ArticleList

All Subjects Newest Articles

Study on Spectral Unmixing Methods for Skin Fluorophores Based on Simulated System

Xinyue Ai, Guoqing Zhu, Yuanzhi Zhang, Meili Dong, Jingshu Ni, Yao Huang, Ruifang Yang, Yikun Wang, and Yang Zhang

ObjectiveThe non-invasive quantification of advanced glycation end products (AGEs) in skin tissue holds significant promise for the early screening and monitoring of diabetes and other age-related chronic diseases. However, this endeavor is fundamentally challenged by the severe spectral overlap among AGEs and other dominant endogenous fluorophores, including reduced nicotinamide adenine dinucleotide (NADH), flavin adenine dinucleotide (FAD), and collagen (COL). Conventional detection strategies, such as single-wavelength excitation with full-spectrum integration, are highly susceptible to cross-interference from these coexisting components, leading to substantial inaccuracies. To overcome this critical bottleneck, this study aims to establish a robust methodological framework for high-specificity quantification of AGEs by systematically evaluating state-of-the-art spectral unmixing algorithms and developing a principled strategy for selecting an optimal set of discrete excitation wavelengths within a well-controlled simulated system.MethodsA comprehensive dataset is constructed, comprising both simulated and experimentally measured excitation-emission matrix (EEM) spectra of the four-component mixture (AGEs, NADH, FAD, and COL). The simulated data is generated under four distinct noise levels (0.00, 0.05, 0.10, 0.20) to mimic varying signal-to-noise ratios, while the experimental data is acquired from prepared aqueous solutions. Three prominent unmixing algorithms—non-negative matrix factorization (NMF), parallel factor analysis (PARAFAC), and Richardson-Lucy spectral unmixing (RLSU)—are rigorously benchmarked against the ground-truth concentrations. Their performance is evaluated using the coefficient of determination (R2) for AGEs quantification. Subsequently, a novel PARAFAC-driven wavelength selection framework is proposed. This framework leveraged multi-scenario ablation experiments across all five datasets (one experimental and four noisy simulations) to compute the average marginal contribution (ΔR2) of each candidate excitation wavelength to the predictive accuracy of the AGEs model.Results and DiscussionsThe algorithmic benchmark clearly identified PARAFAC as the superior performer, consistently achieving the highest R2 values and demonstrating exceptional robustness against increasing noise levels. This advantage is attributed to PARAFAC's inherent trilinear constraint, which enables the unique decomposition of the mixed signal into its constituent pure profiles, effectively isolating the target AGEs component from interfering signals and random noise. The wavelength selection framework yielded an optimal six-wavelength combination: {345, 370, 385, 420, 435, 455} nm. When applied to the experimental solution samples, this combination enabled PARAFAC to achieve high-precision AGEs quantification [R2=0.9756, root mean square error of prediction (RMSEP) of 3.2808 μg/mL]. In a direct comparative analysis, the conventional single-wavelength (370 nm) integration method exhibits a very weak correlation with true concentrations (R2=0.2385), illustrating its inadequacy in complex mixtures. The physical interpretability of the selected wavelengths is strong: 345 nm lies within the characteristic excitation peak of AGEs, offering high specificity; the long-wavelength band (420?435 nm) effectively targets the FAD excitation peak for its accurate resolution; and 370/385 nm provides complementary information that enhances the overall model stability.ConclusionsThis work successfully demonstrates that the PARAFAC algorithm, coupled with a rationally selected set of discrete excitation wavelengths, provides a powerful and robust solution for the spectral unmixing of the four key skin autofluorescent biomarkers in a simulated environment. The identified wavelength combination {345, 370, 385, 420, 435, 455} nm offers a concrete blueprint for the future development of compact, low-cost, and high-specificity optical devices for non-invasive skin AGEs detection. It is crucial to note that this study is based on an idealized "solution model" that does not account for the light scattering and absorption effects present in real biological tissue. Future research will focus on validating this methodological framework in more realistic scattering phantoms and ultimately in in vivo human trials to pave the way for clinical translation.Objective

The non-invasive quantification of advanced glycation end products (AGEs) in skin tissue holds significant promise for the early screening and monitoring of diabetes and other age-related chronic diseases. However, this endeavor is fundamentally challenged by the severe spectral overlap among AGEs and other dominant endogenous fluorophores, including reduced nicotinamide adenine dinucleotide (NADH), flavin adenine dinucleotide (FAD), and collagen (COL). Conventional detection strategies, such as single-wavelength excitation with full-spectrum integration, are highly susceptible to cross-interference from these coexisting components, leading to substantial inaccuracies. To overcome this critical bottleneck, this study aims to establish a robust methodological framework for high-specificity quantification of AGEs by systematically evaluating state-of-the-art spectral unmixing algorithms and developing a principled strategy for selecting an optimal set of discrete excitation wavelengths within a well-controlled simulated system.

Methods

A comprehensive dataset is constructed, comprising both simulated and experimentally measured excitation-emission matrix (EEM) spectra of the four-component mixture (AGEs, NADH, FAD, and COL). The simulated data is generated under four distinct noise levels (0.00, 0.05, 0.10, 0.20) to mimic varying signal-to-noise ratios, while the experimental data is acquired from prepared aqueous solutions. Three prominent unmixing algorithms—non-negative matrix factorization (NMF), parallel factor analysis (PARAFAC), and Richardson-Lucy spectral unmixing (RLSU)—are rigorously benchmarked against the ground-truth concentrations. Their performance is evaluated using the coefficient of determination ( $R^{2}$ ) for AGEs quantification. Subsequently, a novel PARAFAC-driven wavelength selection framework is proposed. This framework leveraged multi-scenario ablation experiments across all five datasets (one experimental and four noisy simulations) to compute the average marginal contribution ( $Δ R^{2}$ ) of each candidate excitation wavelength to the predictive accuracy of the AGEs model.

Results and Discussions

The algorithmic benchmark clearly identified PARAFAC as the superior performer, consistently achieving the highest $R^{2}$ values and demonstrating exceptional robustness against increasing noise levels. This advantage is attributed to PARAFAC's inherent trilinear constraint, which enables the unique decomposition of the mixed signal into its constituent pure profiles, effectively isolating the target AGEs component from interfering signals and random noise. The wavelength selection framework yielded an optimal six-wavelength combination: {345, 370, 385, 420, 435, 455} nm. When applied to the experimental solution samples, this combination enabled PARAFAC to achieve high-precision AGEs quantification [ $R^{2} = 0.9756$ , root mean square error of prediction (RMSEP) of 3.2808 μg/mL]. In a direct comparative analysis, the conventional single-wavelength (370 nm) integration method exhibits a very weak correlation with true concentrations ( $R^{2} = 0.2385$ ), illustrating its inadequacy in complex mixtures. The physical interpretability of the selected wavelengths is strong: 345 nm lies within the characteristic excitation peak of AGEs, offering high specificity; the long-wavelength band (420?435 nm) effectively targets the FAD excitation peak for its accurate resolution; and 370/385 nm provides complementary information that enhances the overall model stability.

Conclusions

This work successfully demonstrates that the PARAFAC algorithm, coupled with a rationally selected set of discrete excitation wavelengths, provides a powerful and robust solution for the spectral unmixing of the four key skin autofluorescent biomarkers in a simulated environment. The identified wavelength combination {345, 370, 385, 420, 435, 455} nm offers a concrete blueprint for the future development of compact, low-cost, and high-specificity optical devices for non-invasive skin AGEs detection. It is crucial to note that this study is based on an idealized "solution model" that does not account for the light scattering and absorption effects present in real biological tissue. Future research will focus on validating this methodological framework in more realistic scattering phantoms and ultimately in in vivo human trials to pave the way for clinical translation.

Sep. 11, 2026
Laser & Optoelectronics Progress
Vol. 63, Issue 17, 1730006 (2026)
DOI：10.3788/LOP260543

Neural Network-Based Channel Estimation Method for Underwater Optical Communication

Wenzheng Ni, Hexi Liang, Xinchen Ding, Hang Li, and Yong Ai

ObjectiveUnderwater wireless optical communication (UWOC) is recognized as a promising technology for high-speed underwater data transmission due to its high bandwidth, low latency, and immunity to electromagnetic interference. However, the UWOC channel is inherently complex, especially in multipath environments, which induces severe frequency-selective fading, leading to inter-symbol interference (ISI) and inter-carrier interference (ICI). These impairments degrade the accuracy of channel estimation, which is crucial for reliable signal recovery at the receiver. Traditional estimation methods, such as least squares (LS) and linear minimum mean square error (LMMSE), exhibit limitations in strong multipath conditions: LS is highly sensitive to noise, while LMMSE relies heavily on accurate channel statistical information, which is difficult to obtain in dynamic underwater environments. Although deep learning approaches, such as deep neural networks (DNN) and gated recurrent units (GRU), are introduced to improve estimation, they often fail to fully exploit the frequency-domain correlations among subcarriers or adaptively focus on key features. Therefore, there is a pressing need for a robust and efficient channel estimation method that can maintain high accuracy under limited pilot overhead and without the reliance on a cyclic prefix (CP), thereby enhancing the spectral efficiency of UWOC systems.MethodsThis paper proposes a novel neural network-based channel estimator named BSD (BiLSTM-Self-Attention-DNN Network) for DC-biased optical orthogonal frequency division multiplexing (DCO-OFDM) systems in UWOC. The BSD network is designed to perform end-to-end mapping from received signals to transmitted symbols, bypassing explicit channel estimation and equalization. The network architecture comprises three main components: a bidirectional long short-term memory (BiLSTM) layer, a self-attention (SA) layer, and a dense neural network (DNN) layer. The BiLSTM layer processes the subcarrier sequence in both forward and backward directions, capturing long-range dependencies and frequency-domain smoothness characteristic of multipath channels. The SA layer then computes attention weights across subcarriers, adaptively emphasizing those with critical information (e.g., near pilot locations or less affected by fading) while suppressing noisy ones. Finally, the DNN layer performs nonlinear transformation and dimensionality reduction to recover the transmitted bit stream. To train the BSD network, a comprehensive dataset is generated using Monte Carlo simulations that emulate realistic underwater optical channels. The simulation models photon propagation in two typical water types: clear ocean and coastal ocean, incorporating effects such as absorption, scattering, and boundary reflections. Key parameters include a wavelength of 470 nm, water depth of 15 m, and anisotropic factor of 0.8. The DCO-OFDM system is configured with 192 subcarriers, 16-QAM modulation, and variable pilot numbers (8 to 64). The network is trained offline using the mean squared error (MSE) loss function and the Adam optimizer under a signal-to-noise ratio (SNR) of 25 dB.Results and DiscussionsThe performance of the BSD estimator is evaluated through extensive simulations and compared with traditional methods (LS, LMMSE) and neural network benchmarks (DNN, GRU). In both clear and coastal ocean channels with CP inserted and 64 pilots, BSD achieves the lowest bit error rate (BER) across various SNR levels. For instance, at 30 dB SNR in clear ocean, BSD attains a BER of 2.25×10-5, outperforming DNN (9.38×10-5), GRU (1.04×10-4), and LMMSE (1.53×10-4). Similar superior performance is observed in coastal ocean. Ablation studies confirm the contributions of each BSD module: the BiLSTM-DNN combination significantly improves upon standalone BiLSTM, and the inclusion of the SA mechanism further enhances estimation accuracy. The robustness of BSD is notably demonstrated under challenging conditions. When the pilot count is reduced to 8 (with CP), BSD maintains a BER of 1.19×10-4 at 30 dB SNR, while LMMSE deteriorates sharply to 1.04×10-1. Moreover, in the absence of CP—which exacerbates ISI and ICI—BSD exhibits remarkable resilience. Without CP and with 64 pilots, BSD achieves a BER of 2.17×10-4 at 30 dB SNR, whereas DNN and GRU degrade to 1.08×10-3 and 9.52×10-4, respectively. Even under the stringent scenario of only 8 pilots and no CP, BSD sustains a BER of 5.76×10-4 at 30 dB SNR, highlighting its exceptional robustness. Complexity analysis measured in floating-point operations (FLOPs) shows that BSD requires approximately 2.23×106 FLOPs per OFDM symbol, which is lower than LMMSE (4.80×106) and acceptable for practical deployment given its performance advantages.ConclusionsThis paper presents a BSD neural network for channel estimation in underwater DCO-OFDM systems. By integrating BiLSTM for sequence dependency modeling, self-attention for adaptive feature weighting, and DNN for nonlinear regression, the proposed method effectively mitigates multipath-induced distortions. Simulation results validate that BSD outperforms existing techniques in terms of BER performance across various water types, pilot counts, and CP conditions. Crucially, BSD demonstrates strong robustness in scenarios with limited pilots and without CP, thereby reducing spectral overhead and enhancing system spectral efficiency. The use of Monte Carlo-simulated channel data ensures that the model is trained on realistic underwater conditions, improving its practical applicability. The BSD estimator offers a promising deep learning-based solution for high-performance and resource-efficient UWOC systems.Objective

Underwater wireless optical communication (UWOC) is recognized as a promising technology for high-speed underwater data transmission due to its high bandwidth, low latency, and immunity to electromagnetic interference. However, the UWOC channel is inherently complex, especially in multipath environments, which induces severe frequency-selective fading, leading to inter-symbol interference (ISI) and inter-carrier interference (ICI). These impairments degrade the accuracy of channel estimation, which is crucial for reliable signal recovery at the receiver. Traditional estimation methods, such as least squares (LS) and linear minimum mean square error (LMMSE), exhibit limitations in strong multipath conditions: LS is highly sensitive to noise, while LMMSE relies heavily on accurate channel statistical information, which is difficult to obtain in dynamic underwater environments. Although deep learning approaches, such as deep neural networks (DNN) and gated recurrent units (GRU), are introduced to improve estimation, they often fail to fully exploit the frequency-domain correlations among subcarriers or adaptively focus on key features. Therefore, there is a pressing need for a robust and efficient channel estimation method that can maintain high accuracy under limited pilot overhead and without the reliance on a cyclic prefix (CP), thereby enhancing the spectral efficiency of UWOC systems.

Methods

This paper proposes a novel neural network-based channel estimator named BSD (BiLSTM-Self-Attention-DNN Network) for DC-biased optical orthogonal frequency division multiplexing (DCO-OFDM) systems in UWOC. The BSD network is designed to perform end-to-end mapping from received signals to transmitted symbols, bypassing explicit channel estimation and equalization. The network architecture comprises three main components: a bidirectional long short-term memory (BiLSTM) layer, a self-attention (SA) layer, and a dense neural network (DNN) layer. The BiLSTM layer processes the subcarrier sequence in both forward and backward directions, capturing long-range dependencies and frequency-domain smoothness characteristic of multipath channels. The SA layer then computes attention weights across subcarriers, adaptively emphasizing those with critical information (e.g., near pilot locations or less affected by fading) while suppressing noisy ones. Finally, the DNN layer performs nonlinear transformation and dimensionality reduction to recover the transmitted bit stream. To train the BSD network, a comprehensive dataset is generated using Monte Carlo simulations that emulate realistic underwater optical channels. The simulation models photon propagation in two typical water types: clear ocean and coastal ocean, incorporating effects such as absorption, scattering, and boundary reflections. Key parameters include a wavelength of 470 nm, water depth of 15 m, and anisotropic factor of 0.8. The DCO-OFDM system is configured with 192 subcarriers, 16-QAM modulation, and variable pilot numbers (8 to 64). The network is trained offline using the mean squared error (MSE) loss function and the Adam optimizer under a signal-to-noise ratio (SNR) of 25 dB.

Results and Discussions

The performance of the BSD estimator is evaluated through extensive simulations and compared with traditional methods (LS, LMMSE) and neural network benchmarks (DNN, GRU). In both clear and coastal ocean channels with CP inserted and 64 pilots, BSD achieves the lowest bit error rate (BER) across various SNR levels. For instance, at 30 dB SNR in clear ocean, BSD attains a BER of $2.25 \times 10^{- 5}$ , outperforming DNN ( $9.38 \times 10^{- 5}$ ), GRU ( $1.04 \times 10^{- 4}$ ), and LMMSE ( $1.53 \times 10^{- 4}$ ). Similar superior performance is observed in coastal ocean. Ablation studies confirm the contributions of each BSD module: the BiLSTM-DNN combination significantly improves upon standalone BiLSTM, and the inclusion of the SA mechanism further enhances estimation accuracy. The robustness of BSD is notably demonstrated under challenging conditions. When the pilot count is reduced to 8 (with CP), BSD maintains a BER of $1.19 \times 10^{- 4}$ at 30 dB SNR, while LMMSE deteriorates sharply to $1.04 \times 10^{- 1}$ . Moreover, in the absence of CP—which exacerbates ISI and ICI—BSD exhibits remarkable resilience. Without CP and with 64 pilots, BSD achieves a BER of $2.17 \times 10^{- 4}$ at 30 dB SNR, whereas DNN and GRU degrade to $1.08 \times 10^{- 3}$ and $9.52 \times 10^{- 4}$ , respectively. Even under the stringent scenario of only 8 pilots and no CP, BSD sustains a BER of $5.76 \times 10^{- 4}$ at 30 dB SNR, highlighting its exceptional robustness. Complexity analysis measured in floating-point operations (FLOPs) shows that BSD requires approximately $2.23 \times 10^{6}$ FLOPs per OFDM symbol, which is lower than LMMSE ( $4.80 \times 10^{6}$ ) and acceptable for practical deployment given its performance advantages.

Conclusions

This paper presents a BSD neural network for channel estimation in underwater DCO-OFDM systems. By integrating BiLSTM for sequence dependency modeling, self-attention for adaptive feature weighting, and DNN for nonlinear regression, the proposed method effectively mitigates multipath-induced distortions. Simulation results validate that BSD outperforms existing techniques in terms of BER performance across various water types, pilot counts, and CP conditions. Crucially, BSD demonstrates strong robustness in scenarios with limited pilots and without CP, thereby reducing spectral overhead and enhancing system spectral efficiency. The use of Monte Carlo-simulated channel data ensures that the model is trained on realistic underwater conditions, improving its practical applicability. The BSD estimator offers a promising deep learning-based solution for high-performance and resource-efficient UWOC systems.

Sep. 11, 2026
Laser & Optoelectronics Progress
Vol. 63, Issue 17, 1706005 (2026)
DOI：10.3788/LOP252657

Study on the Characteristics of Femtosecond Laser-Triggered Discharge Under Low Pressure

Dengfeng Huang, Zheng Lu, Junyan Zhang, Xinguo Sha, Qiang Gao, and Bo Li

ObjectiveLow-pressure gas discharges exhibit broad application prospects in plasma display panels, medical sterilization, surface treatment of materials, and protection of vacuum electronic devices. However, their discharge processes are often characterized by strong randomness of discharge paths and poor triggering stability, which severely limit precise control and practical engineering applications under low-pressure conditions. Femtosecond laser pulses propagating in gases can generate plasma channels with finite electron density, providing a pre-ionized pathway for electrical breakdown and enabling spatial confinement and temporal triggering of the discharge process. Nevertheless, existing studies on femtosecond laser-triggered discharges have mainly focused on atmospheric-pressure environments, while systematic investigations into the guiding capability, spatiotemporal stability, and plasma lifetime under low-pressure conditions remain insufficient. Motivated by this gap, the present study systematically investigates femtosecond laser-triggered discharges in air over a wide pressure range from 1 to 100 kPa. Particular emphasis is placed on elucidating the effects of ambient pressure on discharge breakdown characteristics, spatial stability, and temporal stability. The objective is to reveal the coupling mechanisms between femtosecond laser-produced plasma and external electric fields under low-pressure conditions, thereby providing experimental insights for precise control of low-pressure discharges and the extension of their related applications.MethodsA Ti∶sapphire femtosecond laser system is employed in the experiments, delivering pulses with a central wavelength of 800 nm, a pulse duration of 45 fs, a single-pulse energy of 3.5 mJ, and a repetition rate of 10 Hz. The laser beam is focused into a low-pressure chamber using a lens with a focal length of 300 mm to generate a filamentary plasma channel. A pair of tungsten electrodes with an interelectrode gap of 5 mm is installed inside the chamber and connected to a nanosecond-scale pulsed direct current (DC) high-voltage power supply. The ambient pressure inside the chamber is continuously adjustable in the range of 1 to 100 kPa. A digital delay generator is used to precisely control the time delay Δt between the femtosecond laser pulse and the high-voltage pulse. During the discharge process, voltage, current, and laser signals are synchronously recorded to characterize temporal parameters such as discharge delay time and timing jitter. In addition, a digital single-lens reflex camera is employed to image the discharge channels, enabling analysis of the spatial stability and morphological features of the discharge paths under different pressure and delay conditions.Results and DiscussionsIn terms of spatial characteristics, free discharges exhibit noticeable path bending and nonuniform plasma distributions under both atmospheric and low-to-medium pressure conditions (middle column). In contrast, femtosecond laser-triggered discharges propagate strictly along the laser-produced plasma channel. At a pressure of approximately 10 kPa, a narrow discharge channel with a diameter of about 0.4 mm is formed, featuring a more concentrated and uniform radial emission profile (right column). Compared with the case at 50 kPa, the discharge channel diameter at 10 kPa is reduced to only 44%, indicating a significantly improved spatial resolution. When the pressure is further reduced to 1 kPa, both free discharges and laser-guided discharges degenerate into corona discharges due to insufficient electron density and collision frequency, preventing the formation of a stable breakdown channel. For pressures above 1 kPa, stable laser-guided discharges can be achieved, and the overall variation of the breakdown voltage threshold with pressure follows Paschen's law. In the pressure range of 2.5?10 kPa, the breakdown threshold varies slowly, whereas it increases approximately linearly with pressure at higher pressures. Compared with free discharges, the breakdown voltage threshold of femtosecond laser-triggered discharges is significantly reduced to 50.0%?66.7% of that of free discharges, with a more pronounced advantage observed on the rising branch of the Paschen curve. Regarding temporal stability, the low-pressure environment substantially improves the triggering characteristics of the discharge. At 10 kPa, the discharge delay time is significantly shortened, and the timing jitter remains below 5 ns over a wide range of applied voltages and delay times, reaching sub-nanosecond levels under certain conditions. Compared with atmospheric pressure, the effective triggering time window for femtosecond laser-triggered discharges at low pressure is greatly extended, reaching up to -400 μs under negative delay conditions, which is approximately 12 times wider than that at atmospheric pressure. Notably, for longer negative delay times ranging from -500 μs to -90 ms, although a complete discharge channel is not formed, fluorescence emission induced by residual plasma under the external electric field can still be observed. This phenomenon indicates that femtosecond laser pulses can generate low-density plasma remnants with lifetimes on the order of milliseconds in low-pressure gases. These plasma remnants can be re-excited by an external electric field, producing sustained fluorescence emission and exhibiting a visual effect analogous to "filament re-ignition".ConclusionsThis study systematically investigates the breakdown characteristics and spatiotemporal stability of femtosecond laser-triggered discharges under low-pressure conditions. The results demonstrate that, within an appropriate low-pressure range, femtosecond laser guidance can significantly reduce the breakdown voltage threshold, generate discharge channels with enhanced spatial resolution and superior temporal stability, and substantially broaden the effective triggering time window. Moreover, millisecond-lifetime residual plasma and its electric-field-induced re-excitation fluorescence are observed for the first time under low-pressure conditions. These findings provide experimental insights into the coupling mechanisms between laser-produced plasma and external electric fields at low pressure and offer a new technical pathway for precise control of low-pressure discharges and their applications in other fields.Objective

Low-pressure gas discharges exhibit broad application prospects in plasma display panels, medical sterilization, surface treatment of materials, and protection of vacuum electronic devices. However, their discharge processes are often characterized by strong randomness of discharge paths and poor triggering stability, which severely limit precise control and practical engineering applications under low-pressure conditions. Femtosecond laser pulses propagating in gases can generate plasma channels with finite electron density, providing a pre-ionized pathway for electrical breakdown and enabling spatial confinement and temporal triggering of the discharge process. Nevertheless, existing studies on femtosecond laser-triggered discharges have mainly focused on atmospheric-pressure environments, while systematic investigations into the guiding capability, spatiotemporal stability, and plasma lifetime under low-pressure conditions remain insufficient. Motivated by this gap, the present study systematically investigates femtosecond laser-triggered discharges in air over a wide pressure range from 1 to 100 kPa. Particular emphasis is placed on elucidating the effects of ambient pressure on discharge breakdown characteristics, spatial stability, and temporal stability. The objective is to reveal the coupling mechanisms between femtosecond laser-produced plasma and external electric fields under low-pressure conditions, thereby providing experimental insights for precise control of low-pressure discharges and the extension of their related applications.

Methods

A Ti∶sapphire femtosecond laser system is employed in the experiments, delivering pulses with a central wavelength of 800 nm, a pulse duration of 45 fs, a single-pulse energy of 3.5 mJ, and a repetition rate of 10 Hz. The laser beam is focused into a low-pressure chamber using a lens with a focal length of 300 mm to generate a filamentary plasma channel. A pair of tungsten electrodes with an interelectrode gap of 5 mm is installed inside the chamber and connected to a nanosecond-scale pulsed direct current (DC) high-voltage power supply. The ambient pressure inside the chamber is continuously adjustable in the range of 1 to 100 kPa. A digital delay generator is used to precisely control the time delay $Δ t$ between the femtosecond laser pulse and the high-voltage pulse. During the discharge process, voltage, current, and laser signals are synchronously recorded to characterize temporal parameters such as discharge delay time and timing jitter. In addition, a digital single-lens reflex camera is employed to image the discharge channels, enabling analysis of the spatial stability and morphological features of the discharge paths under different pressure and delay conditions.

Results and Discussions

In terms of spatial characteristics, free discharges exhibit noticeable path bending and nonuniform plasma distributions under both atmospheric and low-to-medium pressure conditions (middle column). In contrast, femtosecond laser-triggered discharges propagate strictly along the laser-produced plasma channel. At a pressure of approximately 10 kPa, a narrow discharge channel with a diameter of about 0.4 mm is formed, featuring a more concentrated and uniform radial emission profile (right column). Compared with the case at 50 kPa, the discharge channel diameter at 10 kPa is reduced to only 44%, indicating a significantly improved spatial resolution. When the pressure is further reduced to 1 kPa, both free discharges and laser-guided discharges degenerate into corona discharges due to insufficient electron density and collision frequency, preventing the formation of a stable breakdown channel. For pressures above 1 kPa, stable laser-guided discharges can be achieved, and the overall variation of the breakdown voltage threshold with pressure follows Paschen's law. In the pressure range of 2.5?10 kPa, the breakdown threshold varies slowly, whereas it increases approximately linearly with pressure at higher pressures. Compared with free discharges, the breakdown voltage threshold of femtosecond laser-triggered discharges is significantly reduced to 50.0%?66.7% of that of free discharges, with a more pronounced advantage observed on the rising branch of the Paschen curve. Regarding temporal stability, the low-pressure environment substantially improves the triggering characteristics of the discharge. At 10 kPa, the discharge delay time is significantly shortened, and the timing jitter remains below 5 ns over a wide range of applied voltages and delay times, reaching sub-nanosecond levels under certain conditions. Compared with atmospheric pressure, the effective triggering time window for femtosecond laser-triggered discharges at low pressure is greatly extended, reaching up to -400 μs under negative delay conditions, which is approximately 12 times wider than that at atmospheric pressure. Notably, for longer negative delay times ranging from -500 μs to -90 ms, although a complete discharge channel is not formed, fluorescence emission induced by residual plasma under the external electric field can still be observed. This phenomenon indicates that femtosecond laser pulses can generate low-density plasma remnants with lifetimes on the order of milliseconds in low-pressure gases. These plasma remnants can be re-excited by an external electric field, producing sustained fluorescence emission and exhibiting a visual effect analogous to "filament re-ignition".

Conclusions

This study systematically investigates the breakdown characteristics and spatiotemporal stability of femtosecond laser-triggered discharges under low-pressure conditions. The results demonstrate that, within an appropriate low-pressure range, femtosecond laser guidance can significantly reduce the breakdown voltage threshold, generate discharge channels with enhanced spatial resolution and superior temporal stability, and substantially broaden the effective triggering time window. Moreover, millisecond-lifetime residual plasma and its electric-field-induced re-excitation fluorescence are observed for the first time under low-pressure conditions. These findings provide experimental insights into the coupling mechanisms between laser-produced plasma and external electric fields at low pressure and offer a new technical pathway for precise control of low-pressure discharges and their applications in other fields.

Sep. 11, 2026
Laser & Optoelectronics Progress
Vol. 63, Issue 17, 1714004 (2026)
DOI：10.3788/LOP252569

Research Progress on Defect Regulation in Energy Field-Assisted Laser Cladding

Rongrong Song, Linsen Shu, Jiasheng Wu, Siyu Mao, and Zhenying Liu

SignificanceLaser cladding technology, as a core means of surface strengthening and remanufacturing of key components in high-end equipment, has become essential for surface strengthening, damage repair, and functional gradient material preparation in high-end manufacturing industries such as aerospace, energy equipment, and medical devices. However, its transient thermal melting characteristics of"rapid heating and rapid cooling"can easily induce defects such as cracks, pores, and uneven microstructure in the cladding layer, seriously restricting the reliability and service life of the formed parts. Traditional process parameter optimization and powder characteristic adjustment encounter a significant"ceiling"effect, making it difficult to achieve deep defect suppression and performance improvement. In this context, external physical fields such as magnetic field, ultrasonic field, and thermal field are introduced to construct a closed-loop system of"energy field-assisted performance control in laser cladding". By actively intervening in the heat and mass transfer and solidification behavior of the melt pool, this approach provides an innovative path to break through technological bottlenecks and prepare high-quality cladding layers, holding great significance for promoting the engineering application of laser cladding technology and upgrading high-end equipment manufacturing.ProgressThe control mechanism of energy fields such as magnetic field, ultrasonic field, and thermal field on defect suppression (such as cracks, pores, and compositional segregation) and microstructure optimization (such as grain refinement) during laser cladding can be explained by"regulating the convective behavior of the melt pool"and"optimizing the solidification process". The magnetic field exerts a synergistic effect of"forced convection"through the thermoelectric magnetic effect, achieving grain refinement by disrupting columnar crystal growth through thermoelectric magnetic force and enhancing solute exchange in the melt pool through the thermoelectric electromagnetic effect. At the same time, the electromagnetic braking effect can effectively suppress splashing and natural convection of the melt pool, stabilize melt pool morphology and reduce defect formation. The ultrasonic field utilizes the cavitation effect and acoustic streaming effect to form a"macroscopic homogenization and microscopic nucleation strengthening" mechanism, breaking the growth mode of coarse columnar crystals under rapid solidification, promoting the transformation of columnar crystals into equiaxed crystals, achieving grain refinement and microstructure homogenization, and ultimately improving the comprehensive performance of the cladding layer such as hardness, wear resistance, and corrosion resistance. The thermal field focuses on temperature regulation, achieving the goals of regulating the temperature field of the melt pool, mitigating thermal stress accumulation, suppressing cracks and pore defects, and optimizing grain growth conditions through preheating or synchronous heating. However, a single energy field also has limitations such as a limited range of action and single performance optimization. The combination of multiple energy fields compensates for the limitations of a single energy field through the"thermal-mechanical" coupling effect, achieving coordinated defect suppression. In the ultrasonic-electromagnetic composite field, the Lorentz force expands and enhances the cavitation effect throughout the entire melt pool, maximizing undercooling of the melt pool and promoting the uniform distribution of the strengthening phase WC to avoid crack initiation caused by particle concentration. Ultimately, a crack-free nickel-based cladding layer with a WC content of 60% was successfully prepared.Conclusions and ProspectsThe core of single energy field-assisted defect suppression and microstructure optimization lies in the reconstruction of solidification conditions, which creates conditions for grain refinement by actively intervening in the convective state of the melt pool and altering the ratio of temperature gradient to solidification rate. Multi-energy field combinations integrate the advantages of each field through the "thermal-mechanical"coupling effect, achieving synergistic suppression of multidimensional defects (cracks, pores, segregation) and yielding a performance improvement of 1+1>2. However, current research still faces challenges, such as the quantitative characterization of magnetic field intensity uniformity, ultrasonic energy attenuation, and the coupling mechanisms of composite energy fields. Additionally, issues such as poor device integration compatibility, high costs, and lack of standardization limit industrial application. Future research should be based on interdisciplinary integration, with"component design"and"external energy field"as core process variables, establishing reliable models through multiphysics field coupled numerical simulation and in situ observation, developing specialized equipment to break through technical bottlenecks, establishing standardized process systems to promote industrialization, expanding green manufacturing applications, and providing support for the high-quality development of the high-end equipment manufacturing industry.Significance

Laser cladding technology, as a core means of surface strengthening and remanufacturing of key components in high-end equipment, has become essential for surface strengthening, damage repair, and functional gradient material preparation in high-end manufacturing industries such as aerospace, energy equipment, and medical devices. However, its transient thermal melting characteristics of"rapid heating and rapid cooling"can easily induce defects such as cracks, pores, and uneven microstructure in the cladding layer, seriously restricting the reliability and service life of the formed parts. Traditional process parameter optimization and powder characteristic adjustment encounter a significant"ceiling"effect, making it difficult to achieve deep defect suppression and performance improvement. In this context, external physical fields such as magnetic field, ultrasonic field, and thermal field are introduced to construct a closed-loop system of"energy field-assisted performance control in laser cladding". By actively intervening in the heat and mass transfer and solidification behavior of the melt pool, this approach provides an innovative path to break through technological bottlenecks and prepare high-quality cladding layers, holding great significance for promoting the engineering application of laser cladding technology and upgrading high-end equipment manufacturing.

Progress

The control mechanism of energy fields such as magnetic field, ultrasonic field, and thermal field on defect suppression (such as cracks, pores, and compositional segregation) and microstructure optimization (such as grain refinement) during laser cladding can be explained by"regulating the convective behavior of the melt pool"and"optimizing the solidification process". The magnetic field exerts a synergistic effect of"forced convection"through the thermoelectric magnetic effect, achieving grain refinement by disrupting columnar crystal growth through thermoelectric magnetic force and enhancing solute exchange in the melt pool through the thermoelectric electromagnetic effect. At the same time, the electromagnetic braking effect can effectively suppress splashing and natural convection of the melt pool, stabilize melt pool morphology and reduce defect formation. The ultrasonic field utilizes the cavitation effect and acoustic streaming effect to form a"macroscopic homogenization and microscopic nucleation strengthening" mechanism, breaking the growth mode of coarse columnar crystals under rapid solidification, promoting the transformation of columnar crystals into equiaxed crystals, achieving grain refinement and microstructure homogenization, and ultimately improving the comprehensive performance of the cladding layer such as hardness, wear resistance, and corrosion resistance. The thermal field focuses on temperature regulation, achieving the goals of regulating the temperature field of the melt pool, mitigating thermal stress accumulation, suppressing cracks and pore defects, and optimizing grain growth conditions through preheating or synchronous heating. However, a single energy field also has limitations such as a limited range of action and single performance optimization. The combination of multiple energy fields compensates for the limitations of a single energy field through the"thermal-mechanical" coupling effect, achieving coordinated defect suppression. In the ultrasonic-electromagnetic composite field, the Lorentz force expands and enhances the cavitation effect throughout the entire melt pool, maximizing undercooling of the melt pool and promoting the uniform distribution of the strengthening phase $W C$ to avoid crack initiation caused by particle concentration. Ultimately, a crack-free nickel-based cladding layer with a $W C$ content of 60% was successfully prepared.

Conclusions and Prospects

The core of single energy field-assisted defect suppression and microstructure optimization lies in the reconstruction of solidification conditions, which creates conditions for grain refinement by actively intervening in the convective state of the melt pool and altering the ratio of temperature gradient to solidification rate. Multi-energy field combinations integrate the advantages of each field through the "thermal-mechanical"coupling effect, achieving synergistic suppression of multidimensional defects (cracks, pores, segregation) and yielding a performance improvement of $1 + 1 > 2$ . However, current research still faces challenges, such as the quantitative characterization of magnetic field intensity uniformity, ultrasonic energy attenuation, and the coupling mechanisms of composite energy fields. Additionally, issues such as poor device integration compatibility, high costs, and lack of standardization limit industrial application. Future research should be based on interdisciplinary integration, with"component design"and"external energy field"as core process variables, establishing reliable models through multiphysics field coupled numerical simulation and in situ observation, developing specialized equipment to break through technical bottlenecks, establishing standardized process systems to promote industrialization, expanding green manufacturing applications, and providing support for the high-quality development of the high-end equipment manufacturing industry.

Sep. 11, 2026
Laser & Optoelectronics Progress
Vol. 63, Issue 17, 1700005 (2026)
DOI：10.3788/LOP252525

All-fiber spectral-overlapping and broadband single-cavity dual-wavelength comb laser

Xiaochun Liao, Guoqing Hu, Junpeng Liu, Xixuan Huang, Zhe Zhang, Huawei Liu, Peng Chen, Hetian Li, Guangwei Chen, and Zhehai Zhou

Sep. 11, 2026
Infrared and Laser Engineering
Vol. 55, Issue 8, 20260043 (2026)
DOI：10.3788/IRLA20260043

Hybrid mode-locked Thulium-doped fiber laser based on Nb₂CT_x MXene saturable absorber and nonlinear polarization rotation

Yibing Chen, Peng Lin, Silun Du, Yunhan Deng, and Tianshu Wang

ObjectiveThulium-doped (Tm) ultrafast fiber lasers operating around 2 μm are attractive for mid-infrared spectroscopy, sensing, lidar, and medical applications. Nonlinear polarization rotation (NPR) mode locking is widely used due to its fast response and high damage tolerance, but NPR-only operation is highly sensitive to polarization-state perturbations, which often leads to a relatively high mode-locking threshold and limited self-starting repeatability. Therefore, it is necessary to develop a hybrid mode-locking scheme that introduces deterministic intensity discrimination while preserving the rapid nonlinear shaping capability of NPR. For this purpose, a Tm-doped hybrid mode-locked fiber laser based on an Nb2CTx MXene saturable absorber (SA) and NPR is proposed and experimentally demonstrated.MethodsAn Nb2CTx MXene SA is fabricated by depositing Nb2CTx onto the waist region of a tapered micro-/nanofiber (MNF) to realize evanescent-field coupling and stable nonlinear absorption. The nonlinear transmission of the Nb2CTx-SA is characterized using a balanced twin-detector measurement and fitted with a fractional saturable-absorption model, yielding a modulation depth of 10.47% and a saturation power of 28 mW. An all-fiber Tm-doped ring cavity is then constructed by integrating the Nb2CTx-SA with an NPR section formed by polarization controllers and a polarization-dependent isolator. To quantify the contribution of the SA under the same platform, a matched NPR-only reference cavity is built by removing the SA segment while keeping the remaining cavity configuration and test criteria unchanged. The output spectrum, pulse train, RF spectrum, and autocorrelation trace (sech2 fitting) are measured to evaluate spectral/temporal characteristics and RF purity. Output stability is assessed by 2 h continuous monitoring of average power and spectrum, and repeatability is examined through 10 power-cycling tests.Results and DiscussionsIn the hybrid mode-locking regime, stable soliton pulses are obtained at a center wavelength of 1949.8 nm with a 3-dB bandwidth of 4.8 nm. The pulse duration is 1.8 ps from sech2 fitting, the repetition rate is 18.02 MHz, and the fundamental RF signal-to-noise ratio (SNR) reaches ~58 dB (Fig.4). In comparison, the NPR-only reference cavity (Fig.7) delivers 2.31-ps pulseswith an RF SNR of ~47 dB and a mode-locking threshold of ~300 mW. After introducing the Nb2CTx-SA, the hybrid cavity reduces the threshold to ~260 mW, compresses the pulse width by 0.51 ps, and improves the RF SNR by ~11 dB, indicating that the introduction of the Nb2CTx-SA was beneficial for improving the mode-locking condition and RF purity under the present cavity configuration. At higher pump powers, second- and third-order harmonic mode locking are achieved at 36.1 MHz and 54.2 MHz (Fig.5). Stability measurements showed an RMS power fluctuation of 0.21% over 2 h with negligible spectral drift, and repeated self-starting mode locking was realized in 10 power-cycling trials (Fig.6), confirming good short-term operational stability and repeatability.ConclusionsA Tm-doped hybrid mode-locked all-fiber ring laser enabled by an Nb2CTx MXene saturable absorber and nonlinear polarization rotation is demonstrated at around 2 μm. The MNF-based Nb2CTx-SA exhibits a modulation depth of 10.47% and a saturation power of 28 mW, and an additional linear insertion loss of about 0.7 dB relative to the bare MNF device, providing deterministic intensity discrimination for intracavity pulse formation. The hybrid laser delivers stable soliton pulses at 1949.8 nm with a 4.8-nm 3-dB bandwidth, a pulse duration of 1.8 ps, a repetition rate of 18.02 MHz, and an RF SNR of about 58 dB. Relative to the matched NPR-only reference cavity, the mode-locking threshold is reduced from about 300 mW to about 260 mW, the pulse duration is shortened from 2.31 ps to 1.8 ps, and the RF SNR is improved from about 47 dB to about 58 Db. These results indicate that the introduction of the Nb2CTx-SA is favorable for improving the mode-locking condition, pulse quality, and RF purity under the present experimental conditions. In addition, second- and third-order harmonic mode locking at 36.1 MHz and 54.2 MHz was observed at higher pump powers. Stability tests further showed an RMS power fluctuation of 0.21% over 2 h and repeatable self-starting operation over ten power-cycling trials, demonstrating good short-term stability and repeatability. Long-term aging behavior and packaging reliability remain to be investigated in future work.ObjectiveThulium-doped (Tm) ultrafast fiber lasers operating around 2 μm are attractive for mid-infrared spectroscopy, sensing, lidar, and medical applications. Nonlinear polarization rotation (NPR) mode locking is widely used due to its fast response and high damage tolerance, but NPR-only operation is highly sensitive to polarization-state perturbations, which often leads to a relatively high mode-locking threshold and limited self-starting repeatability. Therefore, it is necessary to develop a hybrid mode-locking scheme that introduces deterministic intensity discrimination while preserving the rapid nonlinear shaping capability of NPR. For this purpose, a Tm-doped hybrid mode-locked fiber laser based on an Nb₂CT_x MXene saturable absorber (SA) and NPR is proposed and experimentally demonstrated.MethodsAn Nb₂CT_x MXene SA is fabricated by depositing Nb₂CT_x onto the waist region of a tapered micro-/nanofiber (MNF) to realize evanescent-field coupling and stable nonlinear absorption. The nonlinear transmission of the Nb₂CT_x-SA is characterized using a balanced twin-detector measurement and fitted with a fractional saturable-absorption model, yielding a modulation depth of 10.47% and a saturation power of 28 mW. An all-fiber Tm-doped ring cavity is then constructed by integrating the Nb₂CT_x-SA with an NPR section formed by polarization controllers and a polarization-dependent isolator. To quantify the contribution of the SA under the same platform, a matched NPR-only reference cavity is built by removing the SA segment while keeping the remaining cavity configuration and test criteria unchanged. The output spectrum, pulse train, RF spectrum, and autocorrelation trace (sech² fitting) are measured to evaluate spectral/temporal characteristics and RF purity. Output stability is assessed by 2 h continuous monitoring of average power and spectrum, and repeatability is examined through 10 power-cycling tests.Results and DiscussionsIn the hybrid mode-locking regime, stable soliton pulses are obtained at a center wavelength of 1949.8 nm with a 3-dB bandwidth of 4.8 nm. The pulse duration is 1.8 ps from sech² fitting, the repetition rate is 18.02 MHz, and the fundamental RF signal-to-noise ratio (SNR) reaches ~58 dB (Fig.4). In comparison, the NPR-only reference cavity (Fig.7) delivers 2.31-ps pulseswith an RF SNR of ~47 dB and a mode-locking threshold of ~300 mW. After introducing the Nb₂CT_x-SA, the hybrid cavity reduces the threshold to ~260 mW, compresses the pulse width by 0.51 ps, and improves the RF SNR by ~11 dB, indicating that the introduction of the Nb₂CT_x-SA was beneficial for improving the mode-locking condition and RF purity under the present cavity configuration. At higher pump powers, second- and third-order harmonic mode locking are achieved at 36.1 MHz and 54.2 MHz (Fig.5). Stability measurements showed an RMS power fluctuation of 0.21% over 2 h with negligible spectral drift, and repeated self-starting mode locking was realized in 10 power-cycling trials (Fig.6), confirming good short-term operational stability and repeatability.ConclusionsA Tm-doped hybrid mode-locked all-fiber ring laser enabled by an Nb₂CT_x MXene saturable absorber and nonlinear polarization rotation is demonstrated at around 2 μm. The MNF-based Nb₂CT_x-SA exhibits a modulation depth of 10.47% and a saturation power of 28 mW, and an additional linear insertion loss of about 0.7 dB relative to the bare MNF device, providing deterministic intensity discrimination for intracavity pulse formation. The hybrid laser delivers stable soliton pulses at 1949.8 nm with a 4.8-nm 3-dB bandwidth, a pulse duration of 1.8 ps, a repetition rate of 18.02 MHz, and an RF SNR of about 58 dB. Relative to the matched NPR-only reference cavity, the mode-locking threshold is reduced from about 300 mW to about 260 mW, the pulse duration is shortened from 2.31 ps to 1.8 ps, and the RF SNR is improved from about 47 dB to about 58 Db. These results indicate that the introduction of the Nb₂CT_x-SA is favorable for improving the mode-locking condition, pulse quality, and RF purity under the present experimental conditions. In addition, second- and third-order harmonic mode locking at 36.1 MHz and 54.2 MHz was observed at higher pump powers. Stability tests further showed an RMS power fluctuation of 0.21% over 2 h and repeatable self-starting operation over ten power-cycling trials, demonstrating good short-term stability and repeatability. Long-term aging behavior and packaging reliability remain to be investigated in future work.

Sep. 11, 2026
Infrared and Laser Engineering
Vol. 55, Issue 8, 20260080 (2026)
DOI：10.3788/IRLA20260080

Deep learning enabled real-time acoustic monitoring for porosity in laser wire directed energy deposition

Xiangfei Zhang, Huomu Yang and Shouhuan Zhou

ObjectivePorosity defects are prone to occur during the fabrication of components via Laser Wire Directed Energy Deposition (LWDED). These defects severely compromise the mechanical performance of the manufactured parts. Therefore, the implementation of real-time porosity monitoring during the manufacturing process is of great significance for reducing production costs. Acoustic signals generated by laser-material interaction contain critical information regarding the formation of defects such as cracks and pores. Utilizing these acoustic signals for defect identification offers several advantages, including low hardware costs, excellent real-time performance, and strong resistance to intense light interference. In the field of LWDED defect monitoring, current research primarily relies on light intensity, spectral information, and temperature fields. However, the use of acoustic signals for porosity monitoring in LWDED has not yet been reported in the literature. The study represents the application of deep learning-based acoustic monitoring to the real-time detection of porosity in LWDED. The research is highly significant for ensuring the service reliability of components and reducing trial-and-error manufacturing costs.MethodsThis study established a real-time acoustic monitoring workflow for porosity based on deep learning (Fig.1). Acoustic frequency bands containing porosity information were identified through frequency-domain and time-frequency analysis, followed by the application of band-pass filtering and acoustic equalization (Fig.4, Fig.5). Subsequently, Mel-frequency cepstral coefficient (MFCC) spectrograms were extracted from these denoised signals (Fig.9). These MFCC features were then used to train the ResNet18-CBAM model (Fig.10). Finally, the generalization capability and classification performance of this model were compared with other machine learning models.Results and DiscussionsThe t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm was employed to visualize and analyze the extracted acoustic features through dimensionality reduction. The results indicate that the denoising process improved the silhouette coefficient of the feature clusters from 0.017 to 0.352(Fig.8). The MFCC spectrograms were utilized to train the ResNet18-CBAM model, which exhibited excellent generalization capability (Fig.12). The model achieved an accuracy of 0.89 (Tab.4) and an AUC value of 0.965(Fig.13). The ablation experiment further verifies that the CBAM module improves the overall classification performance of the model (the accuracy is increased by 0.04)（Tab.5）. These results confirm that the model can effectively identify specimens with varying porosity levels.ConclusionsThis study applies deep learning-based acoustic monitoring to the real-time detection of porosity in the LWDED process. Acoustic frequency bands containing porosity information were identified through frequency-domain and time-frequency analyses. Subsequently, Mel-frequency cepstral coefficient (MFCC) spectrograms were extracted from the denoised signals. These features were utilized to train the ResNet18-CBAM model, enabling real-time monitoring of LWDED porosity. The experimental results indicate that the model effectively identifies specimens with varying porosity levels. This approach provides a valuable reference for the real-time monitoring of porosity defects in LWDED.ObjectivePorosity defects are prone to occur during the fabrication of components via Laser Wire Directed Energy Deposition (LWDED). These defects severely compromise the mechanical performance of the manufactured parts. Therefore, the implementation of real-time porosity monitoring during the manufacturing process is of great significance for reducing production costs. Acoustic signals generated by laser-material interaction contain critical information regarding the formation of defects such as cracks and pores. Utilizing these acoustic signals for defect identification offers several advantages, including low hardware costs, excellent real-time performance, and strong resistance to intense light interference. In the field of LWDED defect monitoring, current research primarily relies on light intensity, spectral information, and temperature fields. However, the use of acoustic signals for porosity monitoring in LWDED has not yet been reported in the literature. The study represents the application of deep learning-based acoustic monitoring to the real-time detection of porosity in LWDED. The research is highly significant for ensuring the service reliability of components and reducing trial-and-error manufacturing costs.MethodsThis study established a real-time acoustic monitoring workflow for porosity based on deep learning (Fig.1). Acoustic frequency bands containing porosity information were identified through frequency-domain and time-frequency analysis, followed by the application of band-pass filtering and acoustic equalization (Fig.4, Fig.5). Subsequently, Mel-frequency cepstral coefficient (MFCC) spectrograms were extracted from these denoised signals (Fig.9). These MFCC features were then used to train the ResNet18-CBAM model (Fig.10). Finally, the generalization capability and classification performance of this model were compared with other machine learning models.Results and DiscussionsThe t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm was employed to visualize and analyze the extracted acoustic features through dimensionality reduction. The results indicate that the denoising process improved the silhouette coefficient of the feature clusters from 0.017 to 0.352(Fig.8). The MFCC spectrograms were utilized to train the ResNet18-CBAM model, which exhibited excellent generalization capability (Fig.12). The model achieved an accuracy of 0.89 (Tab.4) and an AUC value of 0.965(Fig.13). The ablation experiment further verifies that the CBAM module improves the overall classification performance of the model (the accuracy is increased by 0.04)（Tab.5）. These results confirm that the model can effectively identify specimens with varying porosity levels.ConclusionsThis study applies deep learning-based acoustic monitoring to the real-time detection of porosity in the LWDED process. Acoustic frequency bands containing porosity information were identified through frequency-domain and time-frequency analyses. Subsequently, Mel-frequency cepstral coefficient (MFCC) spectrograms were extracted from the denoised signals. These features were utilized to train the ResNet18-CBAM model, enabling real-time monitoring of LWDED porosity. The experimental results indicate that the model effectively identifies specimens with varying porosity levels. This approach provides a valuable reference for the real-time monitoring of porosity defects in LWDED.

Sep. 11, 2026
Infrared and Laser Engineering
Vol. 55, Issue 8, 20260114 (2026)
DOI：10.3788/IRLA20260114

Semantic prior-guided wavelength-dependent multispectral defogging

Ju Cheng, Feng Huang, Ren Liu, and Yu Liu

ObjectiveUnder foggy conditions, light reflected from a target is corrupted by atmospheric scattering, resulting in image degradation characterized by diminished contrast and loss of detail. Such degradation severely compromises the performance of downstream computer vision tasks. Consequently, developing effective defogging methods to mitigate fog interference is critically important. While traditional RGB images offer limited spectral information due to their narrow band coverage, multispectral imagery spans the visible, near-infrared, and other spectral ranges. This enables a more comprehensive acquisition of scene information, granting multispectral imaging significant advantages in fog-penetrating applications and greater potential for downstream tasks such as anomaly detection. However, the effective exploitation of rich multispectral data is critical; meanwhile, haze concentration and spectral response exhibit significant spatial-spectral heterogeneity across different regions. Existing multispectral defogging methods inadequately exploit rich spectral information and often overlook regional variations in spatial-spectral characteristics. This limitation substantially hinders further improvements in image reconstruction quality. Therefore, implementing a region-adaptive spectral selection strategy following image restoration is a promising direction for advancing multispectral defogging technology.MethodsA semantic feature-guided real-world multispectral defogging method based on the Wavelength-Dependent Extinction Coefficient (WDEC) is employed, termed SFG-WDEC (Fig.2). The method consists of two main modules: the Multispectral Defogging (MSD) module and the Optimal Spectral Selection (OSS) module. In the MSD module, the WDEC algorithm is first applied to generate equivalent dark channel maps and equivalent transmittance maps for each spectral band. Based on the equivalent dark channel maps, the global atmospheric light for each channel is estimated. By integrating these two types of maps, the precise transmittance of each spectral channel is determined, enabling the reconstruction of defogged multispectral sub-images. In the OSS module, pseudo-RGB images are synthesized and subjected to semantic segmentation to extract semantic features—namely scene masks—of different target regions. These regional masks are then mapped onto the defogged multispectral sub-images. Guided by the presented Spatial–Frequency Domain Joint Perception (SFDJP) adaptive spectral selection index, the optimal spectral channel is selected for each region according to the criterion of maximizing comprehensive information. Finally, a grayscale defogged image integrated with optimal spectral information is reconstructed.Results and DiscussionsThe presented SFG-WDEC method consistently outperforms state-of-the-art defogging algorithms, demonstrating distinct advantages in contrast enhancement, multispectral information exploitation, and detail preservation. The effectiveness of the region-wise spectral selection strategy is further validated through comparison with the baseline WDEC method. In qualitative experiments conducted on a self-built real-world foggy multispectral dataset (Fig.7), SFG-WDEC outperforms all other compared methods. While the latter typically suffer from either residual haze or insufficient contrast, SFG-WDEC not only effectively removes fog but also substantially enhances image contrast, resulting in clearly distinguishable boundaries between the sky and mountain regions. Quantitative evaluations (Tab.1) further confirm the superiority of SFG-WDEC across multiple performance metrics, reflecting its enhanced capability in defogging tasks. Compared to the baseline WDEC method, the presented approach shows an increase of 2.9% in entropy, 22.4% in average gradient, and 9.9% in spatial frequency. It also reduces the fog aware density evaluator (FADE) by 8.6% and the integrated local natural image quality evaluator (IL-NIQE) by 8.0%, demonstrating a notable overall enhancement. All in all, the method achieves high levels of image information retention, texture preservation, and detail restoration, demonstrating overall superiority over existing approaches. It is worth noting that, among the five selected blind image quality assessment (BIQA) metrics, SFG-WDEC achieved the best performance in four of them, demonstrating its comprehensive advantage in image quality enhancement. These results indicate that SFG-WDEC not only effectively removes haze interference but also preserves the structure and detail information of the original scene to the greatest extent.ConclusionsExisting multispectral defogging methods fail to fully leverage the spectral richness and often overlook significant variations in spatial-spectral characteristics across regions. To address this limitation, semantic guidance is innovatively integrated into the defogging pipeline, and a region-aware spectral reconstruction strategy that adapts to the spectral characteristics of different targets is proposed. The presented approach combines a multispectral defogging framework with a semantic segmentation model. Guided by the specially designed SFDJP spectral selection index, it adaptively selects the optimal spectral band for each distinct region and ultimately reconstructs a hybrid spectral image that achieves both effective haze removal and high information fidelity. Furthermore, a real-world multispectral dataset for foggy scenes is constructed, effectively filling the gap caused by the scarcity of existing datasets in this field. Qualitative and quantitative experiments demonstrate that the presented method not only effectively mitigates fog-induced degradation — significantly enhancing visual clarity and image contrast — but also improves regional distinguishability. It fully exploits the defogging potential of the WDEC framework and outperforms five other state-of-the-art RGB-based defogging methods in overall performance. Furthermore, the method is compatible with future advanced multispectral defogging algorithms, offering a promising and effective direction for research in this field.ObjectiveUnder foggy conditions, light reflected from a target is corrupted by atmospheric scattering, resulting in image degradation characterized by diminished contrast and loss of detail. Such degradation severely compromises the performance of downstream computer vision tasks. Consequently, developing effective defogging methods to mitigate fog interference is critically important. While traditional RGB images offer limited spectral information due to their narrow band coverage, multispectral imagery spans the visible, near-infrared, and other spectral ranges. This enables a more comprehensive acquisition of scene information, granting multispectral imaging significant advantages in fog-penetrating applications and greater potential for downstream tasks such as anomaly detection. However, the effective exploitation of rich multispectral data is critical; meanwhile, haze concentration and spectral response exhibit significant spatial-spectral heterogeneity across different regions. Existing multispectral defogging methods inadequately exploit rich spectral information and often overlook regional variations in spatial-spectral characteristics. This limitation substantially hinders further improvements in image reconstruction quality. Therefore, implementing a region-adaptive spectral selection strategy following image restoration is a promising direction for advancing multispectral defogging technology.MethodsA semantic feature-guided real-world multispectral defogging method based on the Wavelength-Dependent Extinction Coefficient (WDEC) is employed, termed SFG-WDEC (Fig.2). The method consists of two main modules: the Multispectral Defogging (MSD) module and the Optimal Spectral Selection (OSS) module. In the MSD module, the WDEC algorithm is first applied to generate equivalent dark channel maps and equivalent transmittance maps for each spectral band. Based on the equivalent dark channel maps, the global atmospheric light for each channel is estimated. By integrating these two types of maps, the precise transmittance of each spectral channel is determined, enabling the reconstruction of defogged multispectral sub-images. In the OSS module, pseudo-RGB images are synthesized and subjected to semantic segmentation to extract semantic features—namely scene masks—of different target regions. These regional masks are then mapped onto the defogged multispectral sub-images. Guided by the presented Spatial–Frequency Domain Joint Perception (SFDJP) adaptive spectral selection index, the optimal spectral channel is selected for each region according to the criterion of maximizing comprehensive information. Finally, a grayscale defogged image integrated with optimal spectral information is reconstructed.Results and DiscussionsThe presented SFG-WDEC method consistently outperforms state-of-the-art defogging algorithms, demonstrating distinct advantages in contrast enhancement, multispectral information exploitation, and detail preservation. The effectiveness of the region-wise spectral selection strategy is further validated through comparison with the baseline WDEC method. In qualitative experiments conducted on a self-built real-world foggy multispectral dataset (Fig.7), SFG-WDEC outperforms all other compared methods. While the latter typically suffer from either residual haze or insufficient contrast, SFG-WDEC not only effectively removes fog but also substantially enhances image contrast, resulting in clearly distinguishable boundaries between the sky and mountain regions. Quantitative evaluations (Tab.1) further confirm the superiority of SFG-WDEC across multiple performance metrics, reflecting its enhanced capability in defogging tasks. Compared to the baseline WDEC method, the presented approach shows an increase of 2.9% in entropy, 22.4% in average gradient, and 9.9% in spatial frequency. It also reduces the fog aware density evaluator (FADE) by 8.6% and the integrated local natural image quality evaluator (IL-NIQE) by 8.0%, demonstrating a notable overall enhancement. All in all, the method achieves high levels of image information retention, texture preservation, and detail restoration, demonstrating overall superiority over existing approaches. It is worth noting that, among the five selected blind image quality assessment (BIQA) metrics, SFG-WDEC achieved the best performance in four of them, demonstrating its comprehensive advantage in image quality enhancement. These results indicate that SFG-WDEC not only effectively removes haze interference but also preserves the structure and detail information of the original scene to the greatest extent.ConclusionsExisting multispectral defogging methods fail to fully leverage the spectral richness and often overlook significant variations in spatial-spectral characteristics across regions. To address this limitation, semantic guidance is innovatively integrated into the defogging pipeline, and a region-aware spectral reconstruction strategy that adapts to the spectral characteristics of different targets is proposed. The presented approach combines a multispectral defogging framework with a semantic segmentation model. Guided by the specially designed SFDJP spectral selection index, it adaptively selects the optimal spectral band for each distinct region and ultimately reconstructs a hybrid spectral image that achieves both effective haze removal and high information fidelity. Furthermore, a real-world multispectral dataset for foggy scenes is constructed, effectively filling the gap caused by the scarcity of existing datasets in this field. Qualitative and quantitative experiments demonstrate that the presented method not only effectively mitigates fog-induced degradation — significantly enhancing visual clarity and image contrast — but also improves regional distinguishability. It fully exploits the defogging potential of the WDEC framework and outperforms five other state-of-the-art RGB-based defogging methods in overall performance. Furthermore, the method is compatible with future advanced multispectral defogging algorithms, offering a promising and effective direction for research in this field.

Sep. 11, 2026
Infrared and Laser Engineering
Vol. 55, Issue 8, 20260053 (2026)
DOI：10.3788/IRLA20260053

Very long wavelength quantum cascade detector with variable-barrier structure

Yijing Zhang, Yu Chen, Kai Guo, Yongqiang Sun, Jinlin Yao, Jiagang Chen, Zhiwei Jia, Qingtian Li, Fengmin Cheng, Jinchuan Zhang, Fengqi Liu, and Shenqiang Zhai