NeuraLunaDTNet: Neural-Augmented Lunar DTN

• The paper integrates a lightweight feedforward neural network, ACNet, with PRoPHET to predict next-hop contacts using spatio-temporal features.
• It employs dynamic graph construction from lunar mobility traces, achieving enhanced delivery ratios (0.71) and reduced latency (980 s) compared to standard DTNs.
• The approach ensures low computational overhead (~1 ms) while effectively addressing intermittent connectivity, severe propagation delays, and topology dynamics.

NeuraLunaDTNet is a neural network–augmented routing protocol designed for delay-tolerant lunar communication networks. It introduces a feedforward neural module to enhance PRoPHET-based store-and-forward logic by learning fine-grained spatio-temporal contact plans, thereby improving routing efficiency under lunar mobility, variable connectivity, and long propagation delays (Patel et al., 2024).

1. Motivating Context and Challenges

Lunar communication networks, comprising orbiters, surface rovers, and Earth ground stations, present distinct operational challenges:

• Severe Propagation Delays: One-way delays can reach several seconds due to the Moon-Earth distance.
• Intermittent Connectivity: Contact opportunities are highly variable as a result of dynamic orbital motions, lunar rotation–induced occlusions, and solar interference.
• No Stable End-to-End Path: Network topology is continuously disrupted, with no guarantee of an active route between any node pair at a given time.

Traditional Delay-Tolerant Networking (DTN) protocols, such as PRoPHET, use scalar delivery predictability ($P(A, B)$) based on encounter history and simple heuristic updates. However, these approaches do not exploit a priori trajectory or contextual patterns, nor do they adapt quickly to abrupt topology changes. Their forwarding logic is based solely on comparative probabilities, limiting responsiveness to underlying network dynamics.

2. Protocol Design and Feedforward Neural Augmentation

NeuraLunaDTNet augments the baseline PRoPHET protocol with a lightweight feedforward neural network (ACNet). The protocol preserves PRoPHET's custody-transfer semantics, summary-vector exchanges, and buffer management, but modifies the bundle forwarding decision as follows:

1. Feature Extraction: For each node encounter, the protocol forms a feature vector

$$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$

where each component is normalized to $[0,1]$.

1. Neural Prediction: The feature vector $x$ is input to ACNet—an MLP with four hidden layers (layer sizes: 64, 512, 128, 6 neurons with ReLU; output layer: 1 neuron, linear activation)—yielding a real-valued next-hop ID prediction $\hat{h}$.
2. Forwarding Rule: If the encountered node $B$ satisfies $|ID(B) - \hat{h}| < \epsilon$ (for a configurable $\epsilon$), the bundle is forwarded to $B$, even if PRoPHET's $P(A,D) \geq P(B,D)$. Otherwise, the standard PRoPHET rule is used.

This architecture allows context-sensitive routing decisions that exploit temporal and topological features omitted by standard scalar-based approaches.

3. Spatio-Temporal Contact Plan Learning

A core component of NeuraLunaDTNet is dynamic spatio-temporal network representation and learning:

• Graph Construction: The network builds epoch-wise dynamic graphs $$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$0 from external mobility traces (600 lunar-projected satellite orbits), where $$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$1 denotes nodes and $$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$2 links for each time epoch $$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$3. Edge weights $$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$4 correspond to empirical contact frequencies.
• Learning Target: ACNet is trained to approximate the conditional probability

$$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$5

The network thus captures latent spatio-temporal forwarding patterns embedded in the contact graph.

At runtime, the forwarding rule fuses PRoPHET's predictability $$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$6 and the neural estimate via a tolerance criterion:

• If $$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$7 (equivalently $$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$8) exceeds threshold, forward to $$x = (\text{epoch}, \text{srcID}, \text{dstID}, \text{curHopID})$$9.
• Otherwise, fall back to PRoPHET's traditional $[0,1]$0 test.

A mathematically equivalent fused predictability $[0,1]$1 is discussed, but implementation uses the simpler tolerance-based logic.

4. Training Methodology and Implementation

• Training Data Generation: Data is produced by simulating lunar DTN scenarios in the ONE environment, using a vanilla PRoPHET router with large-buffer configurations. Delivered bundles’ full hop-paths are logged, yielding supervised tuples $[0,1]$2.
• Loss and Optimization: ACNet is trained via Mean Squared Error (MSE):

$[0,1]$3

using Adam ($[0,1]$4), for 70,000 epochs with full-batch updates.

• Performance: MSE converges to $[0,1]$5, which translates to an average prediction error of $[0,1]$6 node ID.

ACNet’s codebase is integrated with PRoPHET’s Java implementation in ONE, leveraging the Deep Java Library for neural inference. The per-contact computational overhead is $[0,1]$7, which is negligible relative to multi-second round-trip delays.

5. Simulation-Based Evaluation

NeuraLunaDTNet is evaluated by simulation under the following conditions:

• Simulator: Opportunistic Network Environment (ONE)
• Nodes: 151 (rovers/orbiters, 1 ground station)
• Mobility and Traffic: Rovers/orbiters follow external-mobility traces (IDAO 2020, 7 days, binned into 168 hour-epochs); ground station is static. Bundle generation follows a Poisson process, all destined to Earth.
• Resources: Buffer limits (50 MB, 100 MB) and TTL 60 minutes.

The key performance metrics and comparative evaluation are summarized below (Table 1):

Protocol	Delivery Ratio	Avg. Latency (s)	Overhead Ratio
PRoPHET	0.62 ± 0.03	1,240 ± 85	3.0 ± 0.2
NeuraLunaDTNet	0.71 ± 0.02	980 ± 70	2.1 ± 0.15
Improvement (%)	+14.5%	−21.0%	−30.0%

Paired t-tests show significance at $[0,1]$8 for all evaluated metrics (Patel et al., 2024).

6. Discussion, Limitations, and Future Prospects

• Benefits: NeuraLunaDTNet learns and exploits spatio-temporal contact patterns beyond scalar encounter counts. It achieves enhanced delivery ratios and reduced latencies by selectively forwarding over links that historical and contextual patterns indicate as promising. The protocol maintains decentralization and low resource consumption, as the neural augmentation is lightweight and interfaces seamlessly with PRoPHET.
• Limitations: Node ID regression can be off by $[0,1]$9 due to the regression objective, resulting in occasional suboptimal forwarding. Current feature sets omit real-time link quality, buffer status, and event-driven disruptions (e.g., solar interruptions). Evaluation has focused on large buffer settings; adaptability under constrained buffers remains to be assessed.
• Future Directions:
  • Incorporation of real-time indicators (link quality, buffer state, contact duration) into the input space.
  • Exploration of Graph Neural Networks to encode $x$0 dynamically and capture multi-hop structural information.
  • Integration with resource-allocation frameworks (e.g., RAPID) for robust performance under bandwidth- and buffer-limited regimes.
  • Validation using hardware-in-the-loop testbeds (e.g., MODiToNeS), extension to multiple destinations (lunar bases, relays), and broader topological scenarios.

A plausible implication is that neural-augmented routing protocols may provide a template for other non-terrestrial DTNs where topology is governed by deterministic motion yet exhibits significant stochasticity. NeuraLunaDTNet represents a step towards AI-driven, context-adaptive interplanetary communication (Patel et al., 2024).