Infrared imaging in the mid-wave infrared (MWIR) (3–5 µm) and long-wave infrared (LWIR) (8–12 µm) bands has attracted significant attention due to its crucial role in defense, remote sensing, and industrial inspection^[1-3]. MWIR imaging offers high spatial resolution, making it well-suited for long-range target recognition under atmospheric transmission windows^[4]. In contrast, LWIR is highly effective in detecting passive thermal radiation, which is essential for night vision and environmental monitoring^[5]. The combination of these spectral bands provides both detailed spatial resolution and enhanced thermal sensitivity, creating a strong demand for MWIR-LWIR dual-band imaging systems that can perform reliably in diverse operational environments.

Conventional MWIR-LWIR dual-band imaging systems rely on refractive lens (RL) assemblies or refractive-diffractive hybrid optics (RDHOs)^[6-8]. While refractive systems use cascaded germanium or chalcogenide elements to correct chromatic aberrations, they often increase system weight and complexity^[9-12]. RDHOs mitigate lens count through diffractive optics elements (DOEs)^[13-15], but their broadband efficiency is constrained by wavelength-dependent phase errors^[16,17]. Furthermore, conventional architectures enforce a common optical path for both spectral bands, exacerbating inter-band crosstalk and restricting independent aberration control.

Recent advancements in meta-optics provide transformative solutions to these challenges. Metalenses, engineered with subwavelength microstructures, enable precise dispersion control and multifunctional wavefront manipulation^[18-22]. Unlike conventional DOEs, dielectric metalenses achieve high efficiency across broad infrared bands by tailoring resonant modes of high-index microstructures^[23-25]. Hybrid refractive-metalens architectures leverage the broadband capabilities of refractive optics while benefiting from metalenses’ compactness and dispersion control, offering a promising pathway for multispectral imaging^[26-32].

In this study, we propose a spatially decoupled MWIR-LWIR dual-band imaging system based on hybrid refractive-diffractive-metasurface optics. The system integrates a silicon-based metalens for the MWIR channel and a high-refractive-index chalcogenide glass lens with embedded diffractive structures for the LWIR channel. The large-aperture chalcogenide lens is optimized for high-efficiency LWIR transmission and broadband aberration correction, while the compact metalens is engineered for MWIR focusing with enhanced dispersion control. By axially aligning these optical components while preserving their spectral independence, the system effectively suppresses inter-band crosstalk and enables tailored chromatic correction for each band. Furthermore, a wavelength-dependent phase compensation strategy is introduced in the metalens design to match the group delay of meta-atoms, thereby improving MWIR achromatic performance. To enhance spatial detail recovery, the system integrates a physics-informed StableSR framework with low-rank adaptation (LoRA) fine-tuning. This architecture demonstrates a high-performance solution for compact, lightweight, and spectrally discriminative dual-band infrared imaging, with promising potential for advanced multispectral applications.

The MWIR-LWIR dual-band imaging system demands a precise optical design to achieve compactness, high efficiency, and minimal aberrations. This section presents the principles and detailed design of our hybrid optical system, which spatially separates the focal planes for MWIR and LWIR imaging. The system architecture integrates a silicon-based metalens with a high-refractive-index chalcogenide refractive-diffractive lens, optimizing the optical configuration for the independent operation of both spectral bands. By leveraging metasurface dispersion engineering and tailored diffractive structures, the system enhances broadband aberration correction.

To achieve compactness and high imaging performance, we propose a spatially decoupled dual-band infrared optical architecture based on hybrid refractive-diffractive-metasurface optics. As shown in Fig. 1(a), the system comprises a large-aperture chalcogenide RL integrated with a DOE for LWIR imaging, a silicon-based metalens for MWIR imaging, and a silicon wafer substrate that supports the metalens.

The LWIR channel employs a double-sided aspheric chalcogenide lens made of Ge-Ga-Ag-Te (GGAT or NBU-10) glass^[33], featuring a high refractive index (n∼3.5) and high broadband transmission across the 3–12 µm range, as depicted in Fig. 1(b). The lens has an 8.7 mm entrance pupil diameter and a focal length of 13 mm, and provides a 30° field of view (FOV). To compensate for chromatic aberrations over the 8–12 µm band, a DOE diffraction ring structure is engraved on the posterior surface of the RL, with a depth of 2.07 µm optimized for a 10 µm central wavelength. The optical surfaces are refined using a damped least-squares optimization algorithm in Zemax. After applying an anti-reflection (AR) coating, the GGAT glass maintains an average transmittance of 81.49% across the 3–12 µm range, ensuring high optical efficiency [Fig. 1(b)].

The MWIR channel is realized by placing a silicon-based metalens posterior to the RL. The metalens is fabricated on a 0.5-mm-thick silicon wafer, which acts as a mechanical support and optical substrate. After AR coating, the silicon wafer exhibits an average transmission efficiency of 88.28% in the 3–12 µm band, as shown in Fig. 1(c). The metalens phase profile is modeled using the binary_2 surface type in Zemax and expressed as an even-order polynomial function of the normalized radial coordinate: φ(r,λ)=∑i=1NAi(rρ)2i,(1)where r is the radial coordinate, ρ is the normalized radius (0.5 mm), and Ai are the polynomial coefficients. The optimized phase distributions for different wavelengths (3.0, 3.2, 3.5, 3.7, and 4.0 µm) are shown in Fig. 1(d).

The MWIR channel performance is critically influenced by the combined action of the silicon-based metalens and the supporting refractive-diffractive elements. As shown in Figs. 1(e) and 1(f), we evaluate how the front-end LWIR RL and DOE structure enhance the imaging quality of the silicon-based metalens. We first assess the MWIR channel without the RL and DOE. As shown in Fig. 1(e), the modulation transfer function (MTF) for 3–4 µm wavelengths sharply decreases with increasing the spatial frequency (SF), dropping to zero near 10 lp/mm. This degradation indicates that the standalone metalens lacks sufficient optical power and aberration correction across the FOV. After introducing the RL and DOE elements, the MWIR imaging performance improves significantly. As illustrated in Fig. 1(f), the MTF for 3.5 µm wavelength remains stable, exceeding 0.5 at an SF of 20 lp/mm across the 0°–10° FOV range. The MTFs for 3.0, 3.2, 3.7, and 4.0 µm are shown in Fig. S1, Supplement 1. In addition, the RL and DOE assist in suppressing residual aberrations and chromatic aberrations throughout the entire MWIR channel. Field curvature and distortion characteristics at 3.0 and 4.0 µm are shown in Fig. S2, Supplement 1. These results demonstrate that the RL and DOE collectively reshape the incident wavefront, enabling the metalens to operate under favorable optical conditions. This hybrid design ensures broadband achromatic focusing across the 3–4 µm spectral range, substantially enhancing both on-axis and off-axis imaging performance.

For the LWIR channel, a DOE structure is designed and engraved on the rear surface of the RL to enhance achromatic performance. The phase profiles provided by the DOE at different wavelengths (8–12 µm) are illustrated in Fig. 1(g). The LWIR channel performance is primarily determined by the chalcogenide RL and its integrated DOE structure. As shown in Figs. 1(h) and 1(i), we evaluate the impact of the DOE on broadband aberration correction and imaging quality across the LWIR spectral band. Without the DOE structure, the LWIR channel exhibits degraded MTF performance. As shown in Fig. 1(h), the MTFs for 8–12 µm wavelengths decline with increasing SF, highlighting the limitations of using a single RL for broadband LWIR imaging. After integrating the DOE onto the posterior surface of the lens, the LWIR MTF improves significantly. As shown in Fig. 1(i), the MTF remains above 0.3 at 10 lp/mm across the entire 0°–30° FOV when centered at a wavelength of 10 µm. Further improvements in MTF performance at wavelengths of 8.0, 9.0, 11.0, and 12.0 µm wavelengths are presented in Fig. S3, Supplement 1. In addition, the DOEs assist in suppressing residual aberrations and chromatic aberrations across the entire LWIR channel. Field curvature and distortion characteristics from 8.0 to 12.0 µm are detailed in Fig. S4, Supplement 1. These results demonstrate that the DOE structure provides effective phase compensation, thereby improving field uniformity and enhancing system resolution. The integration of the DOE enables robust broadband performance within the LWIR channel using a single refractive-diffractive element, achieving partial chromatic aberration correction and maintaining uniform imaging quality over a wide spectral range.

As shown in Fig. 1(a), although the hybrid refractive-diffractive-metasurface optical system is divided into MWIR and LWIR channels, there is no physical filtering separation between them. As a result, MWIR beams can still propagate through the LWIR optical path, and LWIR beams can also pass through the MWIR path. To verify whether beams outside the intended channel introduce significant interference, we investigate the occurrence of inter-band crosstalk in both the MWIR and LWIR channels. Specifically, to assess potential crosstalk in the MWIR channel, the system is illuminated with LWIR-band wavelengths, and the resulting focal plane behavior is observed. When the FOVs are 0° and 30°, the corresponding root mean square (RMS) spot radii are 846.64 and 1026.86 µm, respectively, as shown in Fig. S5, Supplement 1, both significantly exceeding 500 µm. In contrast, the ideal RMS radii for the MWIR channel at these angles are only 17 and 12 µm. The relative intensity of crosstalk is approximately 2%. The highly dispersed spot profiles indicate that LWIR leakage has a negligible impact on MWIR imaging performance. Similarly, to evaluate crosstalk in the LWIR channel, we illuminate the system with MWIR-band wavelengths and observe the resulting focal plane distribution. When the FOVs are 0° and 10°, the RMS spot radii are measured to be 2101.22 and 2015.69 µm, respectively, as shown in Fig. S6, Supplement 1. In comparison, the ideal RMS radii for the LWIR channel are 53 and 143 µm, respectively. The maximum relative crosstalk intensity is approximately 7.1%, indicating that the influence of MWIR leakage on LWIR imaging performance is weak and does not cause significant degradation.

While the methodology in Sec. 2.1 enables the independent optimization of ideal phase profiles for MWIR focusing at different wavelengths through Zemax simulations, it remains a purely numerical approach rooted in ray optics. Zemax offers a powerful environment for designing continuous phase distributions; however, it inherently assumes ideal, smooth phase surfaces and does not account for the discrete and resonant nature of subwavelength-scale meta-atoms. When applied to metalens design, this Zemax-based approximation neglects the fact that each physical meta-atom imparts a specific, wavelength-dependent group delay governed by its geometry and material dispersion. As a result, the phase profile optimized in Zemax—though numerically achromatic—may significantly deviate from what can be physically realized, thereby introducing residual dispersion and chromatic aberrations upon fabrication. This approach effectively imposes idealized dispersion characteristics on the metasurface, which may exceed the practical capabilities of the actual meta-atoms. In particular, the stringent group delay demands derived from the Zemax phase function can surpass the realizable delay range of the designed nanostructures, posing substantial challenges to physical implementation.

To address the group delay mismatch between the Zemax-optimized theoretical phase distribution and the actual meta-atom response, we introduce a spectral phase compensation framework. In this framework, a wavelength-dependent phase correction term C(λ) is incorporated to align the realized group delay with the ideal distribution, thereby enabling precise dispersion engineering and broadband achromatic focusing across the MWIR band. Accordingly, the phase function is reformulated as φc(r,λ)=∑i=1NAi(rρ)2i+C(λ),C(λ)∈[0,2π],(2)where C(λ) compensates for the group delay variations caused by the finite height of the meta-atoms. This correction ensures that the metalens phase response aligns with the desired dispersion profile and matches the group delay of the meta-atom array across the MWIR spectral band.

To determine the optimal values of C(λ) at different wavelengths, we conducted finite-difference time-domain simulations using circular cylinder Si microstructures with a period of 1 µm and a height of 3.1 µm [Fig. 2(a)]. The outer diameter (R) varies from 0.1 to 0.5 µm, while the inner diameter (r) ranges from 0 to 0.48 µm, thereby forming an experimental phase database for the microstructures. A particle swarm optimization (PSO) algorithm is then applied to optimized the C(λ) values by minimizing the wave aberration between the ideal phase and the actual phase provided by the microstructures. The optimized C(λ) values for MWIR wavelengths are as follows: C(3.0μm)=2.59, C(3.2μm)=0.86, C(3.5μm)=5.30, C(3.7μm)=3.91, and C(4.0μm)=2.64. The ideal phase distributions incorporating C(λ) corrections at 3.0, 3.2, 3.5, 3.7, and 4.0 µm are shown in Fig. 2(b). Figure 2(c) presents the comparison between the ideal phase (solid black lines) and the realized phase profiles from the metalens (red circles) at the corresponding wavelengths. The transmittance of the metalens, as arranged by the PSO-optimized meta-atom configuration, exceeds 70% on average across the 3–4 µm spectral range, as shown in Fig. S7, Supplement 1. Furthermore, Fig. 2(d) illustrates the outer and inner radius distributions of the meta-atoms across the metalens diameter. Figures 2(e) and 2(f) present a two-dimensional view of the designed achromatic metalens and an enlarged view of the circular cylinder structures within the marked region. This refined approach enables precise phase control and effective dispersion engineering, ensuring that the metalens meets the rigorous demands of broadband MWIR imaging.

To evaluate the imaging performance of the MWIR-LWIR dual-band hybrid metalens system under realistic conditions, we conducted imaging simulations incorporating both the actual phase profiles of the metalens [represented by red circles in Fig. 2(c)] and continuous FOV conditions (0°–10° for MWIR and 0–30° for LWIR). These simulations closely mimic real-world imaging scenarios by considering both the PSF and the relative illuminance (RI) across the image plane.

For the MWIR channel, Fig. 3(a) and Fig. S8, Supplement 1, present the normalized and raw PSF intensity distributions on a 25 pixel × 25 pixel grid, evaluated at FOV angles of 0°, 2°, 4°, 6°, 8°, and 10°. To facilitate comparative analysis, Fig. 3(b) shows the broadband MWIR raw PSF intensity profiles along with their corresponding RI metrics. The PSF peak intensity and RI remain relatively uniform across the 0°–10° FOV range, with RI values consistently exceeding 0.881. This indicates stable optical performance and high transmittance efficiency in the MWIR imaging system.

Following the PSF and RI evaluation, we assessed the overall imaging quality by constructing a spatially continuous FOV PSF array. Six sub-PSF arrays corresponding to FOV angles of 0°, 2°, 4°, 6°, 8°, and 10° (shown in Fig. S8, Supplement 1) were combined to form a 150 pixel × 150 pixel composite PSF array, as visualized in Fig. 3(c). To simulate realistic imaging conditions, an ideal high-resolution test pattern was segmented into annular regions corresponding to different FOV angles [Fig. 3(d)]. Each segmented region was convolved with its respective PSF array, producing localized imaging results at various FOVs, as shown in Fig. S9, Supplement 1. The resulting images for each FOV were subsequently stitched together to reconstruct the final MWIR imaging result, as shown in Fig. 3(e). This process verifies the MWIR imaging performance of the proposed hybrid metalens system under spatially continuous FOV conditions.

For the LWIR channel, Fig. 4(a) and Fig. S10, Supplement 1, present the normalized and raw PSF intensity distributions on a 40 pixel × 40 pixel grid, computed at FOV angles of 0°, 6°, 12°, 18°, 24°, and 30°. Figure 4(b) provides a comparative analysis of the broadband LWIR raw PSF intensity profiles along with their corresponding RI metrics. The PSF peak intensity and RI values remain relatively uniform within the 0°–15° FOV range. However, beyond 15°, off-axis aberrations become prominent, leading to a 27.3% decline in PSF peak intensity and a reduction of RI to 0.537 at 30° FOV [Fig. 4(b)].

To evaluate full-field imaging performance, a spatially continuous FOV PSF array was constructed by combining six sub-PSF arrays corresponding to the angles of 0°, 6°, 12°, 18°, 24°, and 30°, as shown in Fig. S10, Supplement 1. The resulting composite PSF array, comprising 640 pixel × 480 pixel, is visualized in Fig. 4(c). To simulate realistic imaging conditions, an ideal high-resolution test pattern was segmented into annular regions corresponding to different FOV angles (Fig. S11, Supplement 1). Each segmented region was individually convolved with its respective PSF array to generate imaging results for the corresponding FOVs. These results were then stitched together to reconstruct the final LWIR imaging result, as shown in Fig. 4(e), confirming the LWIR imaging performance of the proposed hybrid metalens system.

Overall, the simulation results validate the imaging capabilities of the hybrid system in both the MWIR and LWIR spectral bands, demonstrating its effectiveness for dual-band infrared imaging applications. Considering potential deviations in practical assembly processes, we systematically analyzed the effects of radial (x-axis) and axial (z-axis) misalignments between the metasurface and the refractive-diffractive element, as shown in Figs. S12 and S13, Supplement 1, respectively. The results indicate that radial and axial misalignments within 0.2 and 0.5 mm, respectively, have a negligible impact on overall system performance across both MWIR and LWIR channels. This confirms that the proposed hybrid refractive-diffractive-metasurface system maintains robust imaging quality under realistic assembly tolerances.

Although the MWIR-LWIR dual-band hybrid metalens demonstrates concurrent imaging across discrete MWIR (3–4 µm) and LWIR (8–12 µm) spectral channels, design-induced optical aberrations lead to blurring as shown in Figs. 3(e) and 4(e). To bridge this performance gap, we employ a StableSR model^[36] integrated with LoRA fine-tuning for infrared-specific image enhancement. The infrared dataset constructed in this study contains approximately 600 MWIR images and 600 LWIR images, totaling around 1200 samples. Experimental results demonstrate that the original model achieves an average peak signal-to-noise ratio (PSNR) of only 24.86 dB, a structural similarity index measure (SSIM) of 0.6281, and a learned perceptual image patch similarity (LPIPS) of 0.3515 on this dataset, revealing its limited detail restoration and insufficient noise suppression in infrared scenarios.

Considering the hardware constraints and dataset scale, we adopted the LoRA fine-tuning strategy. The dataset was split into training and testing sets in an 8:2 ratio, and targeted fine-tuning was performed on the StableSR model. After fine-tuning, the model’s average PSNR on the infrared dataset increased significantly to 26.91 dB, with SSIM rising to approximately 0.7256 and LPIPS dropping to around 0.3787. These results validate the effectiveness of the LoRA-based fine-tuning in adapting to the specific distribution characteristics of infrared images. To further analyze performance improvements, representative MWIR and LWIR samples [Figs. 3(e) and 4(e)] were evaluated separately. Before fine-tuning, the MWIR images achieved PSNR, SSIM, and LPIPS values of 25.83 dB, 0.6943, and 0.3573, while the LWIR images achieved 24.17 dB, 0.5806, and 0.3473, respectively. After LoRA fine-tuning, the MWIR images improved to 27.27 dB (PSNR), 0.7769 (SSIM), and 0.3814 (LPIPS), and the LWIR images improved to 26.66 dB, 0.6888, and 0.3768, respectively. The significantly enhanced imaging quality, as illustrated in Figs. 5(b) and 5(c), demonstrates the strong adaptability and effectiveness of the proposed LoRA fine-tuning strategy for infrared dual-band imaging enhancement. Additional imaging results and image enhancement demonstrations for both the MWIR and LWIR channels under various real-world infrared scenarios are presented in Figs. S14 and S15, Supplement 1, respectively.

Compared with recent works on MWIR-LWIR dual-band imaging systems^[6–15], the system weight is significantly reduced by decreasing the number of refractive elements, while the superiority of dual-band achromatic performance is well maintained. A comparison of the number of optical elements and imaging performance is provided in Table S1, Supplement 1.

In conclusion, we have demonstrated a spatially decoupled MWIR-LWIR dual-band imaging system based on hybrid refractive-diffractive-metasurface optics. A chalcogenide RL with an embedded DOE supports broadband LWIR imaging, while a silicon-based metalens governs MWIR focusing with enhanced dispersion control. The MWIR achromatic metalens achieves broadband aberration correction by incorporating wavelength-dependent phase offsets C(λ) to precisely match the group delay of meta-atoms, enabling achromatic imaging across the 3–4 µm range. The system achieves a compact total track length (TTL) of 11.31 mm, delivering robust performance characterized by a 30° FOV in the LWIR channel and precise dispersion engineering in the MWIR channel. Although this study is primarily simulation-based, all optical designs and analyses were conducted under practical fabrication constraints, including material properties, alignment tolerances, and fabrication feasibility. Furthermore, the image quality is enhanced through a physics-informed StableSR framework with LoRA fine-tuning, further improving the dual-band infrared imaging performance. These innovations present a significant step forward in developing lightweight high-performance dual-band infrared systems, opening new possibilities for both military and civilian applications.