Metrology-grade mid-infrared spectroscopy for multi-dimensional perception Baoqi Shi, Chenxi Zhang, Ming-Yang Zheng, Yue Hu, Zeying Zhong, Zhenyuan Shang, Wenbo Ma, Xiu-Ping Xie, Xue Bai, Yi-Han Luo, Anting Wang, Hairun Guo, Qiang Zhang, Junqiu Liu Abstract The mid-infrared spectral window is essential for molecular fingerprinting and atmospheric sensing, yet unlocking its full potential is currently constrained by a fundamental instrumental trade-off: existing systems cannot simultaneously deliver broad bandwidth, high photon flux, and metrological frequency fidelity. Here, we resolve this bottleneck by demonstrating a metrology-grade spectroscopic system based on difference frequency generation, driven by widely tunable, near-infrared diode lasers traceable to atomic standards. Our system achieves continuous tunability across the 3–3.7 µm atmospheric window and delivers output power exceeding 45 mW with an absolute frequency accuracy of 7.2 MHz. We harness this convergence to overcome a critical barrier in integrated photonics, unambiguously identifying and eliminating hydrogen-induced absorption in silicon nitride microresonators to achieve an 88-fold reduction in optical loss. We further reveal multi-phonon absorption in the silica cladding as the fundamental limit to mid-infrared integrated photonics. Finally, we demonstrate the system’s versatility through scattering-resilient LiDAR capable of penetrating optically dense fog, and dual-modality sensing that simultaneously retrieves target distance and chemical composition. By unifying the rigor of frequency metrology with the versatility of broadband sensing, this architecture establishes a new paradigm for multi-dimensional perception in complex environments. The mid-infrared (MIR) spectral region hosts the fundamental vibrational-rotational transitions of molecules, exhibiting absorption cross-sections orders of magnitude larger than their near-infrared (NIR) overtones 63 . This unique sensitivity underpins transformative applications ranging from planetary-scale greenhouse gas monitoring to non-invasive medical diagnostics 57 ; 11 ; 41 . In demanding scenarios—such as life-support monitoring in space or sensing in dynamic industrial environments—systems must resolve complex chemical mixtures rapidly and unambiguously 9 . Consequently, high-performance spectroscopy requires a convergence of capabilities rarely co-existing in a single instrument: broad bandwidth for multi-species coverage, fine resolution to distinguish overlapping lines, exact frequency determination, high spectral brightness, high signal-to-noise ratio (SNR), and real-time processing speed. Despite this urgent demand, established techniques suffer from fundamental trade-offs, illustrated in Fig. 1 a. Fourier transform infrared spectroscopy (FTIR) 28 offers broad bandwidth but is constrained by low brightness and limited frequency fidelity (i.e., precision and accuracy). Conversely, while MIR dual-comb spectroscopy 81 ; 4 ; 56 ; 75 enables high-precision, time-resolved measurements, it distributes optical power across tens of thousands of comb lines. This results in low optical power spectral density (PSD), severely limiting the photon flux per mode available for long-path remote sensing 19 . In contrast, continuous-wave (CW) tunable lasers—such as quantum cascade lasers (QCLs) 14 , interband cascade lasers (ICLs) 74 , or optical parametric oscillators (OPOs) 85 ; 86 ; 15 —maximize photon flux and interaction lengths that are pivotal for nearly all ultrasensitive optical detection. However, these sources are historically hampered by limited mode-hop-free tuning ranges, precluding the seamless, broadband characterization required for complex molecular fingerprinting. Figure 1: Metrology-grade mid-infrared spectroscopic systems breaking the performance trade-offs . a . Performance benchmarking against established techniques: Fourier transform infrared (FTIR) spectroscopy, dual-comb spectroscopy, and standard tunable CW laser spectroscopy (including QCL, ICL, and OPO). The radar chart illustrates how our approach resolves the fundamental instrumental trade-offs, uniquely combining high optical power spectral density (PSD), high SNR, and seamless broad bandwidth with metrological frequency accuracy and precision via simplified digital signal processing (DSP). b . Schematic architecture. A widely tunable, MIR CW laser is synthesized via difference frequency generation (DFG) of two NIR chirping lasers in a chirped periodically poled lithium niobate (CPLN) waveguide. The instantaneous frequencies of the NIR pump lasers are rigorously calibrated using fiber cavities and referenced to atomic hyperfine transitions. c . Atmospheric transmission 51 (top) and molecular absorption profiles 20 ; 22 (bottom). The system targets the 3–3.7 µm band, exploiting a highly transparent atmospheric window that minimizes H 2 O and CO 2 interference while accessing the strong absorption fingerprints of CH 4 and HCl. d . Emerging applications enabled by our MIR platform: non-destructive semiconductor metrology, scattering-resilient FMCW LiDAR, and simultaneous ranging and gas sensing. An ideal spectroscopic platform would bridge this gap by combining the coherence and high photon flux of CW lasers with the bandwidth of frequency combs and the absolute accuracy of metrology standards. Such a synthesis represents more than an incremental improvement; it is a prerequisite for unlocking new frontiers in MIR science outlined in Fig. 1 d. For instance, in semiconductor manufacturing, while MIR light sources are used to monitor chemicals, a spectrometer with absolute frequency accuracy and fine frequency resolution is critical for decoupling fabrication-induced errors from intrinsic material imperfection. Similarly, for autonomous driving, standard NIR LiDAR fails in heavy fog due to Mie scattering 84 , while MIR LiDAR capable of penetrating complex environments significantly boosts safety standards. Ultimately, unifying spectroscopic resolution with spatio-temporal coherence enables a new paradigm: fusing spatial geometry and chemical composition into a unified multi-dimensional map. Here, we demonstrate a metrology-grade MIR spectroscopic system that simultaneously synergizes broad bandwidth, high brightness, high SNR, and potential for real-time processing. As depicted in Fig. 1 b, our approach leverages difference frequency generation (DFG) driven by two widely tunable, mode-hop-free, NIR external-cavity diode lasers (ECDLs). To ensure frequency accuracy, the two pump lasers (spanning 1035–1086 nm and 1536–1580 nm, individually) are rigorously calibrated against fiber cavities and referenced to atomic hyperfine transitions 59 , yielding a frequency accuracy of 7.2 MHz (see Methods). By amplifying and mixing these pumps in a chirped periodically poled lithium niobate (CPLN) waveguide, we generate up to 45.0 mW of MIR radiation. Details are found in Methods, Extended Data Figs. 1 and 2 , and Supplementary Materials Notes 1 and 2. This configuration delivers an individual mode-hop-free tuning range of 7.49 THz and an aggregate coverage of 19.1 THz (3001–3711 nm). Furthermore, we determine the fundamental frequency resolution by characterizing the MIR laser’s dynamic linewidth, measuring an average of 242 kHz over 100 µs integration time (see Supplementary Materials Note 3). We strategically target the 3–3.7 µm atmospheric window 51 —a highly transparent region that minimizes interference from ubiquitous H 2 O and CO 2 while maximizing sensitivity to hydrocarbons (e.g., CH 4 ) and industrial pollutants (e.g., HCl) 22 , as illustrated in Fig. 1 c. Harnessing this platform, we demonstrate three distinct applications: (1) a metrological probe for integrated photonics, which we utilize to isolate and mitigate material-induced optical losses in silicon nitride microresonators; (2) a frequency-modulated continuous-wave (FMCW) LiDAR capable of penetrating dense scattering media; and (3) a dual-modality sensor that simultaneously retrieves target distance and trace-gas fingerprints from a single waveform. Figure 2: Broadband spectroscopic characterization and loss mitigation in silicon nitride microresonators . a . Broadband transmission spectra of the same Si 3 N 4 microresonator before (as-deposited, red) and after (annealed, blue) annealing. b . Comparative line shape analysis of a single resonance at 95.448 THz. The high-temperature annealing reduces the intrinsic loss \kappa_{0}/2\pi from 1.6 GHz (as-deposited, red) to 0.35 GHz (annealed, blue), while the external coupling rate remains constant at \kappa_{\text{ex}}/2\pi=3.2 GHz. c . Frequency-dependent optical propagation loss \alpha converted from \kappa_{0}/2\pi . The annealing drastically eliminates the N-H bonds whose absorption peak is located at 100 THz ( \approx 3.0 µm). The attenuation \alpha is reduced from 2646 dB m -1 (as-deposited, red) to 30 dB m -1 (annealed, blue)—an 88-fold reduction. d . Integrated dispersion ( D_{\text{int}}/2\pi ) profiles fitted to a fourth-order polynomial. The annealing induces a shift in the free spectral range (FSR, D_{1}/2\pi ) from 97.80 GHz to 97.61 GHz at the central frequency \omega_{0}/2\pi=90.011 THz, reflecting changes in the refractive and group indices. e . Octave-spanning dispersion characterization (1260–3711 nm). The profile is obtained by stitching measurements from the NIR (1260–1640 nm, 546 resonances) and MIR (3001–3711 nm, 192 resonances) bands of the same device and fitting to a seventh-order polynomial. Inset, false-color SEM image of a Si {}_{3}N_{4} microresonator (waveguide highlighted in red). Mid-infrared integrated photonics . Integrated photonics 70 ; 47 ; 32 ; 67 ; 80 ; 27 has matured rapidly in the NIR telecommunication bands, yet the MIR spectral window remains a frontier with immense untapped potential for chemical sensing 87 and free-space communications 10 . Recent progress has yielded critical MIR components, including narrow-linewidth Brillouin lasers 34 , ring lasers 18 , microcombs 76 ; 83 ; 49 ; 55 ; 31 , and tunable filters 30 . Among various material platforms, silicon nitride (Si 3 N 4 ) is a premier candidate for expanding this ecosystem into the MIR, owing to its ultralow loss and high power handling 82 . However, the scalability of Si 3 N 4 devices beyond 3.5 µm has remained obscured by the onset of multi-phonon absorption in the silicon dioxide (SiO 2 ) cladding 68 ; 33 ; 42 . Although this material absorption is well-documented in bulk materials, its specific contribution to waveguide loss—which is governed by the spectral evolution of mode confinement—has not been rigorously quantified. Consequently, progress has been stalled by the absence of metrology capable of distinguishing manageable process imperfections from fundamental material absorption boundaries. Unlocking the potential of MIR integrated photonics demands characterization tools that simultaneously offer high frequency accuracy, resolution, and broad spectral coverage. Such capabilities are indispensable for quantifying high-quality-factor (high- Q ) resonances 73 , validating dispersion engineering 53 , and aligning resonances with molecular transitions 25 . Leveraging our system, we report the metrology-grade, broadband characterization of high- Q Si 3 N 4 microresonators, extending their operation into the uncharted 3–3.7 µm band. Figure 2 a red trace presents the transmission spectrum of a microresonator based on an as-deposited Si 3 N 4 film from low-pressure chemical vapor deposition (LPCVD). As exemplified in Fig. 2 b, the intrinsic loss \kappa_{0}/2\pi for each resonance is extracted via Lorentzian fitting (see Methods). By converting \kappa_{0}/2\pi to linear propagation loss \alpha for all probed resonances within the 3–3.7 µm, we uncover a distinct frequency-dependent loss profile as depicted in Fig. 2 c red dots. We note that the localized discontinuity observed near 95 THz arises from parameter indistinguishability at the critical coupling crossover, detailed in Supplementary Note 4. Crucially, a dramatic increase in \alpha towards 100 THz ( \approx 3.0 µm) is observed, peaking at \alpha=2646 dB m -1 . We identify this feature as the absorption fingerprint of residual N-H bonds in the as-deposited Si 3 N 4 film 37 ; 5 ; 45 whose precursors involve ammonia (NH 3 ). Guided by this spectral fingerprint, we optimize the fabrication process by adding high-temperature annealing (above 1200 ∘ C) on this Si 3 N 4 device to eliminate residual hydrogen 45 ; 43 . Details are found in Supplementary Materials Note 5. Figures 2 a–c blue traces evidence the efficacy of this mitigation and drastic loss reduction in the annealed device. For the mode at \omega/2\pi=95.448 THz in Fig. 2 b, \kappa_{0}/2\pi decreases from 1.6 GHz to 0.35 GHz, corresponding to Q_{0}=\omega/\kappa_{0} improvement to 2.7\times 10^{5} . Specifically, the value of \alpha at 100 THz drops precipitously from 2646 dB m -1 to just 30 dB m -1 —an 88-fold reduction that confirms the suppression of N-H absorption. Eliminating the N-H impurities unmasks a distinct loss profile where attenuation rises as frequency decreases. Finite-element simulations rule out geometric leakage (bending and substrate coupling) as the primary driver of this attenuation (see Methods). At 81 THz ( \approx 3.7 µm), the calculated radiative loss is limited to \sim 11.4 dB m -1 —over an order of magnitude lower than the measured value of \sim 420 dB m -1 . We instead attribute this excess loss to mode delocalization into the absorptive SiO 2 cladding 50 . Poynting vector analysis of the waveguide geometry reveals that 31% of the optical power propagates within the cladding at 81 THz, as shown in Extended Data Fig. 3 . Given the onset of strong SiO 2 multi-phonon absorption in this spectral window 68 ; 33 ; 50 , this substantial mode overlap is sufficient to account for the observed loss floor. Consequently, increasing the waveguide width offers a direct pathway to mitigate this limit by enhancing optical confinement. Characterization results of additional microresonators are found in Supplementary Materials Note 4. Moreover, our metrology-grade accuracy and precision allow for mapping of microresonator dispersion profile (see Methods). Figure 2 d compares the measured integrated dispersion ( D_{\text{int}}/2\pi ) profiles for the Si 3 N 4 microresonator before and after the annealing. While the waveguide geometry is unchanged, the difference in dispersion is also caused by the N-H absorption via the Kramers-Kronig relations 44 ; 29 . These results establish our system as a powerful, non-destructive diagnostic tool for guiding fabrication process iteration. Furthermore, by routing a portion of the NIR pump light, our system enables simultaneous characterization of both NIR and MIR spectral windows. Figure 2 e displays an octave-spanning dispersion profile obtained from one microresonator, stitching data from the NIR (1260–1640 nm) and MIR (3001–3711 nm) bands. Such broadband capability is critical for validating dispersion engineering 53 in applications requiring octave-spanning spectra 61 , including supercontinuum generation 26 ; 24 ; 23 , microcomb generation 76 ; 25 ; 83 ; 7 and frequency translation 17 ; 40 . Beyond Si 3 N 4 , our spectroscopic system is equally pivotal for advancing other MIR platforms, including chalcogenide glass 34 , III-V materials 23 ; 55 , Si 38 ; 83 ; 50 , Ge 6 ; 52 ; 62 , and SiGe 72 . Figure 3: Scattering-resilient mid-infrared FMCW LiDAR . a . Illustration of the fog-penetration concept. Inset, time-frequency representation of the ranging waveforms with a tuning bandwidth of B . The time delay between the emitted signal (solid lines) and the received echo (dashed lines) encodes the target distance. b . Experimental testbed. A controlled fog chamber simulates scattering environments, with density calibrated in real-time by monitoring the optical density of a co-propagating 1064 nm reference laser (OD 1064 ). PD, photodetector. BS, beam splitter. DM, dichromatic mirror. c . Depth imaging comparison of a steel target in dense fog (OD {}_{1064}=4 ) using identical tuning bandwidths ( B=125 GHz). The MIR system successfully resolves the target structure “IQASZ”, whereas the standard NIR system is completely obscured by noise. d . Signal resilience against scattering. The NIR signal (blue, B=125 GHz) is extinguished at OD {}_{1064}=4 , while the MIR signal (red, B=125 GHz) remains distinguishable even at OD {}_{1064}=5 . Green traces represent the MIR mode with B=7.49 THz, yielding 12 dB suppression of the noise floor compared to the red traces. e . Ranging precision histograms. While precision quickly degrades for the narrowband signals (blue and red, B=125 GHz) with increasing opacity, the wideband MIR mode (green, B=7.49 THz) maintains robust micron-level precision even under OD {}_{1064}=5 . Scattering-resilient ranging . LiDAR is the cornerstone of spatial perception in autonomous systems. While FMCW LiDAR offers shot-noise-limited sensitivity and immunity to ambient light 32 ; 39 , standard NIR (e.g., 1.55 µm) implementations suffer severe attenuation in scattering media like fog or smoke, posing a distinct risk to safety-critical applications 84 . To address this vulnerability, we exploit the superior propagation characteristics of the MIR spectrum, as shown in Fig. 3 a. Since scattering strength generally scales inversely with wavelength (e.g., via Rayleigh scattering), shifting operation to the 3–3.7 µm atmospheric window significantly reduces attenuation compared to 1.55 µm. We validate this penetration capability using a comparative testbed incorporating a controlled fog chamber (see Methods), as shown in Fig. 3 b. Scattering environments are simulated via a water mist generator, with the optical density (OD 1064 ) quantified in real-time by a co-propagating 1064 nm probe laser. To ensure a rigorous comparison, we configure a standard NIR system (centered at 1.55 µm) and our MIR system (centered at 3.56 µm) with identical parameters: 1 mW optical power and B=125 GHz tuning bandwidth. We compensate for chirping nonlinearity using intrinsic fine frequency calibration. Details are presented in Supplementary Materials Note 6. The resilience of the MIR signal is striking in Fig. 3 d. As scattering density increases, the NIR signal (blue trace) undergoes rapid extinction, falling below the noise floor at OD {}_{1064}=4 . In stark contrast, the MIR signal (red trace) retains a robust SNR of 23 dB under identical conditions. Even at the dynamic range limit of our detector (OD {}_{1064}=5 ), the MIR system maintains a distinguishable SNR of 19 dB. Figure 3 c evidences that this spectral advantage translates directly into imaging performance: while the NIR point cloud is completely obscured by noise, the MIR system resolves the depth profile of a steel target at 1.62 m through dense fog with high fidelity. Beyond scattering immunity, high-precision ranging demands wide spectral bandwidth. Leveraging our system’s broad mode-hop-free tuning range, we expand the modulation bandwidth to B=7.49 THz. This bandwidth not only refines axial resolution ( \delta R\propto c/2B ) but also dramatically improves SNR through Fourier-domain processing gain, which confines signal energy to a narrower spectral bin against broadband noise background. The enhancement is highlighted in Fig. 3 d: the B=7.49 THz configuration (green trace) yields 12 dB suppression of the noise floor compared to the B=125 GHz baseline (red trace). We calculate the ranging precision \sigma as the standard deviation to the mean, quantified via histogram analysis in Fig. 3 e. Under clear condition OD {}_{1064}=0 , both the NIR and narrowband ( B=125 GHz) MIR systems achieve similar precision of \sigma\approx 15 µm. However, as turbidity increases, precision in narrowband systems degrades significantly due to SNR deterioration: \sigma=97.2 µm for OD {}_{1064}=2 in the NIR, \sigma=87.5 µm for OD {}_{1064}=5 in the MIR. Conversely, the wideband ( B=7.49 THz) MIR system demonstrates exceptional resilience, maintaining \sigma=1.76 µm precision in clear air and degrading only marginally to \sigma=2.19 µm even under extreme attenuation (OD {}_{1064}=5 ). These results establish high-bandwidth MIR FMCW LiDAR as a robust solution for precision ranging in heavily scattering environments. By penetrating dense particulate matter, this technology paves the way for all-weather autonomous driving and critical emergency applications such as imaging through smoke-filled buildings or penetrating dense cloud cover to aid search-and-rescue missions. Figure 4: Simultaneous ranging and chemical identification from a single optical waveform . a . Conceptual illustration of the dual-modality perception: the system measures the physical distance to a target while simultaneously analyzing the chemical composition of the intervening medium. b . Experimental testbed. An industrial emission scenario is simulated using a multi-pass gas cell filled with an HCl mixture. PD, photodetector. BS, beam splitter. c . Signal decoupling principle. The interferometric beat signal carries orthogonal information channels: range is encoded in the frequency modulation (retrieved via fast Fourier transform, FFT), while spectral fingerprints are encoded in the amplitude envelope (extracted via Hilbert transform, HT). d . Reconstructed LiDAR range profile, pinpointing the target location at 1.32 m. e . Quantitative spectroscopic analysis. The extracted transmission spectrum resolves the distinct rotational-vibrational transitions of HCl isotopologues (H 35 Cl and H 37 Cl) and background atmospheric H 2 O vapor. Multi-component spectral fitting yields concentrations of 82 ppm for HCl and 0.50% for H 2 O. Spectroscopic ranging . The escalating complexity of urban environments necessitates a paradigm shift in sensing technologies—moving beyond simple geometric mapping to multi-dimensional perception that simultaneously captures chemical composition 36 . To bridge this gap, we exploit our system’s 3–3.7 µm coverage to function as both a precision ranger and a metrological gas sensor, as illustrated in Fig. 4 a. This dual capability is vital for monitoring hazardous pollutants such as hydrogen chloride (HCl)—a toxic byproduct of fossil fuel combustion and waste incineration that threatens human health and ecosystem. Our approach redefines information extraction from the FMCW LiDAR signal, as shown in Fig. 4 c. Conventionally, LiDAR systems extract range information solely from the beat frequency via fast Fourier transform (FFT), typically discarding amplitude variations as noise or fading. However, in the MIR regime, the signal envelope encodes the spectral absorption “fingerprint” of the medium. To recover this information, we decouple the absorption-induced intensity variations from the frequency-modulated beat signal using Hilbert transform (HT) 48 . Algorithmic details are found in Supplementary Materials Note 7. Crucially, this method leverages the heterodyne gain of coherent detection, where a strong local oscillator amplifies the weak return signal. This enhancement enables metrology-grade spectroscopic retrieval even from non-cooperative, scattering targets. We validate this dual-modality perception using a testbed illustrated in Fig. 4 b. The MIR beam is split. In the measurement arm it traverses a multi-pass gas cell filled with an HCl mixture before reflecting off a steel target. The back-reflected beam traverses the cell again before mixing with the beam in the reference arm. This setup simulates the remote detection of a pollutant cloud against a solid background. Processing a single waveform yields two distinct datasets. First, the FFT of the beat signal accurately resolves the target distance at 1.32 m, as shown in Fig. 4 d. Second, the amplitude retrieval via HT reconstructs the path-integrated transmission spectrum, as shown in Fig. 4 e. The retrieved spectrum agrees with the HITRAN database 22 , resolving fine spectral features including distinct HCl isotopologues (H 35 Cl and H 37 Cl) and interference from atmospheric H 2 O vapor. Multi-component fitting 54 yields concentrations of 82 ppm for HCl and 0.50% for H 2 O, confirming the system’s high sensitivity and spectral resolution. This capability for continuous, broadband, and metrology-grade measurement is vital for analyzing complex gas mixtures. By integrating spectroscopy with LiDAR ranging, our system offers a multi-dimensional perception tool for urban environmental monitoring, enabling not just the detection of hazardous emissions, but their precise localization and identification in three-dimensional space. Figure 5: Performance benchmarking and scalable multi-spectral synthesis . a . Comparison of mode-hop-free (MHF) tuning ranges versus operating wavelengths for state-of-the-art CW lasers 35 ; 16 ; 12 ; 60 ; 15 ; 78 ; 21 ; 13 ; 64 . Data points are classified by frequency calibration capability: absolute (stars), relative (triangles), and none (circles). Our system establishes a new performance benchmark, offering the widest continuous tuning range combined with absolute frequency accuracy. b . Spectral coverage map spanning the visible to MIR (0.4–5 µm). Solid blocks indicate bands experimentally generated in this work, including the primary NIR pumps (1020–1098 nm and 1260–1640 nm) and their nonlinear conversion products: SHG (518–541 nm and 766–795 nm), SFG (618–644 nm), and DFG (3001–3711 nm). Dashed frames denote theoretically accessible regions via Raman amplification or further mixing with auxiliary sources such as 517 THz optically pumped semiconductor lasers, 474 THz HeNe lasers, 400 THz diode lasers or 333 THz diode lasers. This demonstrates the platform’s potential for seamless, gap-free spectral synthesis across the entire 0.4–5 µm landscape. Conclusion . We have demonstrated that the historical trade-off between spectral bandwidth, photon flux, and frequency fidelity in the MIR is no longer a fundamental limitation. By fusing widely tunable NIR CW lasers with rigorous calibration traceable to atomic standards, our DFG-based architecture provides a deterministic tool for continuous interrogation in the 3–3.7 µm spectral window. Benchmarking against state-of-the-art sources listed in Fig. 5 a, our platform occupies a unique performance space: it offers the widest continuous tuning coverage combined with absolute frequency accuracy, essentially democratizing metrology-grade spectroscopy. The transformative potential of this platform is evidenced by its immediate impact across distinct fields. In integrated photonics, the system overcomes the critical spectral limits of Si 3 N 4 by pinpointing and eliminating hydrogen-induced absorption. This elimination of extrinsic impurities unmasks multi-phonon absorption in the silica cladding as the fundamental barrier, establishing enhanced mode confinement as the critical pathway to unlock the platform’s full MIR potential. Simultaneously, in free-space sensing, it bridges the gap between spatial ranging and chemical analysis, enabling LiDAR systems that penetrate opaque scattering media while simultaneously identifying hazardous pollutants within a single optical waveform. Looking forward, this architecture is inherently scalable. As illustrated in Fig. 5 b, we have achieved measurement bandwidths spanning the primary NIR pumps ( 1020 – 1098 nm and 1260 – 1640 nm) and their nonlinear conversion products—including second-harmonic generation (SHG: 518 – 541 nm and 766 – 795 nm, see Supplementary Materials Note 1), sum-frequency generation (SFG: 618 – 644 nm, see Supplementary Materials Note 8), and DFG ( 3001 – 3711 nm)—covering visible to MIR. The current system represents just one node in a broader spectral synthesis scheme. By employing different combinations of nonlinear processes, the same NIR pump lasers can be converted to cover the 0.4–5 µm spectrum. This paves the way for a unified, ultra-broadband spectroscopic engine capable of accessing electronic, vibrational, and rotational transitions across the entire molecular landscape. Ultimately, by converging metrological frequency fidelity with the versatility of nonlinear optics, our work establishes a new foundation for multi-dimensional perception, unlocking the full information capacity of the MIR spectrum. Methods Extended Data Figure 1: Phase-matching mechanisms and broadband mid-infrared synthesis performance . a . Comparison of birefringent phase-matching (BPM), quasi-phase-matching (QPM), and chirped quasi-phase-matching (CQPM). Top: wave-vector ( k ) diagrams; Bottom: MIR power evolution along the waveguide. Unlike BPM and uniform QPM, which exhibit monotonic power growth only at a single phase-matched wavelength (blue regions) and oscillatory behavior elsewhere (red regions), CQPM employs a spatially varying poling period \Lambda(z) to localize phase-matching conditions at different longitudinal positions z , enabling broadband DFG. b . Real-time frequency tuning profiles calibrated via fiber cavities and atomic hyperfine transitions. Chirping laser #1 scans from 1035 nm to 1042 nm (2.01 THz bandwidth), and chirping laser #2 scans from 1580 nm to 1536 nm (5.47 THz bandwidth). The counter-directional chirping of the NIR lasers synthesizes a continuous, mode-hop-free MIR tuning bandwidth of 7.49 THz (3001–3243 nm) in 0.89 s. c . Full spectral coverage. Step-tuning the start wavelength of laser #1 (1035, 1057, and 1079 nm) enables seamless coverage from 3001 nm to 3711 nm. These tuning curves are characterized with laser #1 power P_{1}=2.3 W and laser #2 power P_{2}=3.0 W. d . Power transfer characteristics. MIR output power versus pump power P_{2} with fixed \lambda_{1}=1070 nm, \lambda_{2}=1550 nm, and P_{1}=4.0 W. A maximum MIR power of 45.0 mW is generated at P_{2}=5.6 W. Broadband mid-infrared synthesis . The MIR light is generated via DFG in a CPLN waveguide. To overcome the bandwidth constraints of traditional phase-matching, we employ a chirped poling strategy. In birefringent phase-matching (BPM) and uniform quasi-phase-matching (QPM) 79 ; 77 , momentum conservation \Delta k=0 is strictly satisfied only for a narrow wavelength range. As illustrated in the wave-vector ( k ) diagrams in Extended Data Fig. 1 a, a fixed grating vector 2\pi/\Lambda cannot compensate for the phase mismatch induced by pump detuning, resulting in oscillatory power transfer rather than cumulative growth. In contrast, our device utilizes a chirped quasi-phase-matching (CQPM) profile where the poling period \Lambda(z) varies continuously along the propagation axis z . The spatially dependent grating vector 2\pi/\Lambda(z) ensures that for any pump pair ( k_{1} and k_{2} ) within the design bandwidth, the phase-matching condition is spatially localized to a specific longitudinal position. Consequently, distinct spectral components initiate conversion at different positions along the waveguide while maintaining high output power. To synthesize the broadband MIR laser, chirping laser #1 scans from 1035 nm to 1042 nm (2.01 THz), while chirping laser #2 scans in the opposing direction from 1580 nm to 1536 nm (5.47 THz). This configuration synthesizes an MIR scan from 3001 nm to 3243 nm, achieving a mode-hop-free tuning bandwidth of 7.49 THz in 0.89 seconds, as depicted in Extended Data Fig. 1 b. By discretely stepping the start wavelength of laser #1 (1035, 1057, and 1079 nm), the MIR output achieves seamless coverage from 3001 nm to 3711 nm, as shown in Extended Data Fig. 1 c. The total time for full bandwidth scanning is under 5 seconds. Extended Data Fig. 1 d shows that the system delivers a maximum output power of 45.0 mW with pump parameters \lambda_{1}=1070 nm ( P_{1}=4.0 W) and \lambda_{2}=1550 nm ( P_{2}=5.6 W). Extended Data Figure 2: Experimental setup and frequency metrology performance . a . Schematic of the spectroscopic system, functionally divided into five regions: Regions I and II track relative frequency calibration of the two NIR pump lasers via fiber cavities and UMZIs; Regions III and V establish absolute frequency referencing via frequency doubling to the visible/near-visible bands; and Region IV generates the MIR DFG output. PD, photodetector; PBS, polarizing beam splitter; QWP, quarter-wave plate; DAQ, data acquisition; PM, phase modulator. Bottom-right inset illustrates the principle of saturated absorption spectroscopy (SAS) and modulation transfer spectroscopy (MTS). b . Absolute frequency validation for chirping laser #1 (Region III, 518–541 nm). Resolved hyperfine transitions of I 2 R(53)31-0 line, measured via MTS, shows a residual RMS deviation of \delta_{\text{SH1}}=11.7 MHz relative to reference data (red dashed lines). c . Absolute frequency validation for chirping laser #2 (Region V, 766–795 nm). Resolved hyperfine transitions of K D1 line, measured via SAS, show a residual RMS deviation of \delta_{\text{SH2}}=8.2 MHz relative to reference data (red dashed lines). d . Long-term stability of the fiber cavity’s FSR. Histogram of FSR values measured at 1490 nm over a 24 hours indicates a standard deviation of \sigma_{\text{FSR}}=114.3 Hz. Spectroscopic system setup and relative frequency tracking . The experimental architecture, detailed in Extended Data Fig. 2 a and Supplementary Note 1, is designed for simultaneous multi-spectral operation across five distinct regions (I–V). Regions I (1020–1098 nm) and II (1260–1640 nm) represent the fundamental outputs of the two NIR ECDLs (chirping lasers #1 and #2). These NIR seed sources drive the generation of three derivative bands via nonlinear frequency conversion: visible output (Region III, 518–541 nm, via SHG), MIR output (Region IV, 3001–3711 nm, via DFG), and near-visible output (Region V, 766–795 nm, via SHG). Although the NIR seed lasers possess broad tuning ranges, the system’s effective operational bandwidth is practically confined by the gain profiles of the YDFA (1035–1086 nm) and the EDFA (1532–1590 nm). Consequently, only spectral components residing within these gain windows achieve the power levels necessary to pump the CPLN waveguides, where CQPM conditions further select the efficient conversion bands. This configuration specifically maps the amplified fundamental NIR scans to the target frequencies as follows: 1036–1082 nm to 518–541 nm (Region III); 1532–1590 nm to 766–795 nm (Region V); and combined pumps of 1035–1086 nm and 1536–1580 nm to 3001–3711 nm (Region IV). In Regions I and II, we perform relative frequency calibration by referencing the chirping lasers #1 and #2 to fiber cavities. The method is illustrated in Ref. 46 and details are found in Supplementary Materials Note 1. The fiber cavities’ FSR values (32.32 MHz for laser #1 and 55.58 MHz for laser #2) are precisely characterized as relative frequency rulers. By recording and counting the number of fiber resonances passed by the chirping laser, the relative frequency excursion of the laser is calculated. While the fiber resonances provide discrete frequency markers, sub-FSR instantaneous frequencies are resolved by interpolating interference fringes from an unbalanced Mach-Zehnder interferometer (UMZI) 1 . Using the Hilbert transform 48 to extract the instantaneous phase \phi(t) , we calculate the phase difference \Delta\phi between adjacent sampling points. The corresponding frequency excursion \Delta f is then derived as: \Delta f=\frac{\text{FSR}}{\phi_{\text{FSR}}}\cdot\Delta\phi, (1) where \phi_{\text{FSR}} is the phase change corresponding to one cavity FSR. Atomic referencing for absolute frequency traceability . We frequency-double the fundamental NIR lasers via CPLN waveguides to access atomic reference lines. For chirping laser #1, SHG targets the 518–541 nm band (Region III in Extended Data Fig. 2 a), where iodine molecules (I 2 ) provide a dense manifold of hyperfine transitions. We employ modulation transfer spectroscopy 66 ; 8 to resolve these transitions with sub-megahertz precision, utilizing a counter-propagating pump-probe configuration to generate error signals with high discriminator slopes (see Supplementary Materials Note 1 for details). To establish absolute frequency references, we utilize the hyperfine transitions of the I 2 R(56)32-0, R(53)31-0, and P(28)30-0 lines, for which experimental frequency values are available 2 ; 8 ; 65 and are provided in Supplementary Materials Note 1. The absolute frequency axis is anchored to the I 2 R(56)32-0 line ( f_{\text{ref,1}}=563.2600 THz) 2 . We use the hyperfine transitions of the I 2 P(28)30-0 line ( f_{\text{ref,2}}=560.1518 THz) to validate the relative frequency calibration derived from the fiber cavity. We observe a root-mean-square (RMS) deviation of \Delta_{1}=-35.8 MHz between our measured values and those in ref. 65 , which we attribute to residual dispersion in the fiber cavity’s FSR. Consequently, we apply a correction to the second-harmonic (SH) laser frequency using: f^{\prime}_{\text{SH1}}=\frac{f_{\text{SH1}}-f_{\text{ref,1}}}{1+\epsilon_{1}}+f_{\text{ref,1}} (2) where \epsilon_{1}=\Delta_{1}/(f_{\text{ref,1}}-f_{\text{ref,2}}) . Applying Eq. 2 to the measurement of the I 2 R(53)31-0 transitions ( f_{\text{ref,3}}=561.4205 THz), as shown in Extended Data Fig. 2 b, yields a residual RMS deviation of 11.7 MHz relative to literature values 8 . This determines the SH laser accuracy as \delta_{\text{SH1}}\approx 11.7 MHz, corresponding to a fundamental frequency accuracy for chirped laser #1 of \delta_{\text{laser1}}=\delta_{\text{SH1}}/2\approx 5.9 MHz. For chirping laser #2, SHG targets the 766–795 nm band (Region V in Extended Data Fig. 2 a), where Rubidium (Rb) and Potassium (K) atoms provide distinct hyperfine transitions. We employ saturated absorption spectroscopy to resolve these transitions with sub-megahertz precision (see Supplementary Materials Note 1 for details). To establish absolute frequency references, we utilize the hyperfine transitions associated with the Rb D2, K D1, and K D2 lines, whose frequencies have been experimentally measured 69 ; 71 and are provided in Supplementary Materials Note 1. The absolute frequency axis is anchored to the Rb D2 line ( f_{\text{ref,4}}=384.2286 THz) 69 . The hyperfine transitions associated with the K D2 line ( f_{\text{ref,5}}=391.0165 THz) 71 are used to examine the accuracy of our relative frequency calibration with the fiber cavity. An RMS deviation of \Delta_{2}=28.4 MHz is found between our measured values and the values in ref. 71 . We further correct the SH laser’s frequency as: f^{\prime}_{\text{SH2}}=\frac{f_{\text{SH2}}-f_{\text{ref,4}}}{1+\epsilon_{2}}+f_{\text{ref,4}} (3) where \epsilon_{2}=\Delta_{2}/(f_{\text{ref,4}}-f_{\text{ref,5}}) . Applying Eq. 3 to the measurement of the K D1 line ( f_{\text{ref,6}}=389.2861 THz), as shown in Extended Data Fig. 2 c, yields a residual RMS deviation of 8.2 MHz relative to literature values 71 . This determines the SH laser accuracy as \delta_{\text{SH2}}\approx 8.2 MHz, corresponding to a fundamental accuracy for chirping laser #2 of \delta_{\text{laser2}}=\delta_{\text{SH2}}/2\approx 4.1 MHz. Quantification of accuracy and precision . In Extended Data Fig. 2 a Region IV, the MIR laser is generated via DFG. The absolute frequency accuracy of the MIR laser is determined by the uncorrelated uncertainties of the pump lasers: \delta_{\text{MIR}}=\sqrt{\delta_{\text{laser1}}^{2}+\delta_{\text{laser2}}^{2}}\approx 7.2\text{ MHz}. (4) The system’s precision is ultimately governed by the stability of the fiber cavity FSR. As presented in Extended Data Fig. 2 d, long-term monitoring at 1490 nm over 24 hours reveals an FSR standard deviation of \sigma_{\text{FSR}}=114.3 Hz. Over the full 5.47 THz tuning range of chirping laser #2 ( N\approx 10^{5} FSR steps), the cumulative frequency uncertainty scales as \sigma_{\text{cumulative}}=\sigma_{\text{FSR}}\sqrt{N}\approx 36.1\text{ kHz} . For chirping laser #1, the narrower tuning range (2.01 THz) implies a correspondingly smaller cumulative error contribution. Since the MIR frequency is synthesized via DFG, the aggregate frequency uncertainty remains dominated by laser #2 and is significantly smaller than the dynamic linewidth of the MIR laser (242 kHz, see Supplementary Materials Note 3). Consequently, the system’s precision is limited by the laser linewidth rather than the frequency calibration. Microresonator dispersion and loss characterization . We extract integrated dispersion D_{\text{int}} by measuring and fitting resonance frequencies to the polynomial: D_{\text{int}}(\mu)=\omega_{\mu}-\omega_{0}-D_{1}\mu=\sum_{n=2}^{\cdots}\frac{D_{n}\mu^{n}}{n!} (5) where \omega_{\mu}/2\pi is the frequency of the \mu -th mode relative to the reference mode \omega_{0}/2\pi , D_{1}/2\pi is the microresonator FSR, D_{2} describes group velocity dispersion (GVD), and D_{3},D_{4} etc describe higher-order dispersion terms. The group index is derived as n_{g}=c/(D_{1}L) , with L=1.429 mm being the physical cavity length of the microresonator in Fig. 2 . For loss quantification, resonances are fitted to a Lorentzian model 3 ; 58 : T(\omega)=1-\frac{4\kappa_{\text{ex}}\kappa_{0}}{\Delta\omega^{2}+(\kappa_{\text{ex}}+\kappa_{0})^{2}} (6) where \Delta\omega/2\pi is the laser detuning, \kappa_{0}/2\pi is the intrinsic loss rate, and \kappa_{\text{ex}}/2\pi is the external coupling strength. The optical propagation loss (in dB m -1 ) in the waveguide is calculated as \alpha\approx(27.27\cdot n_{g})/(\lambda\cdot Q_{0}) , where Q_{0}=\omega/\kappa_{0} is the intrinsic Q -factor. Extended Data Figure 3: Waveguide geometry and modal confinement . a . Cross-sectional false-color SEM image of the Si 3 N 4 waveguide encapsulated by SiO 2 cladding. The Si 3 N 4 core is highlighted in red. b . Simulated fundamental TE 00 mode profile at 81 THz ( \lambda\approx 3.7 µm), displaying the normalized electric field magnitude |E| . Finite-element simulations of geometric leakage and optical confinement . To quantify radiative losses arising from waveguide bending and substrate leakage, we perform finite-element method (FEM) simulations in a cylindrical coordinate system. Extended Data Fig. 3 a shows the measured SEM of the waveguide cross-section, which is subsequently used to model the FEM simulation. The computational domain models the Si {}_{3}N_{4} waveguide (width w=3000 nm, height h=1050 nm) with SiO 2 cladding on a silicon (Si) substrate, with boundaries extended 1 µm into the air and Si regions to ensure sufficient field decay. To isolate geometric leakage from material attenuation, we employ lossless material models utilizing only the real parts of the refractive indices for Si 3 N 4 , SiO 2 , and Si. We evaluate the radiative leakage via the imaginary part of the effective refractive index, n^{\prime\prime} . At 81 THz, simulation for a bending radius of r=40 µm yields n^{\prime\prime}\approx 3.5\times 10^{-6} . Increasing the radius to 150 µm significantly reduces this value to n^{\prime\prime}\approx 7.8\times 10^{-7} . Further expansion to r=200 µm results in negligible change ( n^{\prime\prime}\approx 7.7\times 10^{-7} ), indicating that curvature-induced bending loss vanishes at the substrate leakage floor. Consequently, for the experimental microresonator radius of r=227.5 µm, the bending loss contribution is minimal. The residual substrate leakage of n^{\prime\prime}\approx 7.7\times 10^{-7} corresponds to a propagation loss of \alpha=4\pi n^{\prime\prime}/\lambda\cdot 10\log_{10}e\approx 11.4 dB m -1 . This theoretical limit is over an order of magnitude lower than the experimentally measured background of \sim 420 dB m -1 , confirming that material absorption—rather than structural leakage—is the dominant loss mechanism. To examine the interaction between the optical mode and the cladding, we analyze the Poynting vector distributions. Extended Data Fig. 3 b illustrates the normalized electric field magnitude |E| of the fundamental transverse-electric (TE 00 ) mode at 81 THz. Integration of the power flux reveals that the fraction of optical power propagating within the cladding rises from 21% at 100 THz to 31% at 81 THz. This trend confirms reduced optical confinement at lower frequencies. Given the onset of strong multi-phonon absorption in SiO 2 around 86 THz (3.5 µm) 68 ; 33 ; 50 , the increased mode overlap at 81 THz ( \sim 3.7 µm) is consistent with the observed attenuation trend, indicating that interaction with the SiO 2 cladding is a substantial source of loss. Fog chamber . To simulate fog conditions, we use a 1.4-m-long environmental chamber equipped with an ultrasonic atomization unit. Water micro-droplets are introduced via a centrifugal fan, with scattering density regulated by a precision intake valve. A passive exhaust port enables natural atmospheric equilibration to maintain steady-state flow. Optical access is provided by broadband CaF 2 windows mounted at \sim 10 ∘ angle relative to the optical axis to suppress etalon resonances. Resistive heating maintains window surface temperatures above the dew point to prevent condensation-induced scattering. Funding Information : We acknowledge support from the National Key R&D Program of China (Grant No. 2024YFA1409300), National Natural Science Foundation of China (Grant No.12261131503, 12404417, 12404436, 62075233), Quantum Science and Technology–National Science and Technology Major Project (2023ZD0301500), Shenzhen-Hong Kong Cooperation Zone for Technology and Innovation (HZQB-KCZYB2020050), Shenzhen Science and Technology Program (Grant No. RCJC20231211090042078), Guangdong-Hong Kong Technology Cooperation Funding Scheme (Grant No. 2024A0505040008), the Key R&D Plan of Shandong Province (Grant No. 2021ZDPT01), Natural Science Foundation of Shandong Province (ZR2021LLZ013, ZR2022LLZ009), and CAS Project for Young Scientists in Basic Research (YSBR-69).