orbital-chaos-predictor

Orbital Chaos Predictor

Multi-modal deep learning for spacecraft orbit prediction incorporating solar wind perturbations via cross-attention fusion.

Live Dashboard | Dataset | Paper

Model Description

We train bidirectional LSTMs, Transformers, and a novel residual gated multi-modal architecture to predict spacecraft orbits 6 hours ahead using 24 hours of position history. The multi-modal model fuses orbit positions with solar wind measurements via cross-attention, improving predictions by 17% during geomagnetic storms.

Checkpoints

All checkpoints are available in both PyTorch (.pt) and safetensors (.safetensors) formats.

Usage

Loading with safetensors (recommended)

import torch
from safetensors.torch import load_file

# Load ISS LSTM checkpoint
state_dict = load_file("checkpoints/lstm_iss_6h_best.safetensors")

# Build model (match checkpoint architecture)
from your_model_file import OrbitLSTMDirect

model = OrbitLSTMDirect(
    input_dim=6, hidden_dim=128, num_layers=3,
    horizon=360, output_dim=3, dropout=0.0
)
model.load_state_dict(state_dict)

Loading with PyTorch (.pt files)

import torch

ckpt = torch.load(
    "checkpoints/lstm_iss_6h_best.pt",
    map_location="cpu",
    weights_only=False  # Required: checkpoints contain numpy metadata
)
model.load_state_dict(ckpt["model_state_dict"])

Input / Output Specification

Orbit-only models (LSTM, Transformer)

• Input: (batch, 1440, 6) -- 24-hour window at 1-minute resolution
  • Features: [x_gse, y_gse, z_gse, vx_gse, vy_gse, vz_gse] in GSE coordinates, z-score normalized
• Output: (batch, 360, 3) -- 6-hour prediction
  • Positions: [x_gse, y_gse, z_gse], normalized (denormalize with training stats)

Multi-modal model

• Orbit input: (batch, 1440, 6) -- same as above
• Solar wind input: (batch, 1440, 8) -- concurrent OMNI solar wind data
  • Features: [Bx_GSE, By_GSE, Bz_GSE, flow_speed, proton_density, Kp, Dst, AE], z-score normalized
  • Time-shifted by 45 minutes (L1-to-Earth propagation delay)
• Output: (batch, 360, 3) -- same as orbit-only

Training Data

• Orbit positions: NASA Satellite Situation Center (SSC) REST API
• Solar wind: NASA CDAWeb OMNI 1-minute high-resolution dataset
• Period: January 2023 -- December 2025
• Volume: ~4.8 million position records across 3 spacecraft
• Split: 70% train / 15% validation / 15% test (chronological)
• Resolution: 1-minute cadence
• Preprocessing: Velocity derived via finite differences, z-score normalization per spacecraft, gap detection at 10-minute threshold

Full dataset: datamatters24/orbital-chaos-nasa-ssc

Performance

6-Hour Prediction MAE (km)

Storm-Conditioned ISS (by Kp Index)

The multi-modal model improves 17% during geomagnetic storms (164 to 135 km), validating the hypothesis that solar wind data provides meaningful signal for LEO orbit prediction where thermospheric drag dominates.

Horizon Comparison (ISS)

Architecture Details

Residual Gated Multi-Modal Fusion

output = base_prediction + gate * perturbation

• Base prediction: Orbit-only BiLSTM (functionally identical to standalone)
• Perturbation: Cross-attention fusion of orbit and solar wind features, followed by attention-weighted summary and 3-layer MLP
• Gate: Sigmoid network controlling correction strength per timestep
• Guarantee: Gate initialized near zero; model starts as pure LSTM and can only improve

Two-phase training:

1. Phase 1 (20 epochs, LR=1e-3): Freeze solar branch, train orbit encoder only
2. Phase 2 (80 epochs, LR=1e-4): Unfreeze all, fine-tune with lower learning rate

Citation

@article{rubin2026orbital,
  author = {Rubin, Ted},
  title = {Multi-Modal Deep Learning for Spacecraft Orbit Prediction: Incorporating Solar Wind Perturbations via Cross-Attention Fusion},
  year = {2026},
  note = {arXiv preprint arXiv:XXXX.XXXXX}
}

License

MIT