Figure 1 presents a schematic diagram of a genetic MMF laser leveraging wavefront shaping. The MMF laser can be three-dimensionally mode-locked through STML with strong spatial and spectral filtering in addition to high nonlinearity, gain and spatiotemporal dispersion12, wherein ultrashort pulse generation is initiated and stabilized by nonlinear polarization evolution (NPE)25. In brief, the MMF laser has a ring cavity, where the laser beam propagates clockwise. The gain medium is a piece of double-cladding ytterbium-doped MMF (Thorlabs YB1200-25/250DC, ~3 m long) that supports approximately 13 modes at 1060 nm (Table S1). This multimode gain fibre has a core diameter of ~25 µm, a core/cladding numerical aperture (NA) of 0.07/0.48, and a cladding-pump absorption of 2.3 dB/m at 920 nm. Using a cladding-pump scheme, the multimode gain fibre in an optical amplifier configuration can provide a slope efficiency of > 75%. The laser system is pumped by a 976-nm multimode laser diode (27 W maximum power) through a signal–pump combiner (SPC). The lead fibres of the SPC (CorActive DCF-UN-25/250-08), which have core/cladding diameters of 25/250 μm, are designed to match with the gain fibre. A piece of the multimode graded-index (GRIN) fibre (Thorlabs GIF625, 62.5 µm core diameter, ~1 m long) is fusion-spliced to the gain fibre with a core offset (approximately 20 µm, as indicated in Fig. 1) to excite the higher-order modes. Please note that the GRIN fibre, which is designed with reduced modal dispersion of a magnitude comparable to its chromatic dispersion, can support hundreds of transverse modes and is key for successful STML12.
The output of the GRIN fibre is collimated by a lens (L1) and launched into free space. A polarization-maintaining isolator (ISO) is utilized to establish a polarization-dependent transmission mechanism that is associated with the intensity-dependent fibre nonlinearities for NPE. A 50:50 beam splitter (BS) serves as an output coupler and extracts half of the laser beam for characterization. A portion of the laser beam from the BS is launched to the wavefront shaping unit, which consists of a half-wave plate (λ/2), a 5× telescope (lenses L2 and L3) and a fast phase-only SLM (Holoeye PLUTO-2-NIR, 1920 × 1080 pixels, 60 Hz frame rate). The half-wave plate (λ/2) is used to optimize the state of polarization (SOP) of the laser beam and thus improve the modulation efficiency of the SLM, while the telescope magnifies the laser beam to match with the active area of the SLM. The modulated laser beam is returned and subsequently coupled back into the MMF (i.e., lead fibre 1 of the SPC) through the BS, wave plates (λ/2 and λ/4) and L4. The SOP of the laser beam is further adjusted by a half-wave plate (λ/2) and a quarter-wave plate (λ/4). The extracted laser beam from the BS is detected in multiple domains, including cross-section space xy, time t, optical wavelength λ, and optical power P (via a CCD camera, photodiode, spectrometer, and power metre, respectively; see Table S2). The detected multi-dimensional signals serve as the cost values of the corresponding cost functions and are fed to the genetic algorithm for evaluating and generating the phase maps (see "Methods").
In general, this MMF laser can operate in the STML state by increasing the pump power beyond 15 W, together with an appropriate setting of the SOP12. The output optical power can be varied from tens to hundreds of mW, which depends on the SOP of the laser. The laser cavity has a fundamental repetition rate (FRR) of ~16 MHz (i.e., a round-trip time of ~61 ns). The insets of Fig. 1 show the typical mode profiles of the pump and signal beams. In the following experiments, we first investigate the optical power, mode profile, and lasing wavelength of the MMF laser in the quasi-CW state without specific optimization for mode-locking, i.e., Figs. 2–4, respectively. Please note that here, the quasi-CW state of the MMF laser refers to the operating condition without stable mode-locking but does not exclude the occasional generation of broad-pulse fluctuations (or modulations) due to random mode-beating. Then, we study the pulsing ability of the MMF laser through automatic mode-locking manipulation, i.e., Fig. 5.
Lightwaves propagating in MMFs can create complex 3D interference patterns (insets of Fig. 1). Thus, once lasing is established, the interfering lightwaves in all modes are amplified simultaneously, and complex combinations of laser modes (both transverse and longitudinal modes) compete for the limited energy available from the gain medium. As a result, the speckled laser beam experiences uneven gain in both the transverse and longitudinal dimensions, which can compromise the overall optical power. Here, the output optical power of the MMF laser working in the quasi-CW state is optimized by genetically evaluating the cost function Paverage using the reading from a power metre with an attenuation of 10 dB. Please note that all the measurements of the output optical power shown in this work are 10 dB attenuated. The optimization starts from a cavity setting where the MMF laser is adjusted to deliver the "highest" power by only rotating the wave plates. As shown in Fig. 2a, the output optical power can typically be increased by a factor of approximately 2 after approximately 300 generations of breeding, which takes approximately 4 mins. As can be observed, the genetic wavefront shaping exhibits reasonably high sensitivity, and the output power becomes > 5% above the baseline after ~30 generations of breeding. The power enhancement gradually becomes saturated after 300 generations of breeding, after which the relative change is less than 1%. Once the optical power is saturated, the phase maps of the population become closely consistent, and the phase maps exhibit a high overall cross-correlation, i.e., increasing from 0% to ~75%. Figure 2b, c shows the optical spectra and temporal signals before and after optical power optimization. Surprisingly, the optical spectrum was cleaned after genetic optimization, i.e., changing from dual bands to a single band, while both temporal signals exhibit noisy baselines without sharp pulses, which is expected since no specific optimization is present to promote mode-locking; rather, that condition is demonstrated in Fig. 5. These results also indicate that the initial lasing condition favours a multiwavelength operation, which can arise from either the alleviation of gain competition through intensity-dependent nonlinearity (i.e., the Kerr effect)26 or nonlinear coupling between the transverse modes. Figure 2d shows the mode profiles before and after the optical power optimization. As can be observed, the mode profile is slightly changed after the optical power optimization, which, to a certain extent, is expected since here the objective function is set only to optimize the optical power such that the mode profile can vary relatively freely.
The larger mode area of MMFs, on the one hand, can deliver a much greater optical power and mitigate the nonlinear effect, which sets the ultimate limit on the ultrashort pulse propagation; on the other hand, the increased mode area creates much weaker confinement in the transverse space and leads to the coexistence of multiple transverse modes, easily up to hundreds. As a result of the strong nonlinear coupling and interference among these modes, the quality of the mode profile in MMF systems is typically worse than that of the single-mode counterparts—usually evaluated by the M2 factor27. Recently, Teğin et al. demonstrated an MMF laser with M2 < 1.4 through spatiotemporal self-similar evolution28, and Krupa et al. presented laser beam cleaning in the nonlinear propagation regime of the GRIN fibre8. In general, a laser with poor (spatial) beam quality is not suitable for applications requiring diffraction-limited focusing, e.g., precision machining, high-resolution optical microscopy, and long-range metrology. To genetically improve the beam quality of a MMF laser, we set the objective function to the mean intensity of the targeted area, i.e., 〈I(x0, y0)〉, as indicated by the red dotted circle in Fig. 3a, which is measured by a CCD camera (Point Grey GS3-U3-32S4M-C). The mode profile cleaning starts from a mode profile with a random phase map addressed to the SLM. Figure 3a shows the mode profile before genetic optimization, which comprises a complex speckled pattern caused by the coexistence of many transverse modes. The optimized mode profile, fascinatingly, is much cleaner and appears as a single bright spot (Fig. 3b). The final mode pattern exhibits a Gaussian-like profile with a weak background of higher-order modes (Fig. 3c). M2 measurements are conducted to evaluate the beam quality after optimization. As shown in Fig. 3d, M2 factors of 1.17 and 1.25 are obtained for the x and y directions, respectively. The evolution of the objective value, as shown in Fig. 3e, demonstrates that the optimization of the mode profile is less sensitive than that of the optical power, and obvious intensity enhancement (by 5%) of the targeted area occurs only after ~100 generations of breeding. It is worth noting that spatiotemporal self-similar evolution28 and Kerr nonlinear cleaning8 have also been demonstrated as powerful approaches for improving the beam quality in MMF systems. Here, instead, cleaning of the mode profile in the MMF laser using wavefront shaping leads to the generation of a high-quality beam in a genetic manner. As such, a highly MMF laser can also deliver a high-quality laser beam without compromising the optical power, as shown in Fig. 3f, where the optical power exhibits likely random variation at approximately 37 mW (10 dB attenuated) with a standard deviation of ~1.1 mW.
Wavelength tunability is commonly required in different research fields, including optical component testing, fibre optic sensing, metrology, and optical spectroscopy29. To this end, we manipulate the optical spectrum of the MMF laser using genetic wavefront shaping technology. As shown in Fig. 4a, the initial lasing wavelength is approximately 1035 nm. The cost function of the genetic algorithm is evaluated by the intensity at the targeted wavelength, i.e., I(λtarget), which is measured by a spectrometer (Avantes AvaSpec-ULS3648). Please also note that the targeted wavelength can be different from or the same as the initial wavelength, while the latter case is equivalent to optimizing the optical power. We first demonstrate wavelength switching, i.e., from the initial wavelength to another optical wavelength (1060 nm in this case); see Fig. 4a, b. As shown, the optical spectrum begins to oscillate at the targeted wavelength after ~90 generations of genetic iteration. Once lasing at the targeted wavelength is established, the intensity at the original wavelength gradually decreases (Fig. 4c). Figure 4b shows the line plots of the initial and final optical spectra after optimization and clearly illustrates that a new wavelength at ~1060 nm has been established from the noise floor where no initial signal is detected prior to running the genetic algorithm, as shown in Fig. 4c.
However, the optimized optical spectrum has a residual component at the initial wavelength, resulting in a poor side mode suppression ratio (SMSR) of ~10 dB (Fig. 4b). Some specific applications require a clean optical spectrum, as unwanted wavelength components can become noise sources. Thus, strong side mode suppression is highly desired. To this end, the objective function is modified from a simple form of I(λtarget) to I(λtarget)/I(λinitial). As shown in Fig. 4d, the initial wavelength component has indeed been further attenuated, and thus, the SMSR is significantly improved. Figure 4e illustrates the optical spectra before and after enhancement of the SMSR, and an SMSR improvement of 5.6 dB is obtained. Notably, this wavelength switching capability can be further applied to perform wide range wavelength scanning. As shown in Fig. 4f, wavelength scanning can cover the entire operational bandwidth of ytterbium-doped gain media, typically from 1030 to 1070 nm. It is worth noting that the intensities of optical spectra vary across the scanning range, which can be a potential issue for applications requiring superior power stability. This weakness can potentially be mitigated by using power equalization technologies30. Please also note that the speed of wavelength scanning in the current setup is generally slow (tens of mins). If necessary, however, emerging high-speed wavelength wavefront shaping technologies31 can be employed to accelerate the speed.
Finally, the pulsing ability of the MMF laser is investigated using genetic wavefront shaping. It is also worth noting that automatic mode-locking has been intensively studied for single-transverse-mode fibre lasers32-36, while no effort has so far been devoted to MMF lasers—mainly owing to their multi-dimensional complexities, which make directly applying the traditional single-transverse-mode (1D) technologies challenging. In this case, the objective function of the genetic wavefront shaping is evaluated using the radiofrequency (RF) spectrum. In brief, the temporal signal is first detected by a photodiode (PD) and is subsequently digitized by a data acquisition (DAQ) card. The recorded digital temporal signal is converted to the frequency domain using a fast Fourier transformation (FFT), resulting in the RF spectrum of the temporal signal. Because the FRR of the MMF laser can be calculated simply from the total cavity length, the RF intensities of the FRR frequency and its higher-order harmonics, typically up to tens of orders (limited to the electronic bandwidth), are obtained. In this case, the objective function of automatic mode-locking is evaluated by measuring the RF intensity of the 10th harmonic I(f10th). Please note that other alternative objective functions can also be applied, e.g., the intensity of the FRR frequency or the sum of the intensities of all harmonics. However, our experimental study suggests that the case of the intensity of the FRR frequency provides a worse performance for successful automatic mode-locking. This may have been influenced by the weak pulsed signal generated by the beating/interference between the free-running longitudinal modes, which thus increases the failure rate of automatic mode-locking.
In our experiments, automatic mode-locking using genetic wavefront shaping starts from an initial condition without any detectable pulses, which is realized by randomizing the wave plates of the laser cavity to set the laser out of the mode-locking condition. Figure 5a shows the evolution of the objective value during automatic mode-locking using genetic wavefront shaping. Different from previous cases where the evolutions of the objective value are gradually changed, here, the system exhibits a sharp transition within only a few generations of breeding such that the objective value increases by ~20 times from generation 12 to 16. This occurs mainly because the system parameters that can enable STML have a wide range of values, and an approximate wavefront can initiate a strong pulse that subsequently triggers self-evolved mode-locking, as also shown in previous 1D studies33. During the mode-locking manipulation, the mode profile exhibits variation similar to the optical power variation in the case of mode profile cleaning (Fig. S2a, b). However, the mode profile becomes steady once the genetic optimization is finished (Fig. S2c). Figure 5b depicts the temporal signals before and after mode-locking, wherein a pulse train is clearly demonstrated after successful automatic mode-locking. The pulse period is approximately 61 ns, corresponding to an FRR of ~16 MHz. Before mode-locking, the optical spectrum, as shown in Fig. 5c, manifests a complicated scene in which multiple quasi-CW wavelength components coexist. Once mode-locking is established, the third wavelength component from the left is significantly broadened, while the other three quasi-CW components remain unchanged except for intensity variations. It is also noted that other cases involving only one wavelength component have been observed (Fig. S3). The spectral width of the mode-locked component is approximately 8 nm, corresponding to a transform-limited pulsewidth of approximately 150 fs. The pulsewidths before and after dechirping are 1.76 ps and 184 fs, respectively, as shown in Fig. 5e. It is worth pointing out that the mode-locking wavelength is especially sensitive to the initial setting of the wave plates, i.e., the linear polarization offset of the optical system.
Considering the nonlinear phase shift that accumulates in optical fibres, particularly when delivering intense ultrashort pulses with extremely high peak powers, the generation of multiple pulses coexisting in the same round trip is observed. To manipulate the evolution of the multiple pulses, we demonstrate the precise control of the number of coexisting pulses in an MMF laser. To this end, the objective function is first evaluated by the RF intensity of the 10th harmonic at a pump power of approximately 20 W, which supports the generation of multiple pulses. Once multiple pulses are obtained, the pump power is gradually decreased. As shown in Fig. 5f, different numbers of coexisting pulses are generated, i.e., one to five pulses. Please note that the maximum number of coexisting pulses is mainly determined by the pump power level, which can be attributed to the energy quantization effect of dissipative optical solitons37-39. One of the potential applications of multi-pulse mode-locking is the generation of high-repetition-rate pulses through harmonic mode-locking. Andral et al.40 successfully demonstrated automatically harmonic mode-locking in a single-mode fibre laser by incorporating a merit function with a combination of RF harmonics. They also proposed using the pump power as an additional gene and considered the effect of laser hysteresis.