At the heart of this breakthrough is an innovative technique known as radio interferometric multiplexed spectroscopy (RIMS), a method meticulously developed to probe the subtle flux density variations in multiple celestial directions simultaneously by leveraging interferometric datasets. This sophisticated approach breaks new ground by enabling the synthesis of nearly 200,000 dynamic spectra derived from an extensive archive captured by the Low-Frequency Array (LOFAR), one of the world’s flagship radio telescope networks specializing in low-frequency observations. The dataset spans roughly 1.4 years and contains petabytes of information, reflecting an unprecedented commitment to uncovering faint and transient astrophysical signals.

The researchers meticulously applied RIMS to 68 previously identified target systems that showed promise through prior detections in LOFAR’s circularly polarized imaging studies. Remarkably, about 25% of these targets exhibited statistically significant flux variability on timescales of just a few hours. This variability, particularly at such low radio frequencies, underscores dynamic processes potentially connected to the magnetospheric environments of these stars and their orbiting exoplanets. Notably, among these findings were eight novel, weak burst events predominantly associated with low-mass stars, displaying characteristic timescales of variability between 30 minutes and one hour.

These new detections challenge conventional interpretations and open fresh avenues in the study of stellar and planetary magnetospheres. While intrinsic stellar activity remains a viable explanation for the observed bursts, the possibility that these emissions originate from interactions between stars and closely orbiting exoplanets adds profound excitement and complexity to the analysis. Star–planet magnetic interactions are thought to generate distinct radio signatures, reminiscent of Jupiter’s interaction with its moon Io but on vastly larger and more energetic scales. Confirming such phenomena would provide an unprecedented window into daytime weather in distant planetary systems and how magnetic fields influence planetary atmospheres and habitability.

The implications of this research extend deeply into the future of radio astronomy and exoplanet science. The capabilities demonstrated through RIMS and LOFAR’s extensive observational campaigns pave the way toward more refined and sensitive studies with next-generation telescopes such as the Square Kilometre Array (SKA). With its immense collecting area and enhanced frequency range, the SKA promises to surveil the radio sky with unparalleled detail, potentially identifying hundreds or thousands of such planetary magnetospheric phenomena in stars across the Milky Way.

Low-frequency radio emissions have been extensively studied within our Solar System, serving as powerful diagnostic tools for solar flare activity and magnetospheric dynamics around Earth, Jupiter, and other planets. Bringing this diagnostic outside our cosmic vicinity offers a transformative perspective on star-planet coupling, magnetic field strengths, and particle acceleration mechanisms in environments vastly different from our own. Detecting circular polarization is especially crucial, as it confirms the coherent nature of the emission processes, often linked to cyclotron or synchrotron maser emissions—key signatures of magnetic interactions.

The deployment of RIMS represents a methodological leap that harnesses the multiplexing power of radio interferometry to map variability over many sources at once, instead of the traditional approach that relies on isolated target observations. This efficiency not only expands the observable sample size but also enhances sensitivity by integrating dynamic spectral data, allowing for the detection of weaker and more transient bursts that previously evaded capture. The resulting enriched dataset delivers finer temporal, spectral, and spatial resolution, critical for disentangling the signatures of star-planet magnetic dialogues from intrinsic stellar noise and other astrophysical backgrounds.

In practice, the LOFAR observations centered on around 150 MHz provide a sweet spot in frequency, where the ionospheric cutoff imposes limits on lower boundaries but still allows penetration of the key radio frequencies that correspond to magnetospheric emissions in exoplanetary systems. The long baseline of LOFAR’s interferometers intensifies angular resolution capabilities, crucial for isolating point sources and attributing bursts accurately to specific stellar targets. This rigorous spatial mapping complements the temporal resolution gained through spectroscopy, enabling researchers to correlate burst events with orbital phases, stellar activity cycles, and other astrophysical variables.

Critically, the identification of bursts with variability timescales on the order of 0.5 to 1 hour suggests fast-evolving magnetospheric processes influenced either by stellar rotation, planet-induced magnetic reconnection events, or fluctuating stellar wind conditions. These timescales are consistent with expected dynamical intervals in contexts such as magnetic star-planet interactions, where relative motion between a planet’s magnetosphere and the host star’s magnetic field lines can generate periodic bursts of radio emission. Such detections may also shed light on exoplanet magnetic field strengths, essential parameters for modeling their atmospheric retention and potential habitability conditions.

While the evidence draws a tantalizing portrait that some of these detected radio bursts could be consequences of star-planet magnetic coupling, the intrinsic stellar origin remains an alternative explanation that must be carefully vetted through further study. Low-mass stars are well-known for their flaring activity and magnetic complexity, which can also produce circularly polarized radio bursts unrelated to orbiting planets. Disentangling these potential sources demands complementary data from multi-wavelength observations, long-term monitoring, and detailed magnetohydrodynamic modeling to test the physical plausibility of both scenarios.

The advancement presented in this research also reinforces the pivotal role of polarimetry in astrophysical radio studies. Circular polarization measurements act as robust fingerprints indicating coherent emission processes linked to magnetic phenomena. Such measurements provide a powerful discriminator against incoherent synchrotron emission typically associated with broader astrophysical settings, enabling astronomers to isolate the unique radiation signatures of magnetospherically-driven bursts. This precision is vital as researchers strive to build a comprehensive inventory of star-planet interaction systems throughout our galaxy.

Looking ahead, the demonstrated success of RIMS and the enormous LOFAR dataset signal that the era of radio detection of exoplanetary magnetospheric activity has truly dawned. These findings bolster the case for incorporating low-frequency, high time-resolution radio monitoring as a standard tool in exoplanet characterization, alongside traditional optical and infrared approaches. The magnetic environment of exoplanets profoundly influences atmospheric escape, radiation shielding, and planetary evolution—factors fundamental to assessing whether distant worlds might sustain life or even host technological civilizations.

The harnessing of multi-petabyte data volumes and advances in radio interferometric techniques underscore the increasing intersection between astronomy and big data science. Handling, processing, and interpreting petascale datasets from arrays like LOFAR require not only sophisticated algorithms but also computational infrastructures capable of sustained high-throughput analysis. RIMS exemplifies how innovative processing frameworks can unlock the hidden dynamism within vast archives of visibility data, transforming static images into time-resolved spectral cubes brimming with new discoveries.

While the current survey focuses on a relatively small sample of 68 target systems, the scalability of these methods invites application to broader stellar populations, including stars with different spectral types, ages, and planetary architectures. Expanding this search will statistically refine our understanding of how common star-planet magnetospheric interactions are, how they evolve with stellar age and activity, and how they influence exoplanetary system environments over billion-year timescales. Moreover, identifying targets exhibiting recurrent burst patterns will enable targeted, multi-wavelength follow-up campaigns, deepening insight into underlying physical mechanisms.

In sum, this pioneering research by Tasse, Zarka, Hardcastle, and colleagues pioneers a new window onto the dynamic radio universe of distant stars and their planetary companions. By revealing circularly polarized radio bursts produced potentially by star-planet interactions, they have charted an exciting frontier for radio astronomy and exoplanet science. As next-generation facilities like the Square Kilometre Array come online, the prospects for unraveling the magnetic tapestries connecting stars and their planets—and perhaps unveiling new aspects of exoplanet habitability—have never been brighter.

Subject of Research:
The study explores the detection and analysis of low-frequency, circularly polarized radio bursts from stellar and exoplanetary systems, focusing on star–planet magnetic interactions and intrinsic stellar activity signatures.

Article Title:
The detection of circularly polarized radio bursts from stellar and exoplanetary systems

Article References:
Tasse, C., Zarka, P., Hardcastle, M.J. et al. The detection of circularly polarized radio bursts from stellar and exoplanetary systems. Nat Astron (2026). https://doi.org/10.1038/s41550-025-02757-7

DOI:
https://doi.org/10.1038/s41550-025-02757-7

The signal-to-noise problem is not merely an inconvenience; it is a fundamental bottleneck that has historically limited our ability to explore the universe’s most elusive secrets. Whether it’s detecting the faint imprints of dark matter particles, deciphering the gravitational tug of distant black holes, or uncovering the precise mechanisms behind particle interactions at accelerators like the Large Hadron Collider, extracting these faint signals from overwhelming background noise is akin to finding a single needle in an infinitely expanding haystack. This challenge intensifies with every leap forward in experimental sensitivity. As detectors become more precise, they also gather more data, and consequently, more noise. Traditional signal processing techniques, while powerful, often struggle to adapt to the dynamic and often unpredictable nature of this ever-growing background, leading to potential biases in analysis and the obfuscation of truly significant discoveries. The need for a more robust, adaptable, and intelligent method for signal extraction has never been more apparent, and stochastic automatic differentiation appears to be that very innovation.

At the heart of this revolution lies stochastic automatic differentiation, a fusion of machine learning’s advanced computational power and the rigorous mathematical underpinnings of calculus. Automatic differentiation (AD) itself is a powerful technique for efficiently and accurately computing derivatives of functions, which are essential for optimization and sensitivity analysis in many scientific fields. However, when dealing with the inherent uncertainty and randomness present in experimental data – a characteristic often referred to as “stochasticity” – standard AD can falter. This is where the “stochastic” element comes into play. By integrating sophisticated probabilistic models and adaptive learning algorithms, this new approach can effectively navigate and learn from noisy datasets, progressively refining its ability to distinguish true signals from spurious fluctuations with remarkable precision. It’s not just about filtering; it’s about intelligently learning the underlying patterns of noise and signal simultaneously.

The brilliance of this method lies in its ability to adapt and learn. Unlike static algorithms that are pre-programmed with fixed parameters, stochastic automatic differentiation systems are designed to evolve with the data. As they process more information, they refine their internal models of both the expected signals and the characteristics of the background noise. This iterative learning process allows them to become increasingly adept at identifying subtle patterns that might be missed by traditional methods, or even by human analysts. This dynamic adaptation is crucial in fields like high-energy physics, where the nature of background noise can change depending on experimental conditions or unforeseen environmental factors, demanding an analytical tool that can keep pace with these variations. The ability to dynamically adjust is what makes this approach a true game-changer in data analysis.

Consider the realm of gravitational-wave astronomy. Detecting the infinitesimally small ripples in spacetime caused by colliding black holes or neutron stars involves extracting signals buried under immense seismic, thermal, and instrumental noise. Even the most advanced detectors are susceptible to these disturbances, making the identification of genuine gravitational-wave events a monumental task. Stochastic automatic differentiation can be employed to build sophisticated probabilistic models of these detector noises, allowing researchers to more accurately predict and subtract them, thereby enhancing the sensitivity of gravitational-wave observatories and potentially unlocking access to previously undetectable cosmic events. This could mean hearing the faint murmurs of the universe’s earliest moments or observing the mergers of less massive, but perhaps more exotic, astrophysical objects.

Similarly, in particle physics, experiments at facilities like CERN’s Large Hadron Collider generate petabytes of data from trillions of particle collisions. Identifying rare particle decays or the signatures of new, undiscovered particles requires sifting through an avalanche of background events that mimic the desired signal. This new AI-driven approach can learn the intricate patterns of these background processes, allowing physicists to isolate the statistically significant deviations that point towards new physics. It’s like having an AI trained to spot the unique fingerprint of a rare particle amidst the general chaos of a particle accelerator, a feat that would be nearly impossible with older, less nuanced analytical tools. The implications for discovering new fundamental particles or understanding the forces that govern them are profound.

The scientific paper introducing this methodology highlights its potential to significantly improve the accuracy and efficiency of data analysis pipelines across various physics disciplines. The authors demonstrate how stochastic automatic differentiation can outperform conventional techniques in simulated scenarios designed to mimic real-world experimental conditions, showcasing a tangible uplift in signal detection capabilities. This is not just theoretical prowess; it is a practical demonstration of enhanced scientific observation capabilities. The researchers meticulously validated their approach against various noise models, proving its robustness and adaptability, which are critical factors for adoption in the rigorous world of experimental physics where every anomaly must be scrutinized with the utmost care. Their work provides a clear roadmap for implementing this technology.

Furthermore, the implications extend beyond just detection; they touch upon the very precision of our measurements. By more accurately understanding and accounting for noise, scientists can derive more precise values for fundamental constants, particle masses, and interaction strengths. This increased precision is vital for testing theoretical models like the Standard Model of particle physics and searching for deviations that might hint at new physics beyond our current understanding. A tiny adjustment in the measured value of a fundamental constant, achieved through superior noise reduction, could unravel decades of theoretical work or open entirely new avenues of scientific inquiry. The pursuit of ever-greater precision is the bedrock of progress in fundamental physics.

The “viral” potential of this research stems from its broad applicability and its promise of accelerating discovery. In a scientific landscape increasingly reliant on sophisticated data analysis, a tool that can more effectively extract meaningful insights from noisy data is invaluable. It can democratize access to advanced analytical capabilities, potentially empowering researchers at institutions with fewer resources to achieve comparable breakthroughs. The ability to overcome data limitations is a powerful equalizer in the global scientific enterprise, fostering collaboration and accelerating the pace of innovation across the board. This development isn’t confined to a single subfield; its potential impact is felt across the entire spectrum of physics research.

The computational efficiency of stochastic automatic differentiation is another key factor driving its potential for widespread adoption. While machine learning models can be computationally intensive, the underlying principles of AD are inherently efficient. When combined with modern hardware accelerators like GPUs and TPUs, these methods can process vast datasets in a fraction of the time previously required, allowing scientists to iterate more rapidly on their analyses and explore a wider range of hypotheses. This acceleration of the research cycle is critical for staying at the forefront of scientific exploration and for responding quickly to new experimental results or theoretical insights that emerge from the field. The speed of discovery is directly linked to the speed of analysis.

The authors of the pivotal paper emphasize that this is not a replacement for fundamental physics understanding but rather a powerful augmentation. The AI is a tool to help flesh out the details, not replace the core theoretical framework. It empowers physicists to ask more detailed questions of their data, to probe phenomena at finer resolutions, and to explore parameter spaces that were previously inaccessible due to data limitations. This symbiotic relationship between theoretical insight and computational power is the engine of modern scientific progress, and stochastic automatic differentiation is the latest, most potent iteration of that driving force, enabling a deeper interrogation of the universe’s secrets, pushing the boundaries of what is knowable.

Looking ahead, the integration of stochastic automatic differentiation into mainstream scientific analysis workflows is likely to lead to a surge in discoveries across numerous fields. From cosmology and astrophysics to particle physics and condensed matter, the ability to more effectively disentangle signals from noise will undoubtedly unlock new avenues of investigation. Scientists are already exploring its application in areas such as neutrino detection, where signals are notoriously difficult to isolate, and in the search for gravitational waves from the very early universe, where signals are expected to be exceedingly faint. The universe is speaking, and this new technology is providing us with a vastly improved ability to comprehend its language.

The development heralds a new era for experimental science, one where the limitations of data processing are progressively overcome by intelligent algorithms. The ability to train AI on specific noise characteristics of an experiment, and for that AI to then continuously refine its understanding, represents a significant leap forward. It means that as experiments evolve and their data characteristics change, the analysis tools can adapt in real-time, ensuring that no subtle whisper of a new phenomenon is lost in the ever-present clamor of the universe. This adaptability is the hallmark of truly intelligent scientific instrumentation and analysis.

The core innovation lies in creating functions that can learn the statistical properties of both signal and noise simultaneously, allowing for a more holistic and accurate reconstruction of reality from imperfect measurements. This contrasts with older methods that might attempt to model and subtract noise in a separate, often less accurate step. By learning them in tandem, the AI can identify correlations and dependencies between signal and noise that a sequential approach might miss entirely. This integrated learning approach allows for a more nuanced and ultimately more accurate interpretation of the data, a crucial step in understanding the fundamental nature of our universe.

The journey from raw data to fundamental insight is often fraught with peril, primarily in the form of overwhelming noise that obscures the truth. Stochastic automatic differentiation, as pioneered by Catumba and Ramos, acts as a supremely sophisticated sieve, incredibly efficient at separating the wheat from the chaff, the signal from the noise, and the truth from the statistical artifacts. This advancement signifies not merely an incremental improvement but a transformative leap in our capability to extract knowledge from the universe’s data streams, promising a future where the faint whispers of cosmic phenomena are no longer lost, but clearly heard and understood, driving unparalleled progress in our understanding of everything from the infinitesimally small to the unimaginably vast cosmic structures.

Subject of Research: The development and application of stochastic automatic differentiation for enhanced signal-to-noise ratio in scientific data analysis, particularly within physics experiments.

Article Title: Stochastic automatic differentiation and the signal to noise problem.

Article References:

Catumba, G., Ramos, A. Stochastic automatic differentiation and the signal to noise problem.
Eur. Phys. J. C 85, 1037 (2025). https://doi.org/10.1140/epjc/s10052-025-14690-0

Image Credits: AI Generated

DOI: https://doi.org/10.1140/epjc/s10052-025-14690-0

Keywords: Stochastic Automatic Differentiation, Signal-to-Noise Ratio, Data Analysis, Machine Learning, Physics, Scientific Discovery, Particle Physics, Gravitational Waves, Artificial Intelligence.

The scientific community has long been engaged in a relentless pursuit to accurately measure the Hubble constant, a value that sits at the very foundation of our cosmological narrative. Discrepancies between measurements derived from different cosmological probes have led to what is now dubbed the “Hubble Tension,” a persistent anomaly that suggests our current standard model of cosmology might be incomplete or that there are as-yet-undiscovered physical phenomena at play. This new research, however, offers a novel pathway to tackle this cosmic conundrum. By employing artificial neural networks, specifically designed to process and learn from complex, high-dimensional data, the researchers are exploring an entirely different methodology for extracting these crucial cosmological parameters. This algorithmic approach could potentially offer a more objective and efficient way to navigate the intricate web of observational data, bypassing some of the inherent complexities and assumptions that have historically complicated traditional parameter estimation techniques.

At the core of this pioneering work lies the sophisticated application of artificial neural networks. These digital architectures, loosely inspired by the human brain’s intricate network of neurons, are capable of learning complex relationships and patterns directly from data. In this context, the neural networks are trained on simulated cosmic data, known as “mock H(z)” – essentially, synthetic datasets representing the relationship between the expansion rate of the universe and redshift, a measure of how much light from distant objects has been stretched due to the universe’s expansion. By learning from these controlled environments, the AI models gain the ability to infer cosmological parameters from real observational data, mirroring the process astronomers undertake but with a computational engine capable of processing information at an unprecedented scale and speed, potentially uncovering subtle correlations missed by conventional methods.

The researchers meticulously employed a covariance matrix in their methodology, a statistical tool that quantifies the interdependencies between different variables. In cosmology, measurements of various cosmic quantities are rarely independent; they often exhibit correlations due to shared systematic uncertainties or inherent physical relationships. Incorporating the covariance matrix into the neural network’s learning process is crucial for ensuring that the AI’s estimations are not just accurate but also statistically robust, properly accounting for these interdependencies. This rigorous statistical grounding is essential for any scientific endeavor aiming to draw definitive conclusions about the universe, especially when dealing with parameters as fundamental and as hotly debated as the Hubble constant, thereby lending significant weight and reliability to the AI’s deductions.

The use of “mock H(z)” data serves as a crucial validation step for the artificial intelligence models. By training on data generated from known cosmological parameters, the researchers can effectively “test” the AI’s ability to recover these parameters. This controlled environment allows for a precise evaluation of the neural network’s performance, identifying any biases or limitations before applying it to the complexities of real-world astronomical observations. This simulated testing phase is akin to a pilot training on a flight simulator before taking the controls of a real aircraft—it ensures the system is robust, reliable, and capable of handling the demanding task ahead, offering a high degree of confidence in its future real-world applications.

The implications of successfully employing artificial intelligence in cosmological parameter estimation are far-reaching. Beyond potentially resolving the Hubble Tension, these AI-driven techniques could significantly accelerate the analysis of upcoming, massive astronomical surveys, such as the Square Kilometer Array (SKA) or the Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST). These future missions will generate petabytes of data, far exceeding the capacity of traditional analysis methods to process efficiently. AI offers a scalable solution, enabling scientists to extract valuable cosmological information from these data deluge in a timely manner, pushing the boundaries of our cosmic exploration further and faster than ever imagined.

Furthermore, the adaptability of neural networks allows them to be trained on a wide variety of cosmological probes, including Type Ia supernovae, baryon acoustic oscillations (BAO), and cosmic microwave background (CMB) radiation. Each of these probes provides a unique window into the universe’s expansion history and fundamental parameters. By training AI models on diverse datasets, researchers can develop a more comprehensive and robust understanding of cosmology, potentially identifying synergies between different observational methods or even revealing inconsistencies that hint at new physics beyond our current theoretical frameworks, truly unlocking a multipronged approach to cosmic discovery.

This research represents a significant shift in how scientific discovery is pursued. Instead of solely relying on human intuition and analytical frameworks built over decades, scientists are now actively collaborating with intelligent algorithms to probe the deepest mysteries of the universe. This symbiotic relationship between human expertise and artificial intelligence promises to unlock new avenues of inquiry, allowing researchers to explore parameter spaces and complex datasets that would be intractable for human analysis alone. It signifies a powerful evolution in the scientific paradigm, where computation is not just a tool but a partner in scientific exploration, enabling unprecedented levels of insight.

The paper’s findings, while still in their early stages of peer review and further validation, suggest that artificial neural networks can indeed offer competitive, if not superior, accuracies in estimating cosmological parameters compared to traditional methods. The ability of these networks to learn complex, non-linear relationships within the data is particularly beneficial in cosmology, where the interplay of various cosmic constituents and their expansionary effects can be highly intricate. This computational advantage could lead to more precise measurements of fundamental quantities, thereby refining our cosmic inventory and deepening our comprehension of the universe’s evolution.

One of the most exciting prospects of this AI-driven approach is its potential to explore alternative cosmological models beyond the current Lambda-CDM paradigm. The Lambda-CDM model, while highly successful, is known to face certain challenges, including the aforementioned Hubble Tension. Artificial neural networks, unburdened by preconceived theoretical biases, might be able to identify patterns in the data that suggest deviations from Lambda-CDM or even point towards entirely new cosmological frameworks, offering a purely data-driven avenue for theoretical innovation, pushing the boundaries of our understanding into uncharted territories.

The visual representation provided with the research, an AI-generated image, itself symbolizes this marriage of technology and cosmic inquiry. It is a testament to the fact that even the very imagery used to convey these complex scientific concepts is now being augmented by artificial intelligence, hinting at a future where AI plays a role in all facets of scientific endeavor, from data analysis to visualization and conceptualization, blurring the lines between the digital and the empirical. The image serves as a potent symbol of AI’s expanding influence within the scientific landscape, illustrating the abstract concepts with a clarity that resonates visually.

The success of this research could pave the way for dedicated AI-powered cosmological observatories or analysis pipelines, specifically designed to continuously refine our understanding of the universe. Such systems could autonomously identify interesting cosmic phenomena, flag anomalies in observational data, and even propose new avenues of scientific investigation based on emerging patterns. This would mark a significant acceleration in the pace of cosmic discovery, transforming astronomy into a more dynamic and proactive field of scientific research, where insights are generated with unprecedented speed and efficiency.

The specific architecture and training methodology of the neural networks employed in this study are of paramount importance. Understanding how these networks are designed, what features they prioritize, and how they are trained on the mock data will be crucial for their widespread adoption and for building trust in their results. Future work will undoubtedly focus on further optimizing these AI models, exploring different network architectures, and developing robust techniques for interpreting their internal workings, ensuring transparency and interpretability in the process of cosmic inference.

In conclusion, this study by Chen, Zhang, He, and their collaborators is more than just an incremental step forward; it is a bold leap into a new paradigm of cosmological research. By harnessing the power of artificial intelligence, scientists are equipping themselves with tools to tackle humanity’s most profound questions about the origin, evolution, and ultimate fate of the universe. The journey to unraveling the cosmic enigma is far from over, but with AI as a powerful new ally, our understanding of the universe is poised to expand in ways we are only beginning to comprehend, promising a future filled with extraordinary revelations.

Subject of Research: Estimating cosmological parameters and reconstructing the Hubble constant using artificial neural networks.

Article Title: Estimating cosmological parameters and reconstructing Hubble constant with artificial neural networks: a test with covariance matrix and mock H(z).

Article References: Chen, Jf., Zhang, TJ., He, P. et al. Estimating cosmological parameters and reconstructing Hubble constant with artificial neural networks: a test with covariance matrix and mock H(z).
Eur. Phys. J. C 85, 1005 (2025). https://doi.org/10.1140/epjc/s10052-025-14714-9

Image Credits: AI Generated

DOI: 10.1140/epjc/s10052-025-14714-9

Keywords**: Cosmology, Artificial Neural Networks, Hubble Constant, Parameter Estimation, Mock Data, Covariance Matrix, Hubble Tension, Machine Learning, Astrophysics, Scientific Discovery.