The challenge stems from the very technique that makes optical wireless so spectrally efficient: high-order quadrature amplitude modulation, or QAM. Schemes such as 256-QAM and 512-QAM pack enormous amounts of data into each symbol by using extremely dense constellations of amplitude and phase combinations. But as the constellation density grows, the decision boundaries between neighboring symbols shrink to razor-thin margins, making the signal exquisitely sensitive to noise, phase fluctuations, and the nonlinear distortions introduced by optical components such as light-emitting diodes and laser diodes. In intensity-modulated direct-detection systems, where light intensity cannot go negative, engineers must also contend with a high peak-to-average power ratio, which further degrades signal quality. Traditional receivers that rely on analytical models and linear approximations simply cannot capture these nonlinearities, and their performance collapses precisely where 6G needs it most.

The research team built their evaluation around a two-user optical NOMA system, in which both users share the same time and frequency resources through power-domain superposition coding. The user with the weaker channel is allocated 80 percent of the transmit power while the stronger user receives 20 percent, and at the receiver a successive interference cancellation procedure separates the overlapping signals. The channel model incorporated a realistic battery of impairments: distance-dependent path loss with an exponent of 2.2, Rayleigh fading, additive white Gaussian noise, optical nonlinear distortion from the light source, and hardware impairments including synchronization errors. This deliberately harsh environment was designed to emulate the conditions a real 6G optical link would face, with independent channel realizations generated for each user to reflect heterogeneous propagation conditions.

Against this backdrop, the authors evaluated six detectors within a single unified framework. The minimum mean square error (MMSE) detector offers a favorable complexity-performance trade-off but remains fundamentally linear. The QR-decomposition with M-algorithm maximum likelihood detector (QRM-MLD) approaches optimal performance through tree search but becomes computationally prohibitive at high modulation orders. Convolutional neural network (CNN) detectors learn spatial signal patterns efficiently but cannot model temporal dependencies, while recurrent neural network (RNN) detectors capture channel memory and time correlations at the cost of slower convergence. Two hybrids, CNN-MMSE and the newly proposed RNN-MMSE, combine statistical estimation with deep learning refinement. Prior studies had examined these techniques only in isolation; the unified comparison is itself a significant contribution.

The proposed RNN-MMSE detector operates in three stages. First, an RNN learns the channel characteristics directly from pilot symbols, replacing conventional pilot-based estimation and capturing temporal variations in the process. Second, the MMSE stage applies a statistically optimal linear estimate that suppresses noise and linear interference, producing a structured initial symbol estimate. Third, a second RNN refines that estimate by learning only the residual nonlinear distortions and temporal dependencies that the linear stage could not remove. The authors provide a theoretical justification rooted in the orthogonality principle of linear estimation: because the MMSE output extracts all linearly available information, the residual presented to the recurrent network has a substantially smaller variance than the raw signal, which smooths the optimization landscape, lowers gradient variance, and accelerates training convergence.

The simulation results are striking. At 64-QAM, conventional optical NOMA detection required approximately 13.5 dB of signal-to-noise ratio to achieve a bit error rate of 10 to the minus 3, while the proposed RNN-MMSE detector reached the same target at just 6.5 to 7 dB. As the modulation order climbed, the gap widened. At 128-QAM the hybrid detector delivered a 7 to 8 dB improvement, at 256-QAM an 8 to 9 dB gain, and at 512-QAM, the most demanding scheme tested, it achieved a 9 to 10 dB advantage, reaching the target error rate at roughly 12.5 to 13 dB where the conventional baseline needed 21 to 22 dB. Standalone CNN and RNN detectors and the CNN-MMSE hybrid improved on the classical methods but consistently fell short of the proposed architecture.

Spectral behavior told a similar story. Using Welch power spectral density estimation with Hamming windowing, the team measured out-of-band leakage across all detectors. Conventional optical NOMA exhibited in-band power spectral density around minus 40 dB/Hz with poor suppression near minus 60 dB/Hz, while the RNN-MMSE detector pushed spectral leakage below minus 130 dB/Hz at 512-QAM, indicating excellent sidelobe suppression and minimal adjacent-channel interference. Training convergence favored the hybrid as well: over 20 epochs the RNN-MMSE model began at roughly 85 percent accuracy and converged to about 90 percent, outperforming standalone models by 18 to 32 percentage points. Statistical validation across 20 independent Monte Carlo runs confirmed the results were reproducible, with a standard deviation of only 0.22 dB on the key SNR metric and a 95 percent confidence interval of plus or minus 0.10 dB.

Practicality was a central concern. The entire framework was implemented in MATLAB R2024a on a workstation with an Intel Core i7 processor, 32 GB of RAM, and an NVIDIA RTX 3060 GPU. Training on 100,000 generated signal samples took only 14 to 16 minutes, and inference required approximately 3.8 milliseconds per OFDM frame, fast enough for near-real-time deployment. Memory consumption stayed below 800 MB during inference. The authors acknowledge that the hybrid architecture carries more computational overhead than a standalone linear detector, with complexity of the order of N cubed for the MMSE stage plus N squared terms for the recurrent stage, but they argue the performance gains justify the cost, particularly since the heavy training can be performed offline once while deployment requires only lightweight forward inference that GPUs, FPGAs, or edge accelerators can handle.

The study is candid about its limits. All results derive from Monte Carlo simulation rather than a physical testbed, and the authors identify experimental validation on an LED- or laser-based intensity-modulation link with software-defined-radio baseband processing as an immediate priority. Their comparison with Transformer-based detectors draws on literature benchmarks obtained under different channel models rather than a controlled head-to-head experiment, and their robustness analysis of imperfect channel estimation, imperfect interference cancellation, and synchronization errors is mechanistic rather than simulated. The two-user configuration, while modular and scalable in principle, will need extension to denser multi-user networks with adaptive grouping and dynamic power allocation. Even so, the findings make a compelling case that hybrid intelligent detection is not an incremental refinement but a necessary evolution for 6G optical networks, and the roadmap toward testbed validation, attention-based comparisons, and large-scale multi-user trials is already clearly drawn.

Subject of Research: Neural network-based signal detection for 6G optical non-orthogonal multiple access systems with high-order QAM modulation

Article Title: Comprehensive analysis of 6G optical NOMA waveforms using neural network-based detection for high-order QAM modulation

Article References: Kumar, A., Revathi, V., Gaur, N., & Nanthaamornphong, A. (2026). Comprehensive analysis of 6G optical NOMA waveforms using neural network-based detection for high-order QAM modulation. Results in Optics, 25, Article 101159. https://doi.org/10.1016/j.rio.2026.101159

Image Credits: AI Generated

DOI: 10.1016/j.rio.2026.101159

Keywords: 6G, optical wireless communication, NOMA, QAM, recurrent neural network, MMSE detection, deep learning, bit error rate, power spectral density, IM/DD, OFDM, signal detection

The framework, called FAGNet, short for Frequency-Guided Adaptive Graph Network, begins with a deceptively simple observation. Financial time series are not single, uniform phenomena. They are superpositions of processes operating at different speeds: fast, high-frequency fluctuations driven by intraday trading and noise, medium-frequency swings tied to weekly and monthly cycles, and slow, low-frequency drifts that reflect macroeconomic fundamentals. When a graph neural network ingests these mixed signals directly, it must somehow learn to disentangle all of these scales simultaneously, a task the authors argue is fundamentally ill-suited to time-domain processing. Their solution is to apply the Fourier transform, a mathematical technique with a century of pedigree in physics and engineering, to project the coupled stock price signals into the frequency domain, where each oscillation speed becomes a distinct, measurable component.

But simply moving to the frequency domain is not enough, and this is where the technical heart of the paper lies. A naive frequency decomposition, such as splitting the spectrum into bands of equal width, would not guarantee that each band carries comparable information. Two bands of identical width can contain wildly different amounts of energy, and therefore wildly different amounts of signal. FAGNet addresses this with an adaptive energy-based spectral partitioning mechanism guided by the power spectral density, a function that describes how the power of a signal is distributed across frequencies. Rather than carving the spectrum into equal slices, the model divides it into sub-bands that each contain roughly equal energy. The result, the authors contend, is genuine scale decoupling at the level of information content: each sub-band represents a comparable share of the market’s total dynamical activity, so no scale is systematically over- or under-represented in the downstream analysis.

Once the spectrum has been partitioned, the framework constructs a dedicated graph neural network for each decoupled sub-band. This is a crucial architectural decision. Risk contagion, the authors argue, is scale-dependent: the network of relationships through which a fast-moving shock propagates may look very different from the network along which slow-moving fundamentals diffuse. A rapid panic might spread through algorithmic trading links and sentiment spillovers, while a slow repricing might follow industrial supply chains and shared analyst coverage, a phenomenon documented in the finance literature. By modelling each frequency band with its own graph network, FAGNet can learn the contagion topology that is specific to that scale, rather than forcing a single graph structure to explain contagion at every speed simultaneously. The separated scale signals thus become the raw material for a family of specialised graph learners, each attuned to one slice of the market’s rhythm.

The final stage of the pipeline tackles a problem that arises whenever multiple models are combined: how should their outputs be fused? A static weighted average would be blind to changing market conditions, yet the relative importance of different scales is precisely what shifts during a crisis. FAGNet therefore employs a cross-scale interaction and context-aware dynamic fusion module. This module does two things. First, it captures synergistic dependencies between risk contagions at different scales, recognising that a slow-moving fundamental weakness can amplify a fast-moving panic, and vice versa. Second, it adaptively adjusts the prediction weight assigned to each scale according to the instantaneous state of the market. In turbulent conditions, the model can lean more heavily on the frequency bands that carry the most relevant information; in calmer periods, it can rebalance toward the slower components. The fusion is not a fixed formula but a learned, state-sensitive mechanism.

The motivation for this frequency-domain perspective draws on a substantial body of prior research. Economists have long known that systematic risk behaves differently at different time scales, with wavelet-based studies of emerging stock markets showing that the relationship between risk and horizon is far from uniform. More recently, the machine learning community has embraced frequency methods for time series: frequency-enhanced decomposed transformers have improved long-term forecasting, frequency-domain multilayer perceptrons have proven to be effective learners, and Fourier-based graph networks have been applied to multivariate prediction problems. FAGNet extends this line of work by combining frequency decomposition with graph-based contagion modelling in a principled way, using the energy content of the spectrum, rather than arbitrary frequency cutoffs, to define the scales themselves.

The empirical case for the approach rests on extensive experiments on real-world stock market datasets, where the authors report that FAGNet significantly outperforms state-of-the-art baseline methods. The baselines against which such models are typically judged include temporal relational ranking systems, dynamic graph attention networks, integrated convolutional and recurrent architectures built on knowledge-incorporated graphs, and hybrid frameworks that combine graph attention with recurrent memory cells. The claim that a frequency-guided design beats these established approaches suggests that the multi-scale entanglement of risk contagion is not merely a theoretical nuisance but a genuine bottleneck limiting the accuracy of existing models. The authors argue that the results validate frequency domain decoupling as a new paradigm for financial risk modelling, one that is both more accurate and more dynamically interpretable than time-domain alternatives.

That word, interpretable, deserves emphasis, because it points to what may be the framework’s most valuable property for practitioners. Deep learning models in finance are often criticised as black boxes: they may predict well, but they offer little insight into why. FAGNet’s architecture is different in kind. Because each graph network operates on a well-defined frequency band, its learned contagion topology can be read as a statement about how risk propagates at that particular scale. The dynamic fusion weights, meanwhile, provide a running record of which scales the model considers most informative at any moment. In principle, an analyst could inspect these weights during a market stress event and see whether the model is relying on fast, panic-like components or slow, fundamental ones. This is a form of transparency that emerges naturally from the design, rather than being bolted on afterwards, and it aligns with a broader movement in the field toward models whose internal structure mirrors the phenomena they describe.

The study also situates itself within a rapidly growing literature on graph-based financial prediction. Graph neural networks have become the dominant approach for modelling risk contagion, with applications ranging from recommending profitable stocks via financial graph attention networks to predicting stock movements using hybrid-relational market knowledge graphs. Yet the authors identify a persistent limitation: the vast majority of deep graph models operate directly in the raw time domain, making it difficult for them to handle the multi-scale entanglement and dynamic variability of contagion. Related work by some of the same authors, including a frequency-domain graph learning framework for understanding risk propagation and a frequency-decoupled progressive graph learning approach for heterogeneous contagion, indicates that this research group has been systematically developing the ideas that culminate in FAGNet. The new paper consolidates that programme into a single adaptive architecture.

Caveats remain, as they always do in financial machine learning. The paper reports that data will be made available on request, and independent replication on different markets, asset classes and time periods will be needed to establish how general the gains are. Markets are adversarial environments where any published predictive edge tends to erode as it is arbitraged away, and no architecture, however elegant, escapes that dynamic. Still, the conceptual contribution stands on its own merits. By treating financial contagion as a multi-scale signal processing problem, partitioning the spectrum by energy content, and letting each scale have its own graph model with state-aware fusion, FAGNet offers a template that could plausibly extend beyond equities to other networked dynamical systems, from cryptocurrency markets to interbank lending networks. For a field searching for ways to make deep learning both sharper and more transparent, the message is that sometimes the answer is not a bigger network, but a better coordinate system in which to run it.

Subject of Research: Frequency-domain graph neural networks for financial risk contagion and stock market prediction

Article Title: Fagnet: an adaptive frequency-guided multi-scale graph network for financial market prediction

Article References: Zhou, C., Xiao, S., & Hu, C. (2026). Fagnet: an adaptive frequency-guided multi-scale graph network for financial market prediction. International Journal of Machine Learning and Cybernetics, 17(10), Article 485. https://doi.org/10.1007/s13042-026-03301-3

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03301-3

Keywords: graph neural networks, stock market prediction, frequency domain analysis, risk contagion, Fourier transform, power spectral density, multi-scale learning, financial networks, deep learning, dynamic fusion, interpretable AI, machine learning

The scientific and legal momentum behind tools like TUNE stems largely from the European Union’s Marine Strategy Framework Directive, the landmark 2008 legislation that obliges member states to achieve Good Environmental Status in their marine waters. Under the Directive’s Descriptor 11, which covers the introduction of energy, including underwater sound, two distinct criteria apply. The first addresses low- and mid-frequency impulsive sounds, such as the hammering of pile drivers during offshore construction, while the second targets continuous low-frequency noise, dominated overwhelmingly by commercial shipping. Shipping noise is the most pervasive form of underwater noise pollution and dominates ambient sound levels below one kilohertz across much of the world’s oceans. Criterion D11C2 specifically requires monitoring programs to report sound pressure levels within decidecade frequency bands centered at 63 and 125 hertz, two narrow slices of the acoustic spectrum that serve as proxies for overall shipping-related noise exposure.

Meeting that regulatory requirement sounds straightforward, but in practice it has been anything but. Reliable estimation of these acoustic quantities depends on robust and reproducible signal processing workflows, and the software landscape has long been fragmented. Raven Pro, a mainstay of bioacoustics research, offers limited flexibility in calibration and requires a commercial license. MATLAB-based tools such as CHORUS provide advanced processing capabilities but are not distributed as freely available standalone applications. The open-source MANTA package has improved accessibility and standardization for long-term ocean soundscape analysis, but it produces outputs in hybrid millidecade bands rather than the decidecade bands adopted for Marine Strategy Framework Directive reporting. Many other packages remain proprietary or closed-source, preventing users from inspecting and verifying the algorithms that ultimately feed into official environmental assessments. TUNE was engineered to close this gap by combining calibration accuracy, usability and regulatory compliance in a single freely available package distributed under the GNU General Public License.

At its core, TUNE processes calibrated acoustic recordings in wav format and computes a suite of standard underwater noise metrics. These include broadband sound pressure level, decidecade sound pressure level, zero-to-peak sound pressure level and power spectral density estimates. The software embraces recent international standardization: it uses the term decidecade in place of one-third octave band, following the 2025 update of the relevant ISO standard, which considers the two metrics approximately equivalent, and it adopts the sound pressure symbol conventions of ISO 18405:2017 for the definitions and equations of all four metrics. Processing is carried out over user-defined time intervals called snapshots, a concept aligned with the temporal observation windows introduced in recent ISO documents. Each recording can be segmented into consecutive, skipped or overlapping snapshots, allowing fully time-resolved estimates of how noise levels rise and fall, for example as a vessel approaches and then recedes from a monitoring station.

A distinguishing technical feature of TUNE is its support for frequency-dependent calibration. Hydrophones, the underwater microphones that record ocean sound, do not respond uniformly across frequencies, and their sensitivity can vary substantially across the acoustic spectrum. Many existing tools accept only a single calibration factor, which introduces errors when the hydrophone response is uneven. TUNE instead accepts a two-column text file containing frequency values and corresponding hydrophone sensitivity, expressed either as voltage sensitivity in decibels relative to one volt per micropascal or as an equivalent calibration factor in pascals per volt, with a minimum of three frequency points required. Spectral estimates are corrected bin by bin using this sensitivity curve, improving the accuracy of the resulting sound pressure levels. The software also allows optional high-pass, low-pass or band-pass digital filtering before analysis, with filters designed to fixed passband ripple and stopband attenuation specifications, and it supports flexible FFT parameters including window type, overlap and transform length.

Internally, the software follows a modular architecture organized into five stages: user input, data import, calibration, signal processing and output generation. A single-window graphical user interface lets users configure directories, calibration files, time ranges, spectrogram settings and filtering options, and the entire parameter set can be saved and reloaded as a setup file to guarantee reproducibility across monitoring campaigns. Processing proceeds through three nested loops at the file, channel and snapshot levels, with multi-channel recordings handled channel by channel. For each snapshot, the software computes a series of Fast Fourier Transforms over signal subsegments and averages them into a single representative spectrum, then corrects that spectrum against the calibration curve. Decidecade integration is performed using the nearest available FFT bins rather than exact theoretical band edges, a pragmatic approximation that benchmarking described in the paper’s supplementary materials shows produces negligible discrepancies for standard monitoring applications, while dramatically accelerating computation. One creative addition is the option to overlay Knudsen curves for sea states zero through six, letting analysts compare measured ambient noise against classic empirical predictions of wind-driven sea-state noise.

To demonstrate the tool, the researchers analyzed a real recording of a ship passing through a harbor, captured with a Teledyne Reson TC4040 hydrophone and VP1000 preamplifier feeding a Tascam DR-100 recorder at 96 kilohertz sampling with 24-bit resolution. The ten-minute recording was processed in ten-second snapshots using 131072-point FFTs with a Hann window and fifty percent overlap, and the entire computation completed in just two minutes and thirty-five seconds. The results were vivid: mean sound pressure levels in the 63 and 125 hertz decidecade bands reached 111 and 117 decibels respectively, the mean broadband level was 123 decibels, and the maximum zero-to-peak level hit 156 decibels at 256 seconds into the recording, precisely the closest point of approach of the vessel. All four metrics rose clearly during the first half of the recording, matching the ship passage documented in the campaign logs, and every output was delivered both as figures and as numerical text files ready for statistical aggregation across recordings, sites and seasons.

The impact of the software is already visible in practice. By providing a consistent, transparent and reproducible framework for acoustic data processing, TUNE directly addresses a long-standing problem: differences in signal processing workflows can produce substantial discrepancies in reported sound pressure levels, undermining the comparability of environmental assessments across institutions and countries. Perhaps most significantly, the tool has been adopted by multiple Regional Environmental Protection Agencies in Italy, establishing a shared national framework for underwater noise monitoring and overcoming methodological fragmentation among regional programs. This software-driven harmonization improves the consistency of reported indicators, increases transparency in reporting and supports the coordinated implementation of national monitoring obligations under the Marine Strategy Framework Directive. Because TUNE can process long-term acoustic datasets and generate decidecade-band time series, it also supports the estimation of baseline conditions, trend analysis and threshold exceedance assessments in line with recent European Commission guidance on threshold values.

The developers see TUNE as a foundation rather than a finished product. Future versions will extend the range of computed metrics to include the Sound Exposure Level, which would allow the software to cover the Directive’s impulsive-sound criterion in addition to continuous noise, along with greater flexibility in temporal analysis. Beyond European regulatory monitoring, the authors note that the tool’s applicability extends to a broad range of marine and freshwater acoustic studies, from low-noise calibration sites to research linking low-frequency noise shifts to changes in fin whale song. In a field where proprietary licenses and closed algorithms have often stood between environmental agencies and their legal obligations, TUNE represents a quietly radical proposition: that the measurement of a pollutant threatening ocean life should be accurate, transparent and free for anyone who needs it.

Subject of Research: A free standalone software tool for standardized evaluation of underwater noise pollution under the Marine Strategy Framework Directive

Article Title: TUNE: a Tool for Underwater Noise Evaluation

Article References: Caradonna, V., Buogo, S., Azzellino, A., Vicinanza, D., & Borsani, J. F. (2026). TUNE: a Tool for Underwater Noise Evaluation. SoftwareX, 35, Article 103008. https://doi.org/10.1016/j.softx.2026.103008

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103008

Keywords: underwater noise, TUNE software, Marine Strategy Framework Directive, shipping noise, hydrophone calibration, sound pressure level, decidecade bands, power spectral density, ocean acoustics, marine monitoring, MATLAB, environmental protection