eXplaNet in Networking & Exoplanet Science

• eXplaNet is a dual-use term that defines an XAI-driven reinforcement learning pipeline for networking and a conceptual integration framework in exoplanet science.
• In networking, it employs surrogate models and explainability techniques like SHAP and LIME to refine RL policies, improving performance metrics under dynamic conditions.
• In exoplanet research, it unifies diverse models, databases, and observational tools to advance climate modeling, biosignature analysis, and mission planning.

eXplaNet is a term used in two distinct research contexts. In networking, it denotes a pipeline grounded in explainable artificial intelligence for reinforcement-learning-based networking solutions, designed to use explanations to refine an agent’s policy (Murcia et al., 25 Sep 2025). In exoplanet research, the term is not consistently used as the name of a single instrument, project, or consortium; instead, it appears as a contextual label for a broader exploration and characterization ecosystem spanning catalogs, climate models, radio-emission simulators, host-star databases, retrieval frameworks, and observational benchmarks (Renaud et al., 2022, Bhatnagar et al., 28 Jan 2026). This suggests that eXplaNet is best understood as a domain-dependent designation whose meaning depends on whether the reference is to a specific XAI workflow or to an integrated exoplanet research infrastructure.

1. Terminological scope and domain usage

The explicit, formal use of the name occurs in the paper on RL-based networking, where eXplaNet is introduced as an XAI-driven pipeline for RL-based networking solutions (Murcia et al., 25 Sep 2025). Its stated purpose is to help networking researchers and practitioners gain deeper insights into the decision-making processes of RL-based solutions and to use those insights to improve the policy itself.

By contrast, the exoplanet-oriented material does not present a single formal entity named eXplaNet. One review explicitly states that it does not name a project, tool, or consortium called eXplaNet, while also arguing that the term is strongly implied as a conceptual exoplanet exploration or characterization network rather than a single instrument (Bhatnagar et al., 28 Jan 2026). In that exoplanet usage, the associated literature emphasizes modeling terrestrial exoplanet climates, detecting and interpreting biosignatures, atmospheric retrievals, and preparing future observations and missions.

The distinction is substantive. In networking, eXplaNet is a concrete methodological pipeline. In exoplanet science, the term functions more like an organizing concept for interoperable resources, models, and observational strategies. A plausible implication is that the same label has been used to denote two different forms of integration: integration of explainability into RL optimization, and integration of tools and data for exoplanet characterization.

2. eXplaNet as an XAI-driven pipeline for networking RL

In its networking usage, eXplaNet is presented as a modular, data-driven, topology-agnostic pipeline. The paper argues that RL is attractive for tasks such as routing, resource allocation, and SDN control because it can adapt to dynamic conditions and learn policies from experience, but that real deployment is limited because the resulting agents are often black boxes (Murcia et al., 25 Sep 2025). The stated motivation for explainability is therefore trust, debugging, manageability, and broader operational adoption.

The workflow is organized into four logical steps. The first is data creation, in which data are collected while the original RL system is running. In a Q-learning setting, the dataset may be derived from the agent’s Q-table and should include input features describing the network state, expected outputs such as Q-values or action values, and enough samples to represent behavior under different operating conditions. The same description emphasizes cleaning, normalization, categorical encoding, and removal of inconsistencies and outliers as preprocessing requirements.

The second step is the surrogate model. eXplaNet uses interpretable proxy models that imitate the RL agent rather than replace it. The stated surrogate options include decision trees, linear models, and ensemble models, and fidelity is checked to verify that the proxy faithfully reproduces the RL policy. Because networking conditions change over time, the paper also recommends periodically updating surrogates through incremental training, sliding windows, experience replay, retraining, or weighted learning.

The third step is the application of XAI techniques. The framework focuses on SHAP for global and local feature attribution based on Shapley values, LIME for local explanation around one prediction using a simple surrogate, PDP for the average effect of a feature on the output, and ICE for instance-level views of that effect. The paper argues that combining methods is preferable to relying on a single explanatory lens.

The fourth step is analysis and improvement. Explanation results are used to identify which features matter most and how they affect decision-making, after which the original RL model can be refined by removing irrelevant features, changing feature weights, redefining the reward function, changing exploration or exploitation behavior, or reducing dimensionality. The paper identifies this feedback step as the core innovation: explainability is not treated only as post-hoc interpretation, but as a mechanism for policy improvement.

3. RSIR, surrogate fidelity, and the transition to XRSIR

The principal case study applies eXplaNet to RSIR, a Q-learning-based SDN routing system (Murcia et al., 25 Sep 2025). RSIR is organized around the SDN data plane, control plane, management plane, and knowledge plane. The control plane periodically queries the data plane, the management plane computes metrics such as available bandwidth, delay, and packet loss, and the Q-learning agent in the knowledge plane uses those metrics to choose routes whose paths are then installed in switch flow tables.

The dataset for this case study is built from RSIR behavior on a GEANT topology with 23 nodes and 37 links, using Mininet 2.3.0 and Open vSwitch 2.13.8, and 16 distinct real intra-domain traffic matrices representing different load conditions, including peak hours. Each sample includes available bandwidth, delay, packet loss, route information, and the Q-value as target output. Preprocessing removes duplicates, imputes missing values, handles outliers using an interquartile-range threshold, and target-encodes origin and destination categories.

Several surrogate regressors are trained to approximate the reward or Q-value behavior of RSIR. The comparison reported in the paper is summarized below.

Surrogate model	$R^2$	MSE
XGBoost	0.91	225.39
Random Forest	0.87	641.78
Ridge regression	0.82	882.44
Linear regression	0.80	887.41

XGBoost is selected because it provides the best fidelity while remaining interpretable enough for feature-based analysis. The paper states that ensemble tree methods capture nonlinearities in the reward function better than linear models.

The RSIR reward is given as

$$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$

Here, $\hat{bwd}$ is normalized available bandwidth, $\hat{delay}$ is normalized delay, and $\hat{pkloss}$ is normalized packet loss. The metrics are normalized so that no single link state dominates learning, and the function is minimized to guide the agent toward optimal paths.

SHAP, PDP, and ICE are then used to interpret the surrogate. The reported finding is that available bandwidth has the strongest influence, high bandwidth is associated with lower Q-values, higher delay and packet loss generally increase Q-values, and packet loss is the least significant of the three features although still relevant. On that basis, the reward is revised to create XRSIR. The two best configurations reported are XRSIR$_1$ with $\beta_1 = 0.6$, $\beta_2 = 0.3$, $\beta_3 = 0.1$, and XRSIR$_2$ with $$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$0, $$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$1, $$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$2. The revised reward remains

$$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$3

and is likewise minimized to identify optimal paths.

The reported performance effects are that XRSIR selects shorter routes than RSIR in all traffic matrices, reduces delay and loss more effectively under high traffic, improves throughput in many traffic matrices, and that XRSIR$$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$4 achieves the best overall loss reduction, about 9% lower loss ratio. The paper also notes a qualification: in low-traffic conditions, RSIR can sometimes do slightly better on delay because balanced weights may allow shorter-hop paths while bandwidth is less of a bottleneck.

4. eXplaNet as a conceptual exoplanet characterization framework

In the exoplanet literature, eXplaNet is associated not with a single formally named system but with a network-like scientific program centered on habitability, biosignatures, retrievals, and mission preparation (Bhatnagar et al., 28 Jan 2026). The review literature presents a model hierarchy for terrestrial exoplanets comprising 1-D radiative-convective models, 1-D energy balance models, and 3-D global climate models.

The 3-D models highlighted include THOR and the Generic Planetary Climate Model. THOR is described as an open-source GCM built from scratch for exoplanets that solves the non-hydrostatic Euler equations on an icosahedral grid and avoids Earth-centric assumptions. Generic-PCM is an LMD-based model adapted for Mars, early Mars, Venus, Pluto, and exoplanets, and is identified as the only GCM in that discussion that can treat water vapor as a major atmospheric component. It now includes a dynamical slab ocean module incorporating sea ice and snow evolution, Ekman transport, horizontal diffusion, and convective adjustment.

The same framework presents habitability as requiring liquid water, an energy source, nutrients or building blocks, protection from harmful conditions, and long-term stability. The habitable zone is described as depending on stellar type, atmospheric composition, pressure, rotation state, ocean and ice dynamics, planetary mass, and interior evolution. The review discusses runaway greenhouse, moist greenhouse, maximum greenhouse, atmospheric collapse, and water condensation limits, and notes that slow rotators can keep habitable conditions closer to the star than fast rotators, that hydrogen-rich atmospheres can extend the outer habitable zone via collision-induced absorption, and that CO$$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$5 collapse on the nightside is crucial for tidally locked planets.

Biosignature analysis in this literature distinguishes gaseous biosignatures, surface biosignatures, and temporal biosignatures, while emphasizing false positives. The gases discussed include O$$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$6, O$$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$7, CH$$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$8, N$$R_{RSIR} = \beta_{1} \cdot \hat{bwd} + \beta_{2} \cdot \hat{delay} + \beta_{3} \cdot \hat{pkloss}.$$9O, H$\hat{bwd}$0S, SO$\hat{bwd}$1, OCS, DMS, CH$\hat{bwd}$2Cl, CH$\hat{bwd}$3Br, and CH$\hat{bwd}$4I. CH$\hat{bwd}$5Cl, CH$\hat{bwd}$6Br, and CH$\hat{bwd}$7I are discussed as capstone biosignatures: weaker than a primary biosignature like O$\hat{bwd}$8, but lower in false-positive risk and more convincing when combined with other indicators. Surface biosignatures include the Vegetation Red Edge, and temporal biosignatures include seasonal or periodic variations in CO$\hat{bwd}$9, O$\hat{delay}$0, and CH$\hat{delay}$1.

Atmospheric retrievals are presented as an enabling methodology consisting of a forward model and a parameter inference algorithm that searches prior space and returns a posterior probability distribution. The tools mentioned include petitRADTRANS, pyMultiNest or MultiNest, HELIOS-RETRIEVAL, and machine-learning methods for fast parameter searches. This retrieval-centered view is linked directly to mission preparation for JWST, VLT, ELT, RISTRETTO, the Habitable Worlds Observatory, and LIFE, the Large Interferometer For Exoplanets, which is described as a space-based mid-infrared nulling interferometer designed to detect and characterize temperate terrestrial planets through thermal emission rather than reflected light.

5. Infrastructure, catalogs, and data services in the exoplanet ecosystem

The exoplanet-oriented interpretation of eXplaNet is strongly infrastructural. EMAC, the Exoplanet Modeling and Analysis Center at NASA Goddard Space Flight Center, is described as a web- and mobile-accessible catalog, repository, and integration platform for exoplanet modeling and analysis resources (Renaud et al., 2022). It fills a long-standing gap by providing a single, searchable hub for tools, models, datasets, documentation, and tutorials relevant to exoplanet characteristics and environments. The paper states that EMAC has cataloged over 170 tools and resources across 11 scientific categories and many subcategories, including “Planetary Atmosphere Models,” “Planetary Interior Models,” “Planet Formation and Dynamics Tools,” and “Observatory/Instrument Models.”

EMAC defines resources broadly as any software, data visualization tool, or collection of model inputs or outputs related to exoplanet science. Hosted resource types include planetary interior models, visualization tools, code bases, web tools, tutorials, notebooks, documentation, and video demonstrations. Each appears as a resource block with metadata such as authors and summaries plus links to external repositories and supporting materials. A notable indexing feature is the unique ID format [2-Digit Year] [2-Digit Month]-[3-Digit Sequential Number], with permanent links of the form https://emac.gsfc.nasa.gov?cid=UniqueID. The platform also provides multi-category search and filtering, a “Related” linking function for tools with shared heritage or code bases, subscription services, an RSS feed, and a Twitter account. It can assist developers in building web-based applications and Jupyter notebook/MyBinder-based tutorials, and is described as a key project of SEEC, funded through NASA’s Internal Scientist Funding Model by the Planetary Science Division, Astrophysics Division, and Heliophysics Division.

A parallel but distinct infrastructure is ExoplANETS-A, a public, Virtual Observatory-compliant archive built by the EU Horizon-2020 ExoplANETS-A project (Morales-Calderón et al., 2024). Its stated aim is to provide a uniform, curated, and interoperable set of stellar and planetary properties for transiting exoplanet systems, with special emphasis on the host-star XUV environment. The database contains more than 700 columns for 205 planets around 114 host stars, selected as all transiting-exoplanet systems observed by HST or Spitzer up to June 2019. It includes basic identification and astrometry, stellar parameters, planetary properties, X-ray data, UV data, spectra, TESS light curves, and derived quantities such as preferred X-ray and UV values, XUV luminosities and fluxes, and SED-based stellar quantities.

The archive can be queried through a web interface, ConeSearch, and SAMP, and its products can be downloaded as CSV, VOtable, or JSON. ExoplANETS-A also developed exoVOSA, a tool based on VOSA that reads user-provided photometry tables, queries VO-accessible photometric catalogs, queries theoretical VO-compliant spectral model grids, computes synthetic photometry, determines the best-fitting model statistically, and infers bolometric luminosity, effective temperature, surface gravity, and metallicity. As a science demonstration, the database is used to recover a radius valley at 1.8 Earth radii, in agreement with previous studies, and to test photoevaporation in 14 multiplanet systems, with K2-3 identified as the one significant contradiction.

6. Specialized models and benchmark instruments within the exoplanet landscape

Several specialized tools and benchmark concepts populate the broader exoplanet landscape associated with eXplaNet-like integration. ExPRES, the Exoplanetary and Planetary Radio Emission Simulator, is a forward-modeling code for the time-frequency appearance of radio emissions generated by the Cyclotron Maser Instability in planetary magnetospheres, exoplanetary systems, and star-planet interaction systems (Louis et al., 2019). Its core task is not direct measurement but the prediction of source visibility to a distant observer once beaming and geometry are taken into account. The paper emphasizes that CMI emission is highly anisotropic and typically emitted in a narrow hollow cone, so a source may radiate continuously but be detected only when the beam intersects the observer’s line of sight.

ExPRES reproduces arc-like patterns in the time-frequency plane by computing whether radio waves emitted from user-defined source regions are visible at each time and frequency. The resonance condition is written as

$\hat{delay}$2

with the relativistic cyclotron frequency

$\hat{delay}$3

The paper states that radio arcs can constrain the location and dynamics of radio sources, the type of current system responsible for electron acceleration, the energy of the accelerated electrons, and, for exoplanetary systems, magnetic field strength, orbital period, rotation period, magnetic-field tilt, and magnetic-field offset. The code is described as useful for Jupiter, Saturn, Io–Jupiter, Europa–Jupiter, Ganymede–Jupiter, Cassini, Voyager, Galileo, Juno/Waves, BepiColombo, JUICE, exoplanets, magnetic white dwarf–planet systems, and white dwarf–brown dwarf systems.

ExoPlaSim is a fast, intermediate-complexity 3D climate model built by extending PlaSim for terrestrial exoplanets that are synchronously rotating, irradiated by non-solar stellar spectra, or characterized by non-Earth-like surface pressures (Paradise et al., 2021). The modifications include synchronous rotation support by fixing the Sun in the sky at a user-specified longitude, non-solar stellar spectra, surface-pressure dependence, optional vertical grids, and physics filters to suppress spectral ringing or Gibbs artifacts. The vertical-grid options are pseudolinear, pseudologarithmic, and hybrid, with pseudolinear recommended as the default. The model introduces spectrally dependent surface albedo and Rayleigh-scattering corrections for arbitrary stellar spectra and pressures, and implements Cesàro, exponential, and Lander–Hoskins physics filters, with exponential filtering at both transforms identified as a practical default. Validation is performed through model intercomparisons with ExoCAM, LMDG, ROCKE-3D, and benchmark suites such as Yang et al. 2019 and THAI. The principal significance of ExoPlaSim is speed: the paper reports about 30–60 seconds per model year at T21 on cluster nodes, about 15–16 minutes per year at T42, and the possibility of running thousands of models per week with a modest cluster allocation.

LOUPE, the Lunar Observatory for Unresolved Polarimetry of Earth, occupies a different but complementary niche as an observational benchmark for exoplanet characterization (Karalidi et al., 2012). It is a small, robust spectropolarimeter proposed for the lunar surface to observe Earth as an unresolved exoplanet. The scientific rationale is that future direct observations of rocky exoplanets will generally be single-pixel, disk-integrated signals, so retrieval algorithms require benchmark data from Earth observed under exoplanet-like conditions. The paper describes the full Stokes flux vector as

$\hat{delay}$4

and the degree of polarization as

$\hat{delay}$5

with

$\hat{delay}$6

LOUPE is designed to operate at visible wavelengths from 400–800 nm, with about 20 nm spectral resolution for polarimetry, hourly cadence, and at least a month of operation to cover a full range of phase angles. The benchmark target is a time-resolved, disk-integrated dataset of Earth’s reflected flux and polarization as functions of wavelength, time of day, phase angle, and potentially season.

7. Conceptual significance, boundaries, and recurrent misconceptions

A recurring misconception is that eXplaNet refers to one unified exoplanet instrument or consortium. The literature on terrestrial exoplanet climates and biosignatures explicitly states otherwise: the name is not directly attached there to a project, tool, or consortium, but is instead most plausibly interpreted as an exoplanet exploration or characterization concept or network (Bhatnagar et al., 28 Jan 2026). Another misconception is that platforms such as EMAC, ExoplANETS-A, ExPRES, ExoPlaSim, and LOUPE are interchangeable. They are not. EMAC is a catalog, repository, and integration platform; ExoplANETS-A is a VO database and host-star environment archive; ExPRES is a radio-emission forward model; ExoPlaSim is a 3D climate model; and LOUPE is a benchmark spectropolarimetric observing concept (Renaud et al., 2022).

In the networking literature, a parallel misconception is that explainability is merely post-hoc reporting. The defining claim of eXplaNet is precisely the opposite: explanations are treated as feedback that can refine the reward function or policy, as demonstrated by the transition from RSIR to XRSIR (Murcia et al., 25 Sep 2025). The paper therefore locates explainability inside the optimization loop rather than outside it.

Across both domains, the significance of the term lies in integration. In exoplanet science, the associated resources are described as improving accessibility, comparison, validation, and dissemination, and as linking host-star physics, climate modeling, biosignature theory, retrievals, and mission preparation. In networking, the corresponding integration links surrogate modeling, SHAP, LIME, PDP, ICE, and RL reward design into a single operational workflow. A plausible implication is that eXplaNet, despite its domain-specific meanings, denotes an infrastructural strategy: turning distributed technical components into a coherent system for inference, interpretation, and decision support.