The research team, led by Hoa T. Le of the University of Transport and Communications in Vietnam together with collaborators including Peter LoPresti of the University of Tulsa, built a complete FSO-SLIPT testbed in which atmospheric turbulence was deliberately generated, measured, and varied as an experimental parameter. This distinguishes the work from earlier laboratory demonstrations of optical data-and-power transfer, which typically ran over short, still-air paths where turbulence was never controlled or characterized. In those prior experiments, such as demonstrations of gigabit-per-second data-and-power links using gallium arsenide photovoltaic cells, the focus fell on receiver electronics and modulation formats rather than on the physics of a fluctuating channel. The new study instead treats turbulence as the primary variable, bringing experimental practice into line with the statistical channel models that theorists have relied on for years.

The underlying system architecture is elegant in its simplicity. At the transmitter, an on-off keying generator produces an alternating-current signal carrying the information, while a separate fiber-coupled laser at 1550 nanometers supplies a direct-current bias that represents pure optical energy. A wideband coupler merges the two into a single composite beam, so the light leaving the transmitter is simultaneously a data stream and a power beam. At the receiver, an indium gallium arsenide photodetector converts the incoming light back into an electrical signal, which is then fed into an AC-DC separator circuit. A high-pass branch, built from a 0.47-microfarad capacitor and a 560-ohm resistor, routes the rapidly varying data component to the decoder. A low-pass branch, using a 10-picofarad capacitor and a 120-kilo-ohm resistor, extracts the slow direct-current component that charges the energy-harvesting circuit. This AC/DC separation scheme is the simplest of the three main SLIPT receiver architectures, which also include time-splitting and power-splitting designs, and it allowed the team to measure both communication quality and harvested power from the same received signal.

The turbulence itself was created inside a customized chamber 77 centimeters long, fitted with two heaters along its base and two fans on its sides. By heating the air and stirring it with fans, the researchers produced convective pockets of air with different temperatures and densities, mimicking the refractive-index fluctuations that plague real outdoor and airborne optical links. Two scenarios were tested. Weak turbulence was generated with the chamber at 22 degrees Celsius and fans at maximum speed, yielding a scintillation index of 0.007 and a refractive-index structure constant of 5.41 times ten to the minus eleven. Moderate turbulence, produced at 100 degrees Celsius with heaters and fans both at full power, raised the scintillation index to 0.096 and the structure constant to 7.32 times ten to the minus ten. Both values fall in the weak-to-moderate regime, where the received intensity is well described by a log-normal statistical distribution, a fact the team confirmed by fitting the measured voltage histograms with log-normal probability density functions that achieved root-mean-square errors of just 0.016 and 0.013 respectively.

On the communication side, the results followed the expected and sobering pattern. The team transmitted on-off-keyed data at bit rates from 0.5 to 4 megabits per second and estimated the bit-error rate from the Q-factor, a standard quality metric derived from the mean and standard deviation of the voltages representing digital ones and zeros. At every bit rate, moderate turbulence produced a consistently higher bit-error rate than weak turbulence. The mechanism is visible in the voltage statistics: under weak turbulence the probability distributions for ones and zeros remain narrow and well separated, preserving a wide decision margin at the receiver. Under moderate turbulence, scintillation broadens both distributions and pulls them into overlap, shrinking the margin between the decision thresholds and driving errors upward. Theoretical BER curves computed from the measured voltage statistics tracked the experimental curves closely, with the small residual gap attributed to circuit noise and non-idealities in the AC-DC separator that the simplified linear model does not capture.

The energy results told a very different story. Average harvested power, calculated from the root-mean-square voltage on the DC branch across a one-megaohm load, showed no clear dependence on bit rate, fluctuating in both turbulence scenarios. But when the researchers compared turbulence strengths, the moderate-turbulence case consistently delivered more harvested power than the weak-turbulence case. Far from being a measurement artifact, this counterintuitive outcome follows directly from the mathematics of fluctuating signals. The harvested-power estimate depends on the second moment of the received intensity, and the scintillation index is defined as the normalized second moment. For a fixed mean received power, increasing the scintillation index mathematically increases the second moment, raising the RMS voltage and therefore the estimated harvested power. The effect requires no exotic nonlinear harvesting circuitry; it emerges purely from second-order signal statistics.

The voltage distributions on the DC branch make the physics tangible. Under weak turbulence, the DC voltage forms a sharp, narrow peak, indicating a steady and predictable power level. Under moderate turbulence, the distribution spreads dramatically, and occasional voltage samples climb to notably higher values. Those intermittent bursts of strong received intensity inflate the average harvested power, even though they also make the instantaneous power less reliable, as reflected in the higher power variance measured in the moderate case, 9.998 times ten to the minus eleven compared with 8.371 times ten to the minus eleven under weak turbulence. In other words, turbulence trades stability for intensity: the receiver sees a rockier energy supply, but on average a richer one.

The experimental protocol was designed for statistical rigor despite the modest scale of the benchtop link. Data were recorded at 10 million samples per second across 200 trials per measurement, totaling one second of capture, which corresponds to roughly one million bits at the 1-megabit-per-second setting and scales proportionally at higher rates. All measurements were conducted in total darkness to eliminate ambient light, and the scintillation metrics were computed from signals measured after the photodetector but before the separation circuit, with noise removed. The authors are candid about one methodological limitation: the Q-factor approach to BER estimation assumes approximately Gaussian decision statistics, whereas log-normal intensity fluctuations make the AC voltage slightly non-Gaussian. They argue that the method remains a widely accepted practical estimator when large sample counts are available, and the close agreement between experimental and theoretical BER trends supports that claim, while direct bit-by-bit error counting is deferred to future work.

The broader significance of the study lies in what it implies for the design of future optical wireless networks. Sixth-generation systems envision connectivity not only across massive terrestrial and underwater sensor fleets but also for unmanned aerial vehicles, high-altitude platforms, and satellites, all competing for scarce radio-frequency spectrum. Free-space optics offers fiber-like data rates and immunity to RF interference, and SLIPT extends it into a genuine data-plus-power delivery system for energy-constrained nodes. Until now, engineers designing such systems had only simulation results to guide them, and those simulations uniformly treated turbulence as a pure impairment. The new experimental evidence shows that this framing is incomplete. A system optimized purely for communication reliability would suppress turbulence effects wherever possible, yet the same fluctuations that corrupt the data can be exploited to boost energy delivery. The authors conclude that SLIPT-enabled FSO systems demand joint optimization of communication and energy objectives rather than conventional communication-centric design, and they point toward adaptive transmission strategies that dynamically balance reliability against harvested power as channel conditions change. While the 77-centimeter laboratory path is far shorter than any real deployment, the team notes that scintillation statistics over longer links obey the same underlying physics, so the observed trade-off is expected to persist. If the finding holds at scale, the turbulent air that once seemed destined to ruin laser links may instead become a resource to be managed, turning an old adversary of free-space optics into an unexpected ally for the power-hungry devices of the 6G era.

Subject of Research: Experimental evaluation of simultaneous lightwave information and power transfer over turbulent free-space optical links

Article Title: Laboratory evaluation of simultaneous lightwave information and power transfer in turbulent FSO environments

Article References: Le, H. T., Seifi, E., Nguyen, T. V., Pham, H. T., Dang, N. T., & LoPresti, P. (2026). Laboratory evaluation of simultaneous lightwave information and power transfer in turbulent FSO environments. Results in Optics, Article 101182. https://doi.org/10.1016/j.rio.2026.101182

Image Credits: AI Generated

DOI: 10.1016/j.rio.2026.101182

Keywords: free-space optics, SLIPT, atmospheric turbulence, wireless power transfer, 6G networks, bit-error rate, scintillation, energy harvesting, optical communications, log-normal fading, photodetector, laboratory experiment

Plasma bubbles form over equatorial and low-latitude regions and are among the most common nighttime plasma density depletion structures in the ionosphere. When they occur, they can cause severe scintillation interference on high-frequency communications, degrading satellite navigation, radar, and radio links that depend on signals traversing the affected region. That makes them a priority target for space weather monitoring. The conventional approach to finding them relies on statistical thresholds derived from temporal fluctuations in ion density measured by satellites. Analysts apply linear detrending to short data segments and flag an event when the density fluctuation exceeds a fixed cutoff. The problem is that plasma bubbles span spatiotemporal scales ranging from kilometers to hundreds of kilometers, and their occurrence frequency is strongly modulated by solar activity, season, and local time. A single fixed threshold simply cannot accommodate that diversity.

The consequences of this inflexibility are striking. According to the research team, fixed-threshold methods suffer miss rates of up to 40 percent during periods of low solar activity, precisely when understanding the baseline behavior of the ionosphere becomes critical for operators of communication and navigation systems. Missed bubbles mean unanticipated scintillation events, and false alarms erode confidence in automated alerts. The challenge, the researchers recognized, resembles problems that deep learning has already solved impressively in other domains, such as medical image segmentation and signal peak detection, where algorithms learn to extract and recognize complex waveform features automatically. Yet a systematic application of these methods to ionospheric plasma bubble detection had remained largely unexplored until now.

To build their model, the team first assembled a rigorous dataset. They used ion density measurements from the ion trap sensor aboard the FORMOSAT-1 satellite, which provides high-resolution data, covering the years 1999 through 2004. Combining the outputs of traditional detection methods with expert manual annotations, they established a dataset containing 13,675 plasma bubble events. Crucially, they also defined a new identification criterion to standardize what counts as a bubble: a density deviation from the background trend exceeding twice the standard deviation, an amplitude of at least 0.2, a duration of at least 5 seconds, and a transverse scale of no more than 500 kilometers. In ion density time series, bubbles manifest as sudden density drop features, a signature that lends itself naturally to the kind of pattern recognition at which neural networks excel.

The architecture of IncResUnet builds on U-Net, a convolutional network design originally developed for biomedical image segmentation and widely adopted for tasks requiring precise localization of features in data. The encoder portion of the network comprises five groups of convolutional modules and Inception-residual modules, progressively extracting multi-scale features through four max-pooling operations, while the decoder restores the original resolution through four upsampling operations. The key innovation lies in the Inception-residual modules themselves. These employ parallel one-dimensional convolution kernels of multiple sizes, allowing the network to capture density variation features across different temporal spans simultaneously, an essential capability given the wide range of bubble durations, which in the annotated examples ranged from as short as 18 seconds to more than a minute. Residual connections concatenate the processed features with the original input, effectively alleviating the vanishing gradient problem that can hamper the training of deep networks.

Training followed a carefully designed time-split strategy intended to test the model on conditions it had never seen. Data from 2001 to 2003 were used for training, data from 1999 served for validation, and data from 2000 and 2004 were reserved for testing. Because bubble events are relatively rare compared with quiet ionospheric conditions, the team applied random undersampling to balance the ratio of positive and negative samples. After the model produces its output, a dual-threshold refinement strategy cleans up the results: events shorter than 5 seconds are filtered out as noise, and adjacent events separated by less than 1 minute are automatically merged into complete large-scale plasma bubble structures. This post-processing pipeline ensures that the final detections correspond to physically meaningful events rather than fragmented fragments or spurious spikes.

The performance results are compelling. On the 2000 test set, IncResUnet achieved an F1-score of 0.914, a recall of 0.958, and a precision of 0.874. On the 2004 test set, it attained an F1-score of 0.914, a recall of 0.924, and a precision of 0.905, outperforming all comparison models in terms of recall. The benchmark comparisons included the traditional threshold method as well as a range of machine learning architectures, among them ResNet-18, LSTM, U-Net, and U-Net variants such as U-Net++ and Attention U-Net. In robustness experiments across different data partitions, IncResUnet achieved the highest F1-scores, 92.5 percent and 89.0 percent. In typical detection examples, the traditional method suffered from missed detections and false alarms, whereas IncResUnet accurately identified the onset and offset boundaries of plasma bubbles.

Perhaps the most important test was whether a model trained on one satellite could generalize to entirely different instruments, since a practical monitoring tool must not be tied to a single data source. When the model trained on FORMOSAT-1 data was applied directly to data from the C/NOFS satellite, it achieved an F1-score of 0.804 without any retraining. After a small amount of incremental training on the new data, the score improved to 0.859, with a recall rate as high as 98.8 percent, meaning nearly every true bubble was caught. Applied to COSMIC-2 data, the model successfully detected all 36 events, achieving a recall of 100 percent. The researchers found that the differences in generalization performance across satellite datasets primarily stem from the degree of data distribution shift, a finding that offers practical guidance for deploying the approach on future missions.

The significance of this work extends beyond a single algorithmic achievement. The study represents the first systematic application of deep learning methods to automatic ionospheric plasma bubble detection, validating the feasibility of learning density depletion waveform features directly from time-series data. By replacing rigid statistical thresholds with a learned model that adapts to multi-scale structure, the approach addresses the core adaptability deficiency of traditional methods, including their high miss rates during low solar activity. For space weather monitoring, that translates into more dependable warnings of scintillation interference affecting high-frequency communications and satellite systems. For ionospheric physics research, it provides a consistent, automated way to catalog bubble occurrences across years and instruments, enabling more robust statistical studies of how these depletions respond to solar cycle, seasonal, and local-time influences.

The research also illustrates a broader trend in the geosciences, in which architectures proven in medical imaging and signal processing are being repurposed for environmental monitoring problems with similar mathematical structure. The combination of Inception-style parallel kernels and residual connections proved particularly well suited to the multi-scale character of plasma bubbles, whose durations and spatial extents defy any single detection window. As satellite constellations such as COSMIC-2 continue to deliver vast streams of ionospheric measurements, models like IncResUnet point toward a future in which automated, learning-based systems provide continuous, high-fidelity surveillance of the equatorial ionosphere, protecting the communication and navigation infrastructure that modern society increasingly depends upon.

Subject of Research: Automatic detection of equatorial ionospheric plasma bubbles using a deep learning model applied to satellite ion density data

Article Title: IncResUnet: A model for automatic ionospheric plasma bubble detection

Article References: IncResUnet: A model for automatic ionospheric plasma bubble detection. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: ionosphere, plasma bubbles, deep learning, space weather, U-Net, FORMOSAT-1, C/NOFS, COSMIC-2, scintillation, ion density, satellite monitoring, Inception-residual modules

The central challenge of any direct-detection dark matter experiment is separation. If weakly interacting massive particles, or WIMPs, exist, they should occasionally strike a xenon nucleus, producing a tiny recoil that scintillation light and ionization electrons record. But gamma rays and beta particles from residual radioactivity in detector materials produce electronic recoils that look superficially similar. LZ’s main line of defense has been charge-to-light discrimination, which exploits the fact that nuclear recoils and electronic recoils produce different ratios of ionization charge to scintillation light. That method already rejects electronic recoil backgrounds with high efficiency, but every additional layer of rejection translates directly into better sensitivity to the vanishingly rare signals the experiment is built to find.

Pulse shape discrimination offers a second, independent handle. When energy is deposited in liquid xenon, scintillation light emerges from excited xenon dimers that can occupy either a singlet or a triplet molecular state. These states decay with different lifetimes, and crucially, nuclear recoils preferentially excite the singlet state, producing a flash with a steeper rising edge, while electronic recoils favor the triplet state and produce a longer tail. In liquid argon detectors, where the two lifetimes differ by more than two orders of magnitude, this effect is so pronounced that it serves as the dominant background rejection method. In liquid xenon, however, the singlet and triplet lifetimes are only about 3 and 26 nanoseconds respectively, a difference so small that most experiments have treated pulse shape information as marginal at best.

The LZ team attacked this limitation with an unusually rigorous timing analysis. The first hurdle was instrumental: differences in cable lengths, electronics, and operating voltages mean that each of the 494 photomultiplier tubes lining the detector responds with its own timing offset. Using an embedded LED calibration system, the collaboration measured these channel-to-channel offsets to better than one nanosecond. The deeper challenge was that individual photomultiplier waveforms often contain multiple overlapping photons, each contributing a pulse roughly 60 nanoseconds wide. Rather than treating the waveform as a single lump, the team developed an N-photon model that fits the waveform as a sum of up to three single-photoelectron templates, using Bayes’ theorem to decide how many photons each waveform actually contains, while accounting for the roughly 23 percent chance that a single photon produces two photoelectrons.

With photon arrival times reconstructed event by event, the collaboration compared timing distributions from two calibration sources: tritiated methane, which produces electronic recoils, and deuterium-deuterium neutrons, which produce nuclear recoils. As expected, the nuclear recoil spectra showed photons clustered earlier in time, while electronic recoil events exhibited longer tails. To convert this difference into a usable discriminator, the team defined a tail fraction, the ratio of photons counted in an optimized early time window to the total number of photons in the pulse. Counting individual photons rather than integrating raw current eliminated the smearing effects of the photomultiplier response and improved resolution. The boundaries of the early window were then optimized for each pulse size using a raster scan followed by a stochastic gradient descent algorithm, with the data repeatedly split into optimization and verification sets to guard against overfitting.

The performance exceeded expectations for a xenon detector. For scintillation pulses above 20 detected photons, the fraction of electronic recoil events leaking into the region occupied by half of the nuclear recoils ranged from 15 to 35 percent, reaching as low as 14 percent for the largest pulses studied. For a specific and troublesome background, the double electron capture of xenon-124, whose suppressed charge yield causes it to masquerade as a nuclear recoil under charge-to-light analysis, simulations showed that pulse shape discrimination reduces leakage to roughly 5 percent and outperforms charge-to-light discrimination above 95 detected photons. The team also derived a position-dependent correction to account for the fact that photons generated deeper in the detector spend longer in transit, extending the technique across the full detector volume rather than just the upper third where neutron calibration data were available.

To understand the physics underlying these results and to project performance, the collaboration tuned the photon timing model of the NEST simulation framework against the calibration data. The fit confirmed the singlet and triplet lifetimes previously measured in smaller experiments and extracted recombination time constants of essentially zero for nuclear recoils and about 3.7 nanoseconds for electronic recoils, consistent with the field-dependent recombination model built into NEST. The transit of photons through the detector, including reflections and scattering, was modeled with a combination of direct and exponential components whose parameters vary smoothly with depth. The agreement between simulation and data validated the emission parameters and provided an estimate of the best possible pulse shape performance, while also confirming that the depth correction applies equally to both interaction types.

The most consequential result comes from combining the two methods. The team constructed a two-factor discriminator, a linear combination of the charge-to-light output and the corrected tail fraction, with the weighting between them optimized as a function of pulse size. Because calibration data contained too few leaking events for a traditional counting-based evaluation, the collaboration instead quantified the separation between electronic and nuclear recoil distributions using a Z-score approach, previously employed by Xenon10 and other pulse shape studies. In the largest pulse bin, the separation improved from 3.65 to 3.83 standard deviations, cutting the corresponding false positive rate from 1.3 times ten to the minus four to 0.6 times ten to the minus four, a factor of two reduction. Applied to the WIMP search data collected between March 2023 and April 2024, the combined discriminator halved the number of events falling in the nuclear recoil band, from four to two, and all four events in the charge-to-light band were classified as electronic recoils by their pulse shapes alone.

The implications reach beyond LZ. The demonstration that photon timing carries exploitable information in a detector of this scale establishes a feasibility case for next-generation liquid xenon experiments, where every increment in background rejection buys sensitivity to ever fainter dark matter signals. The technique may prove even more powerful at lower drift fields, where enhanced recombination lengthens the timing differences between interaction types. And for backgrounds like xenon-124 double electron capture, which will only grow more relevant as exposures increase, pulse shape discrimination offers a complementary tool precisely in the energy range where other methods weaken. For now, the null result from the WIMP search data stands, but the machinery built to interpret those flickers of xenon light has become measurably sharper, and the hunt for the universe’s missing mass has gained a second pair of eyes.

Subject of Research: Pulse shape discrimination for electronic recoil background rejection in the LUX-ZEPLIN liquid xenon dark matter detector

Article Title: Pushing the limits of pulse shape discrimination in a large liquid xenon detector

Article References: LZ Collaboration, Akerib, D. S., Al Musalhi, A. K., Alder, F., Almquist, B. J., Amarasinghe, C. S., Ames, A., Anderson, T. J., Angelides, N., Araújo, H. M., Armstrong, J. E., Arthurs, M., Baker, A., Balashov, S., Bang, J., Bargemann, J. W., Barillier, E. E., Beattie, K., Bhatti, A., … Zhou, Y. (2026). Pushing the limits of pulse shape discrimination in a large liquid xenon detector. The European Physical Journal C, 86(9), Article 1060. https://doi.org/10.1140/epjc/s10052-026-16216-8

Image Credits: AI Generated

DOI: 10.1140/epjc/s10052-026-16216-8

Keywords: dark matter, LUX-ZEPLIN, liquid xenon, pulse shape discrimination, WIMPs, background rejection, scintillation, photon timing, NEST simulation, two-factor discrimination, xenon-124 double electron capture, direct detection