Femtosecond pulse amplification on a chip

The study addresses the long-standing challenge of achieving high-peak-power femtosecond pulse amplification directly on a photonic chip. While femtosecond pulses underpin numerous applications (biomedical imaging, femtochemistry, precision spectroscopy, sensing, low-noise signal generation, timekeeping), chip-scale sources to date provide insufficient peak power (around or below 1 W), forcing reliance on table-top amplifiers. The strong optical nonlinearity inherent to tightly confining integrated waveguides typically causes pulse distortion or damage at modest peak powers, preventing on-chip amplification. The authors propose and demonstrate a chip-based approach akin to chirped pulse amplification (CPA) that manages nonlinearity and dispersion to simultaneously amplify and compress pre-chirped pulses within the amplifier, targeting ultrashort, high-peak-power output directly at the chip facet. The goal is to unlock chip-integrated ultrafast technology with peak powers comparable to bench-top systems for applications ranging from spectroscopy to navigation.

Recent chip-integrated femtosecond sources include microresonator frequency combs, on-chip mode-locked lasers, and integrated electro-optic pulse generators, offering compact and scalable platforms but limited peak power. Prior integrated amplifiers have been shown using rare-earth-doped waveguides (erbium, thulium; up to ~1 W CW-levels), heterogeneous III–V semiconductor gain, and parametric amplification. However, amplification of femtosecond pulses on chip remained difficult due to strong nonlinearities in waveguides (often ~1000× higher than in fibers), with nonlinear distortion observed around 20 W peak power in integrated rare-earth-doped devices. The advent of chirped pulse amplification (CPA) in bulk systems overcame similar challenges by stretching, amplifying, and recompressing pulses, but implementing analogous functionality on chip requires careful dispersion and nonlinearity control and has not been realized for high-power femtosecond pulses. The authors build on these foundations, leveraging large mode-area (LMA) gain waveguides, all-normal dispersion design, and Tm-doped gain to overcome previous limitations.

Device design and fabrication: The amplifier is implemented in a CMOS-compatible silicon nitride (Si3N4) platform with an 800 nm-thick waveguide layer. The total on-chip propagation length is 12 cm folded into a footprint below 15 mm². To form the gain region, the SiO2 top cladding is selectively thinned to 200 nm, and a 1000 nm-thick thulium-doped alumina (Tm³+:Al2O3, n≈1.72) layer is deposited as an active top cladding, then overcladded by 1000 nm SiO2. The estimated Tm concentration is 3.5 × 10^20 cm^-3. Straight gain sections use a narrowed 300 nm-wide Si3N4 core producing a single-mode, large mode-area (LMA) waveguide where most optical power resides in the gain layer. Curved sections use 1000 nm-wide cores for tight confinement and low-loss bends (Euler bends with minimum radius 200 µm). Sections are connected via adiabatic tapers; input/output access waveguides are 2000 nm wide with inverse tapers for low-loss coupling. Design optimization: Finite-element mode simulations using ellipsometry-derived, wavelength-dependent refractive indices determine mode confinement, effective mode area, and group-velocity dispersion (GVD). The design targets (i) high power fraction in Tm:Al2O3 to raise gain and saturation power while reducing nonlinearity (lower intensity and ~10× lower nonlinear index than Si3N4), (ii) all-normal GVD across gain, taper, and bend sections to enable monotonic compression and avoid modulation instability, and (iii) manageable mode area to prevent parasitic lasing/spontaneous emission issues. The selected Tm:Al2O3 thickness (1000 nm) and geometries achieve these goals. The effective mode area in the gain layer is ~7 µm² with nonlinear parameter γ ≈ 0.007 W^-1 m^-1 in the LMA gain section (γ ≈ 0.7 W^-1 m^-1 in 1000 nm-wide waveguides). The chip’s total group-delay dispersion (GDD) is 7.79 × 10^-26 s², sufficient to support ps-level pre-chirping; distinct third-order dispersion contributions from bends and tapers are exploited to reduce residual TOD. Experimental setup: A 1 GHz ultrafast source provides ~80 fs pulses centered at 1815 nm (generated by launching 150 fs, 1560 nm pulses into HNLF to form a Raman-shifted soliton; a 1700 nm LPF removes shorter wavelengths). The input average power coupled to the chip is 1.81 mW (1.81 pJ per pulse). Pulses are pre-chirped in a short segment of anomalous-dispersion fiber; the fiber length is tuned (optimal 146.5 cm) to counterbalance the chip’s normal dispersion so that pulses compress near the output. The amplifier is forward-pumped at 1610 nm using a diode laser amplified by an EDFA; pump and signal are combined and coupled onto the chip. Output is collected in free space to avoid additional dispersion/nonlinearities and characterized via frequency-resolved optical gating (FROG). Pump and signal fiber-to-chip coupling losses are measured on passive reference waveguides (pump: 2.9 dB/facet; signal: 4.2 dB/facet). Net on-chip gain is extracted by comparing calibrated input/output signal powers (pump filtered out with a long-pass filter). Numerical modeling: Pulse propagation and amplification are simulated with a generalized nonlinear Schrödinger equation (NLSE) incorporating measured dispersion, nonlinearity, and a Tm gain model. In time domain, Kerr nonlinearity is modeled via ∂A/∂z = iγ|A|²A. In frequency domain, dispersion and position-dependent gain are applied to Â(ω). The gain model includes level populations (H6, F4, H4), active/quenched ion fractions, cross-relaxation and ETU terms (tested and simplified per parameter sensitivity), and coupled pump/signal propagation with overlap factors Γp, Γs and propagation loss α. Simulations use experimental inputs (pulse parameters, pump power, waveguide geometry) and reproduce measured spectra, FROG traces, and temporal profiles, allowing extraction of dispersion length LD and nonlinear length LNL along the chip to assess the dominance of linear versus nonlinear dynamics. CPA-inspired on-chip strategy: Pre-chirped pulses enter the amplifier with reduced peak power; as they propagate through all-normal dispersion gain sections, they are amplified while monotonically compressing, maintaining low peak power until near the output, where the highest peak power is reached, minimizing nonlinear distortions over most of the 12 cm device.

- Demonstration of on-chip femtosecond pulse amplification to high peak power: >50× on-chip amplification of 1 GHz pulses at 1815 nm, achieving 95 mW average output power, 95 pJ per pulse, and >800 W peak power with 116 fs duration directly at the chip output. - Net on-chip gain and pump dependence: Up to 17 dB net gain at 700 mW on-chip pump power (1610 nm), with an on-chip signal output of 95 mW. Optical spectra before/after amplification confirm gain and bandwidth evolution. - Dispersion-managed compression: Optimal pre-chirping fiber length of 146.5 cm counteracts the chip’s normal GDD (7.79 × 10^-26 s²), yielding near-transform-limited output pulses; FROG reconstructions show largely flat temporal phase with a weak asymmetric tail indicative of small residual TOD. - Bandwidths and durations: Input pulses (~80 fs) had broader bandwidth (~5.1 THz) than the amplified output (~3.5 THz), consistent with output pulse duration of 116 fs and finite gain bandwidth limitations. - Waveguide parameters: LMA gain waveguides achieve effective mode area ~7 µm² in the gain layer and low nonlinear parameter γ ≈ 0.007 W^-1 m^-1, reducing nonlinear accumulation; bends (1000 nm width) have γ ≈ 0.7 W^-1 m^-1 but are short and dispersion-engineered. - Propagation regime: Simulations indicate linear effects dominate over most of the >12 cm propagation (LD ≪ LNL), with nonlinear effects becoming significant only in the last ~1.5 mm, enabling stable monotonic compression and high-quality pulse output. - Agreement between experiment and simulation: Measured FROG traces and reconstructed temporal intensity/phase (original, pre-chirped, and amplified pulses) agree closely with simulations (FROG errors on the order of 10^-5), validating the physical model and design approach.

The work directly addresses the central barrier to fully integrated ultrafast photonics—the inability to amplify femtosecond pulses on chip without nonlinear degradation—by combining large mode-area rare-earth-doped gain waveguides with all-normal dispersion to support CPA-like operation in an integrated device. Pre-chirping lowers the instantaneous intensity, while engineered dispersion and low γ ensure that pulse propagation remains largely linear during most of the amplification path. As the pulse compresses monotonically, high peak power is reached only near the output, minimizing nonlinear phase accumulation and distortion. The demonstrated 17 dB net gain and 800 W peak power at 1 GHz repetition rate exceed previous on-chip sources by 2–3 orders of magnitude in peak power, enabling applications such as on-chip supercontinuum from IR to UV and self-referencing for precision metrology. The close match between experiment and NLSE-plus-gain modeling provides confidence in scaling and design generality. The approach is compatible with other dopants (e.g., erbium for telecom) and suggests pathways to higher average power via longer devices and larger mode areas, or to even higher peak powers by increasing dispersion or employing shorter pulses while staying within linear propagation over most of the device.

This study demonstrates the first on-chip amplification of femtosecond pulses to hundreds of watts peak power using a CMOS-compatible Si3N4 platform with a Tm:Al2O3 gain cladding, LMA waveguides, and all-normal dispersion. The device achieves >50× amplification, 17 dB net gain, 95 pJ pulse energy, ~800 W peak power, and 116 fs pulse duration at 1 GHz, with experimental results corroborated by detailed NLSE-based simulations. The concept bridges the gap between integrated sources and table-top ultrafast systems, opening avenues for chip-scale supercontinuum generation, self-referenced frequency combs, and integrated microwave photonics. Future research can focus on: (i) increasing average power via longer amplifiers and larger mode areas; (ii) pushing peak power through enhanced dispersion management or shorter seed pulses; (iii) fully on-chip pre-chirping/dispersion control using integrated gratings or anomalous-dispersion waveguides; (iv) exploring self-similar amplification in normal dispersion followed by anomalous-dispersion and nonlinear compression to realize high-power few-cycle pulses; and (v) adapting the platform to other rare-earth dopants (Er, Nd, Yb) for different wavelength bands and applications.

- Residual dispersion: FROG reconstructions reveal a weak residual third-order dispersion causing an asymmetric low-intensity tail; complete TOD compensation is not yet achieved. - Finite gain bandwidth: Output pulses (116 fs) are longer than the input (80 fs), consistent with narrower amplified spectra (~3.5 THz vs ~5.1 THz), indicating gain-bandwidth-limited compression. - External pre-chirping and pumping: The demonstration relies on off-chip anomalous-dispersion fiber for pre-chirping and an EDFA-pumped 1610 nm source; while integrable, these functions are not yet monolithically implemented. - Nonlinearity near output: Although most propagation is linear (LD ≪ LNL), nonlinear effects become non-negligible in the final ~1.5 mm; scaling to higher peak power will require increased dispersion or shorter pulses to maintain linear dominance. - Risk of parasitic lasing/spontaneous emission: Very large mode areas or high excited-ion densities can increase thresholds for parasitic processes, constraining geometry and pumping schemes; careful design is required, especially for forward pumping. - Device-specific optimization: The approach depends on precise dispersion engineering (including bends/tapers) and material indices; variations in fabrication may require re-optimization for consistent performance.