Directly Modulated Laser Platforms

• Directly modulated laser platforms are photonic sources where the laser structure itself produces modulation without external modulators, reducing hardware complexity.
• They employ diverse techniques such as semiconductor rate equations, master–slave injection locking, and intracavity Pockels modulation to control chirp, phase, and dynamic behavior.
• Differentiable modeling and end-to-end optimization enable precise control over modulation parameters, enhancing performance in optical communications, metrology, and quantum key distribution.

Searching arXiv for papers on directly modulated laser platforms and related modeling/optimization. A directly modulated laser platform is a photonic source architecture in which the laser cavity itself, or an intracavity electro-optic element, is driven electrically so that modulation is produced without a separate external optical modulator. In the literature represented here, the term spans several distinct but related implementations: single-mode semiconductor directly modulated lasers governed by carrier–photon rate equations; master–slave injection-locked transmitters that use direct current perturbations to control phase and time-bin structure; and a hybrid InP/thin-film lithium niobate Pockels laser in which the frequency-selective cavity and the high-speed Pockels actuator are co-located on chip (Xue et al., 2024, Hernandez et al., 2023, Woodward et al., 2021). Across these realizations, the platform is used for IM/DD links, chirp-aware transmission, FMCW metrology, direct frequency stabilization, and quantum communication.

1. Architectural classes and physical operating principles

The conventional semiconductor directly modulated laser is described by coupled rate equations for carrier density $N(t)$ and photon density $S(t)$, driven by a time-dependent injection current $I(t)$. In the formulations used for large-signal modeling, the dynamics take forms such as

$$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$

or, with gain compression,

$G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$

These equations encode the essential fact that direct modulation is not merely an amplitude operation: the injected current changes carrier density, which in turn perturbs gain, refractive index, chirp, relaxation oscillations, and extinction ratio (Hernandez et al., 2023, F. et al., 2024).

A second architectural class uses direct modulation in an injection-locked pair. In the gigahertz MDI-QKD transmitter, each user employs two DFB semiconductor lasers in a master–slave arrangement: the primary laser is gain-switched at $1\ \mathrm{GHz}$ and the secondary at $2\ \mathrm{GHz}$, with the injected seed forcing the secondary’s optical phase to follow that of the primary while reducing chirp and jitter (Woodward et al., 2021). A related design for high-dimensional QKD also uses master–slave DFBs, with a $312.5\ \mathrm{MHz}$, $3\ \mathrm{ns}$ gain-switch pulse train for the master, $400\ \mathrm{ps}$ gating pulses for the slave, and $S(t)$0 amplitude perturbations on the master RF drive to imprint relative phases onto slave pulses (Zhou et al., 13 Mar 2026). In the directly phase-modulated light source, the same functional separation appears in a different form: a phase-preparation laser provides the phase state, and a pulse-generation laser supplies $S(t)$1 gain-switched pulses at $S(t)$2, coherently seeded through an optical circulator (Yuan et al., 2016).

A third class replaces carrier-induced index modulation with an intracavity Pockels element. The Pockels-laser platform hybrid-integrates an InP reflective semiconductor optical amplifier with a thin-film lithium niobate photonic integrated circuit. The thin-film lithium niobate chip forms both the frequency-selective cavity and the high-speed Pockels element; an extended-distributed Bragg reflector grating etched in an overlying SiO$S(t)$3 cladding provides narrow reflection bandwidth, while a contiguous $S(t)$4 TFLN waveguide serves as an electro-optic phase shifter and a co-located $S(t)$5 eDBR section carries gold electrodes with a $S(t)$6 electrode–waveguide gap (Xue et al., 2024).

Taken together, these variants define a platform family rather than a single device type. The unifying feature is that modulation authority resides in the laser structure itself, either through carrier–photon dynamics, injection-mediated phase transfer, or intracavity electro-optics.

2. Nonlinear dynamics, chirp, and threshold-region behavior

Large-signal DML behavior is difficult because no closed-form solution exists for the nonlinear rate equations, the equations are stiff, and the drive waveform can be arbitrary. Small-signal linearization around an operating point fails to capture waveform distortion, chirp, relaxation oscillations, and extinction-ratio trade-offs at high bitrates (Hernandez et al., 2023). This is the basic reason directly modulated platforms remain both attractive and difficult: they reduce hardware complexity, but the modulation process is inseparable from the laser’s internal nonlinear dynamics.

Near threshold, the dynamics become especially rich. Under pump modulation

$S(t)$7

and with spontaneous-emission coupling retained, modulation around threshold includes below-threshold dynamics that are not captured if the photon population is allowed to collapse to zero. For the macroscopic, low-$S(t)$8 laser studied in the threshold-region bifurcation analysis, $S(t)$9 is sufficient to keep the rate equations coupled even when $I(t)$0, so deterministic oscillatory trajectories extend into the below-threshold region (Zou et al., 2021).

The reported bifurcation structure is explicit. At fixed $I(t)$1 and bias just above threshold, one sees strong harmonic generation for $I(t)$2, period doubling between $I(t)$3 and $I(t)$4, small intervals of period-4 from $I(t)$5 to $I(t)$6, period-8 from $I(t)$7 to $I(t)$8, and a broadened continuous RF band indicating fully developed chaos in $I(t)$9–$$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$0. Beyond $$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$1, the output locks back to the sinusoidal modulation. At fixed $$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$2, increasing $$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$3 gives a single-loop attractor for $$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$4, then period doubling, then higher-order cascades, a transition region with two distinct period-2 attractors around $$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$5–$$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$6, and a broadband chaotic-like regime by $$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$7 (Zou et al., 2021).

These observations correct a common simplification: direct modulation is not generically equivalent to a memoryless amplitude transfer function. The threshold-region study shows that the spontaneous-emission term can be decisive for whether below-threshold portions of the drive remain dynamically active, while the communications-oriented models show that chirp and relaxation behavior remain important well above threshold (Zou et al., 2021, Hernandez et al., 2023).

3. Differentiable modeling and end-to-end optimization

Because large-signal DML dynamics lack an analytical solution suitable for gradient-based learning, several works replace the physical ODE block with differentiable surrogate models trained on rate-equation traces or directly on experimental waveforms. In the rate-equation-based studies, the surrogate maps an input current waveform $$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$8 to an optical-power waveform $$\frac{dN}{dt} = \frac{I(t)}{qV} - v_g g(N-N_{tr})S - \frac{N}{\tau_n}, \qquad \frac{dS}{dt} = \Gamma v_g g(N-N_{tr})S + \beta \frac{N}{\tau_n} - \frac{S}{\tau_p},$$9, using training data generated from interleaved super-Gaussian and random pulse families, 4-PAM amplitudes, 32 samples/symbol, and $G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$0 training sequences with ground truth obtained by solving the rate equations with RK4,5 at a time-step of $G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$1 (Hernandez et al., 2023, Fernandez et al., 2023).

Four surrogate families are compared: a second-order Volterra filter with memory $G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$2, a time-delay neural network using a 32-sample input window, an LSTM unrolled over the 1024-sample sequence, and a decoder-only Convolutional Attention Transformer with learned positional embeddings, convolutional-attention sublayers, and a 2-layer MLP head (Hernandez et al., 2023). The CAT achieves $G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$3 NRMSE across $G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$4–$G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$5 and about $G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$6 in the high-speed region $G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$7; equalizers trained on CAT outputs generalize almost identically to the true ODE outputs, with $G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$8, whereas TDNN and Volterra show up to $G[N,S] = g_0 (N-N_0)/(1+\varepsilon S).$9 mismatch (Hernandez et al., 2023). In related results, CAT achieves NRMSE $1\ \mathrm{GHz}$0 up to $1\ \mathrm{GHz}$1, with inference and training per epoch on NVIDIA A100 within $1\ \mathrm{GHz}$2–$1\ \mathrm{GHz}$3 of an LSTM and both about $1\ \mathrm{GHz}$4 faster than the ODE solver used for data generation (Fernandez et al., 2023).

These differentiable channel models are then embedded in end-to-end optimization loops. One formulation jointly optimizes transmitter pulse shaping, receiver equalization, bias current, and peak-to-peak modulation current for symbol rates between $1\ \mathrm{GHz}$5 and $1\ \mathrm{GHz}$6, while another chirp-aware formulation includes both amplitude and phase in the surrogate to address the interplay between chirp and chromatic dispersion (F. et al., 2024, F. et al., 2024). In the simulation study over $1\ \mathrm{GHz}$7, $1\ \mathrm{GHz}$8, and $1\ \mathrm{GHz}$9, the end-to-end autoencoder jointly learns constellation points, a 2-tap pulse-shaping filter, $2\ \mathrm{GHz}$0, $2\ \mathrm{GHz}$1, and receiver processing; at $2\ \mathrm{GHz}$2 and $2\ \mathrm{GHz}$3 it reports $2\ \mathrm{GHz}$4 versus $2\ \mathrm{GHz}$5 for VNLE, $2\ \mathrm{GHz}$6 for FFE, and $2\ \mathrm{GHz}$7 for the baseline, with $2\ \mathrm{GHz}$8 (F. et al., 2024). The same study reports that the optimized bias rises with symbol rate, from $2\ \mathrm{GHz}$9 at $312.5\ \mathrm{MHz}$0 to $312.5\ \mathrm{MHz}$1 at $312.5\ \mathrm{MHz}$2 and $312.5\ \mathrm{MHz}$3 at $312.5\ \mathrm{MHz}$4, reflecting the bandwidth–extinction-ratio trade-off (F. et al., 2024).

Experimental end-to-end demonstrations extend this methodology to measured DML links. In one back-to-back IM/DD system using an NTT NLK1551SSC directly modulated laser, the joint transmitter–receiver optimization at $312.5\ \mathrm{MHz}$5 and $312.5\ \mathrm{MHz}$6 outperforms nonlinear receiver-only equalization while using only 9 total taps, with measured $312.5\ \mathrm{MHz}$7 bandwidth compressed by $312.5\ \mathrm{MHz}$8 at $312.5\ \mathrm{MHz}$9 and $3\ \mathrm{ns}$0 at $3\ \mathrm{ns}$1 versus RRC-shaped signals (Hernandez et al., 2024). In a subsequent experimental study over $3\ \mathrm{ns}$2, $3\ \mathrm{ns}$3, and $3\ \mathrm{ns}$4 SSMF, end-to-end optimization using a 5-tap TX filter and 15-tap RX FFE reports, for $3\ \mathrm{ns}$5 back-to-back, a best E2E SER of $3\ \mathrm{ns}$6 at $3\ \mathrm{ns}$7 and $3\ \mathrm{ns}$8, compared with $3\ \mathrm{ns}$9 for the best 20-tap FFE benchmark at $400\ \mathrm{ps}$0 and the same bias, while reducing electrical RX-side $400\ \mathrm{ps}$1-dB bandwidth from $400\ \mathrm{ps}$2 to $400\ \mathrm{ps}$3 (Hernandez et al., 27 Aug 2025).

The main significance of this line of work is methodological. It shows that the directly modulated laser platform can be treated not only as a device but as a learnable differentiable block, enabling joint optimization of laser drive, waveform, and DSP under realistic nonlinear dynamics.

4. Pockels-laser realization for ultrafast optical metrology

The Pockels-laser platform is a directly modulatable semiconductor laser that replaces the usual external electro-optic modulator with an intracavity thin-film lithium niobate element. Its eDBR grating, etched in an SiO$400\ \mathrm{ps}$4 cladding rather than in the high-index LN core, provides a Bragg coupling strength $400\ \mathrm{ps}$5 down to a few $400\ \mathrm{ps}$6 over a $400\ \mathrm{ps}$7 grating, suppresses side modes by more than $400\ \mathrm{ps}$8, and yields an intrinsic white-noise floor of only $400\ \mathrm{ps}$9, corresponding to $S(t)$00 (Xue et al., 2024).

When a voltage $S(t)$01 is applied across the $S(t)$02 electrode gap, the Pockels effect induces

$S(t)$03

which shifts the Bragg-cavity resonance by

$S(t)$04

With $S(t)$05 and $S(t)$06, this predicts about $S(t)$07–$S(t)$08, in excellent agreement with the measured $S(t)$09 (Xue et al., 2024). Because the entire external cavity is Pockels-active, the same electrodes can chirp the optical frequency directly over tens of gigahertz at modulation frequencies up to and beyond $S(t)$10, without an external modulator or optical alignment. The platform maintains a mode-hop-free tuning range of at least $S(t)$11 for $S(t)$12 up to $S(t)$13, and the ultimate chirp rate is expressed as

$S(t)$14

that is, $S(t)$15 (Xue et al., 2024).

This directly modulated source is then used for FMCW LiDAR and Doppler velocimetry. During each linear chirp,

$S(t)$16

so that

$S(t)$17

At a $S(t)$18 chirp rate, corresponding to $S(t)$19, a $S(t)$20 target produces $S(t)$21; velocities up to $S(t)$22 are cleanly resolved, and the measurable velocity extends up to the first cosmic velocity at $S(t)$23 away. Two-dimensional ranging images are obtained with $S(t)$24 resolution, consistent with $S(t)$25 and $S(t)$26 (Xue et al., 2024).

The same intracavity Pockels element is also used to implement Pound–Drever–Hall locking directly. A single-tone RF drive at $S(t)$27 imprints sidebands on the cavity, the reflected light from an H$S(t)$28C$S(t)$29N gas cell produces an error signal, and feedback is applied to the same Pockels electrodes. Over three one-hour intervals, frequency excursions shrink from $S(t)$30 free running to $S(t)$31 locked, with the locked result limited principally by mechanical drift of the unpackaged chips and the $S(t)$32 resolution of the wavemeter (Xue et al., 2024).

This platform directly challenges the assumption that ultrafast chirping, narrow linewidth, wide tuning, and stabilization require a chain of external AOMs, EOMs, circulators, and beamsplitters. In the reported implementation, these functions are collapsed into a single chip-scale device.

5. Communication-system realizations

In short-reach optical links, directly modulated lasers are attractive because they are compatible with low-power, cost-constrained IM/DD architectures, but their limited modulation bandwidth and chirp can undermine throughput (F. et al., 2024). A standard communication realization places the DML between digital pulse shaping and square-law direct detection, with the optical field written as

$S(t)$33

and the chirp approximated by the Henry $S(t)$34-factor model

$S(t)$35

while chromatic dispersion acts through

$S(t)$36

In this view, chirp is not an incidental defect but an integral part of the channel model (F. et al., 2024).

The numerical chirp-aware optimization framework jointly tunes $S(t)$37, $S(t)$38, $S(t)$39, and $S(t)$40 through a differentiable autoencoder. In the example summarized for $S(t)$41 over $S(t)$42 SSMF, the AE is compared with a 20-memory-tap FIR equalizer and a second-order Volterra nonlinear equalizer. At $S(t)$43 the reported SERs are $S(t)$44 for FIR EQ, $S(t)$45 for VNLE, and $S(t)$46 for the AE; at $S(t)$47 they are $S(t)$48, $S(t)$49, and $S(t)$50, respectively. Mutual-information gains of about $S(t)$51–$S(t)$52 are also reported in favor of the AE, especially where chirp–dispersion interplay dominates (F. et al., 2024).

The experimental DML IM/DD testbeds make the same point from a hardware perspective. One setup uses an NTT NLK1551SSC directly modulated laser with small-signal $S(t)$53-dB bandwidth of about $S(t)$54, temperature controlled at $S(t)$55, and treats $S(t)$56 and RF drive power $S(t)$57 as optimization variables (Hernandez et al., 2024). Another experimental platform uses a NEL NLK1551SSC DML with threshold current $S(t)$58, peak differential external efficiency near bias $S(t)$59, and measured small-signal $S(t)$60-dB bandwidths of $S(t)$61, $S(t)$62, and $S(t)$63 at biases of $S(t)$64, $S(t)$65, and $S(t)$66, respectively (Hernandez et al., 27 Aug 2025).

A further architectural dimension is source integration. A monolithic InAs/GaAs quantum-dot DFB laser array on silicon demonstrates CW threshold currents as low as $S(t)$67, SMSRs as high as $S(t)$68, well-aligned channel spacing of $S(t)$69, and a wavelength coverage range of $S(t)$70 spanning the O-band (Wang et al., 2018). The same report states explicitly that it does not report small-signal modulation bandwidth or large-signal eye diagrams (Wang et al., 2018). This is important context: monolithic integration of laser arrays on silicon is compatible with DML-oriented system design, but source integration alone does not establish direct-modulation performance.

6. Quantum communication implementations and integration outlook

Directly modulated laser platforms are prominent in quantum communication because they can replace external phase and intensity modulators with laser-internal dynamics. In the directly phase-modulated light source, a phase-preparation laser is biased into quasi–steady-state emission and directly modulated by a small electrical perturbation, while a pulse-generation laser is gain-switched at $S(t)$71 to produce $S(t)$72 pulses that inherit the phase-preparation laser’s instantaneous optical phase. The accumulated phase shift obeys

$S(t)$73

and the cavity-enhanced electro-optic effect yields an effective half-wave voltage of about $S(t)$74. The measured phase-transfer visibility reaches $S(t)$75, saturating at $S(t)$76 seed power, with modulation bandwidth $S(t)$77 limited by the electrical driver (Yuan et al., 2016). In BB84 phase-encoded QKD, the source is reported to operate at $S(t)$78 sifted-key clock rate with secure key rates of several kbps at $S(t)$79 channel loss and tolerance up to $S(t)$80, while over $S(t)$81 with no active feedback the measured QBER remains at $S(t)$82 (Yuan et al., 2016).

The MDI-QKD transmitter of Woodward et al. uses gain-switching and injection locking to encode phase-modulated time-bin bits with free-running semiconductor laser sources, without spectral or phase feedback (Woodward et al., 2021). The primary laser is biased below threshold except when gain-switched at $S(t)$83 with $S(t)$84 duty; the secondary is gain-switched at $S(t)$85 with $S(t)$86 duty so that each primary envelope contains two secondary pulses. A $S(t)$87 band-pass filter removes spontaneous-emission wings, and detection is performed with SNSPDs of about $S(t)$88 efficiency and dark counts of about $S(t)$89. The measured free-running detuning between independent DFBs drifts by about $S(t)$90 over $S(t)$91, which corresponds to $S(t)$92 over the $S(t)$93 time-bin spacing and only $S(t)$94 worst-case X-basis QBER increase. Reported finite-size secure key rates are $S(t)$95 at $S(t)$96, $S(t)$97 at $S(t)$98, $S(t)$99 at $I(t)$00, and $I(t)$01 at $I(t)$02 channel loss, with an asymptotic rate of $I(t)$03 at $I(t)$04 (Woodward et al., 2021).

The same design logic extends to high-dimensional encoding. The directly modulated laser platform for HD-QKD realizes 4-dimensional time-bin BB84 states using master–slave injection-locked DFBs, a repetition rate of $I(t)$05, and only two fiber Faraday–Michelson interferometers on the receiver side, with delays $I(t)$06 and $I(t)$07 (Zhou et al., 13 Mar 2026). It reports experimental 4D QKD over $I(t)$08 and $I(t)$09 of standard fiber, with finite-key secret key rates of $I(t)$10 and $I(t)$11, respectively, and states that the four-dimensional encoding outperforms the two-dimensional counterpart in secret key rate under identical hardware conditions (Zhou et al., 13 Mar 2026).

A plausible implication is that the directly modulated laser platform has become a unifying hardware strategy across classical and quantum photonics: direct modulation reduces the need for bulk external modulators, but its success depends on whether the underlying dynamics can be shaped, stabilized, or learned. The integration outlook in the cited work is correspondingly consistent. The quantum platforms point to co-packaged DFBs, bias tees, RF routing, on-chip delay lines, and on-chip SNSPD integration (Zhou et al., 13 Mar 2026), while the silicon quantum-dot DFB array demonstrates that monolithic laser integration on silicon can already achieve CWDM-compatible wavelength grids and high SMSR, even though direct-modulation bandwidth remains to be characterized in that specific platform (Wang et al., 2018).