datamatters24/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.

Architecture	Parameters	ISS MAE (6h)
BiLSTM	660K	125 km
Transformer	690K	282 km
Residual Gated Multi-Modal	1.9M	163 km (135 km during storms)

Checkpoints

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

Checkpoint	Architecture	Spacecraft	Orbit Type
lstm_iss_6h_best	BiLSTM	ISS	LEO (~408 km)
lstm_dscovr_6h_best	BiLSTM	DSCOVR	L1 (~1.5M km)
lstm_mms1_6h_best	BiLSTM	MMS-1	HEO (magnetosphere)
transformer_iss_6h_best	Transformer	ISS	LEO
transformer_dscovr_6h_best	Transformer	DSCOVR	L1
transformer_mms1_6h_best	Transformer	MMS-1	HEO
multimodal_iss_6h_best	Residual Gated Multi-Modal	ISS	LEO
multimodal_dscovr_6h_best	Residual Gated Multi-Modal	DSCOVR	L1
multimodal_mms1_6h_best	Residual Gated Multi-Modal	MMS-1	HEO

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)

Model	ISS (LEO)	DSCOVR (L1)	MMS-1 (HEO)
LSTM	125	12,797	18,832
Transformer	282	13,517	19,296
Multi-Modal	163	25,059	19,277
Ensemble (LSTM+MM)	126	--	--
SGP4 Keplerian baseline	575	--	88

Storm-Conditioned ISS (by Kp Index)

Model	All (n=657)	Quiet Kp<=3 (n=489)	Active Kp 4-5 (n=71)	Storm Kp>=6 (n=7)
LSTM	125 km	126 km	122 km	113 km
Multi-Modal	163 km	164 km	158 km	135 km
Ensemble	126 km	126 km	124 km	111 km

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)

Model	1 hour	3 hours	6 hours
LSTM	54.5 km	82.1 km	125 km
Multi-Modal	171.5 km	176.2 km	163 km

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