LP}_{11}$ outputs were set in orthogonal linear polarization states. Both signals having identical path lengths were then combined using a PBC. This combining method enables a theoretically lossless signal combination through a unitary transformation and is feasible due to the single-polarization modulation (OOK) at the transmitters. However, practical losses arose from imperfections in the PBS/C components, fiber connectors, and slow polarization state drift. This approach avoids the inherent 3 dB loss associated with simple beam combiners. The signals from both channels were optically filtered using a narrow (0.3 nm bandwidth) WDM filter around $\lambda =1543\mathrm{nm}$ (Lumentum DWS-1F3323L93) and analyzed using a bit error rate tester (BERT) (model J-BERT N4903A) evaluating the eye diagrams and BER.

The transmitted signal was a $2^{23}-1$ pseudo-random binary sequence (PRBS) operating at 12.5 Gb/s. The ${LP}_{01}$ channel, as shown in Fig. 6(b), displays an eye opening of 140 mV alongside a BER of $6\times {10}^{-8}$, with an optical power of 6.5 mW reaching the detector (system loss of $-5.9\mathrm{dB}$). The ${LP}_{11}$ channel, depicted in Fig. 6(c), exhibits a smaller eye opening of 115 mV with a BER of $1.4\times {10}^{-7}$ and an optical power of 4.7 mW the detector (loss of $-7.3\mathrm{dB}$). It is worth noting that the ${LP}_{11}$ channel demonstrates the poorer performance, because this channel is suffering from an additional 1.4 dB of loss relative to the ${LP}_{01}$.

A system OSNR compared to BER was evaluated. An ASE source and a variable optical attenuator (VOA) were used to introduce noise loading. OSNR was measured for each attenuation level using an optical spectrum analyzer, while BER was assessed utilizing the BERT instrument. The evaluation of both channel systems, including the modulator, detector, EDFA, and other components, was conducted using SMF, with the BER versus OSNR serving as the baseline performance. The two SMF transmitter channels exhibit nearly identical performance, as shown in Fig. 6(d). Each mode channel was assessed individually and in mode-multiplexed transmission [as depicted by the “${LP}_{01/11}$ only” curves in Fig. 6(d), illustrating the XT impact when the other channel is present]. The measurement results presented in Fig. 6(d) indicate superior performance of the ${LP}_{01}$ channel compared to the ${LP}_{11}$ channel primarily due to its lower losses and reduced XT measured in the PL system’s input/output, relative to the ${LP}_{11}$ channel. We extract from the measurement the power penalty for each channel relative to the SMF channel at $\mathrm{BER}{=10}^{-6}$ (Table 2), with mode group multiplexing penalty lower than 2 dB. This outcome aligns with expectations, as the latter is subject to higher XT and IL.Table 2.

Power Penalty for Each Channel at BER ${=10}^{-6}$

We designed, fabricated, and characterized a diminutive, 3D-printed three-mode-selective photonic lantern spatial multiplexer with a total length of 300 μm, utilizing high-refractive-index core-cladding contrast direct laser writing (DLW) waveguides. We successfully integrated this device with a large core 3-MF and conducted a comprehensive system experiment involving two PLs serving as Mux/Demux and a 3-MF equipped with mode-matching tapers at both ends. The PL device exhibited average power losses of $-1.35\mathrm{dB}$ without fiber coupling and $-2.61\mathrm{dB}$ with a 3-MF, indicating efficient power transfer within the PL. Additionally, PDL was found to be less than $-0.25\mathrm{dB}$, and modal XT was below $-16\mathrm{dB}$ across a wavelength range of 1520–1600 nm. Due to its compact size, the device demonstrates minimal wavelength dependent effects, ensuring nearly uniform performance across the measured wavelength range. The ability to realize MIMO-free communication with 3D-printed photonic lanterns introduces new opportunities for short-reach, ultra compact, fiber-optic communication links for intra-DC applications. Additionally, their compatibility with 3D nanoprinting technology allows for direct integration with PIC transceivers or dense, micron-scale VCSEL arrays.

Our simulations suggest that the IL of the PL system (PL+taper printed on the 3-MF) can be as low as 0.6 dB including fiber coupling. We experimentally observed PL losses (measured without a fiber) below 2 dB, suggesting further reduction is achievable by optimizing the fabrication process—namely, accurate source alignment, minimizing surface roughness, and mitigating waveguide shrinkage. At the system level, transmission losses can be reduced through improved fiber coupling strategies. For practical implementation, both the PL multiplexer and demultiplexer can be printed directly within optical transceivers, converting between multiple SM sources, likely on a silicon-photonic integrated circuit, and FM-fiber ends using the PL disposed in between, as currently being suggested for photonic wire bonds [35,36]. This approach eliminates the need for active fiber alignment, which can otherwise degrade system performance. The XT can theoretically be improved to $-21\mathrm{dB}$ per device, with system-level crosstalk largely dictated by the fiber coupling interface. While the experimental system exhibits greater XT than the theoretical limit, it already supports mode group multiplexed, IM-DD communication experiments for short-reach transmission, demonstrating a doubling of system capacity through spatial multiplexing without any need for MIMO processing overhead.

Future work will focus on refining the fabrication process, improving fiber alignment techniques, and increasing and integrating the PL with photonic-integrated-circuit-based systems.

The electromagnetic simulations carried out using Ansys Lumerical solvers encompass the utilization of FDTD, EME, and FDE tools. The optimization and design process involves utilizing a Python API wrapper that has been developed in house for the Lumerical solver.

A Nanoscribe Photonic Professional GT printer was utilized in the fabrication process, employing an IP-Dip photoresist, which is specifically formulated to match the refractive index of the microscope objective ($63\times$) that focuses the laser beam. This results in ideal laser beam focusing and fine lateral resolution (200 nm) of the fabricated structures [37]. Prior to fabrication, a silanization process was used to enhance polymer adhesion to the optical fiber facet (silica). The fabricated structure was developed in PGMEA (propylene glycol monomethyl ether acetate) for 20 min, then cleaned with IPA (isopropyl alcohol) for 2 min, and dried with Novec 7100 for 1 min. To achieve the desired shape and size of the waveguide with minimum surface roughness, we optimized various writing parameters, such as laser power set to 35% of the maximum (Nanoscribe’s laser), scanning speed of 10,000 μm/s, and the distance of laser scanning hatching set to 0.05 μm (lateral spacing between lines) and slicing set to 0.1 μm (spacing between layers in the $z$-axis). The fabrication time for the PL structure is approximately 20 min.

To measure the coupling matrix of the PL, we employed off-axis digital holography to capture the complex electric field of the device’s output. Using two orthogonally polarized reference beams ($x$ and $y$ polarized) enables capturing the two interference field components simultaneously, denoted as $E_x$ and $E_y$, respectively. Using a polarization controller before the PL, we launch two orthogonal input modes to each of the PL inputs depicted as $I{n}_i^x$ and $I{n}_i^y$, where $i\in [\mathrm{1,2,3}]$. We then performed modal decomposition (MD) using six digital modes that were simulated and supported by the three-mode PL’s output (with polarization). In our experimental setup, we utilized the following equipment: a Yenista optics model $-1560$ ECL (external cavity laser) source, Thorlabs MPC320 polarization controllers, a Thorlabs TC25FC-1550 fiber collimator for collimating the reference beam, a $20\times$ T1.1 Mitutoyo LCD Plan Apo NIR infinity corrected objective with a 200 mm tube lens for focusing the output beam of the PL onto the camera plane, and an Allied Vision Goldeye G-033 TECless InGaAs camera.

Figure 3(a) illustrates the configuration used to measure the power transmission efficiency. The setup included a Yenista optics model $-1560$ tunable ECL source, which operated across the wavelength range of 1520–1600 nm, and an HP-8153A optical power meter. To account for losses caused by optical devices such as the optical switch and fanout, we initially measured the power transmission of the entire system using the seven-core fiber, excluding the PL structure. This measurement was performed on each core, resulting in a total of three baseline measurements for the system. Each measurement vector covered the required wavelength range. We denote the measurement of core $i$ as $\mathrm{P}{\mathrm{S}}_i$.

After fabricating the PL on the measured fiber, we repeated the same measurement procedure, measuring the power emerging from the PL device. We denote these measurements as $\mathrm{P}\mathrm{P}{\mathrm{L}}_i$. The loss measurement of the PL per input $i$ [as shown in Fig. 3(a)] was calculated using the normalization formula and converted to dB scale: $$\mathrm{P}{\mathrm{L}}_i=10·{\log }_{10}\left(\frac{\mathrm{P}\mathrm{P}{\mathrm{L}}_i}{\mathrm{P}{\mathrm{S}}_i}\right)[\mathrm{dB}].$$(7)

The experimental arrangement is depicted in Fig. 6(a). In our setup, we employed the Agilent J-BERT 12.5 Gb/s for the pattern generator, oscilloscope, and bit error rate (BER) measurement. The Mach-Zehnder modulator utilized is the 43 Gb/s Fujitsu FTM7937EZ, and the photo-detectors are ${\mathrm{u}}^2\mathrm{t}$ BPDV2150R, supporting at up to 40 Gb/s.

Acknowledgment. Y.D., A.K., and D.M.M. thank the Israel Innovation Authority for funding parts of this work via the VCSEL Consortium. Y.D. and D.M.M. thank the Peter Brojde Laboratory for Miniature Integrated Systems for 3D nanoprinting on the Nanoscribe tool.