Fine tuned LLM on AWS Bedrock + physics simulator for recommending thermal mitigation strategies in deep space photonic instruments

Physics simulation · Bedrock fine tuning · XGBoost classification · Streamlit demo

Photonic Integrated Circuits (PICs) are the backbone of next generation space probe instruments that operate in deep space — spectrometers, laser communication terminals, waveguide sensor arrays, and photonic signal processors. But space is brutal:

• 🌡️ Spectral drift — temperature swings shift refractive indices, pushing resonant wavelengths off-target and corrupting measurements
• 📐 Waveguide misalignment — differential thermal expansion between chip layers destroys optical coupling, killing signal throughput
• 💥 Mechanical cracking — repeated thermal cycling fatigues bonding interfaces and dielectric layers until catastrophic failure

A spectrometer on a Jovian probe faces 180 K temperature swings. An optical link in the outer solar system endures 240 K. The wrong mitigation strategy means mission failure.

This project combines deterministic physics with AI-driven recommendations to prescribe optimal thermal mitigation strategies for any instrument-material-environment combination:

┌─────────────────────────────────────────────────────┐
│              Streamlit Interactive App               │
│         (Physics Simulator  ·  AI Advisor)          │
├──────────────────────┬──────────────────────────────┤
│   ThermalDriftSim    │    BedrockInferenceClient    │
│   Δn = dn/dT × ΔT   │    Fine-tuned Titan Express  │
│   ε  = α × ΔT       │    + streaming responses      │
├──────────────────────┼──────────────────────────────┤
│  XGBoost Classifier  │    DataPrepPipeline          │
│  P(Passive|Active|   │    HuggingFace → JSONL → S3  │
│    Hybrid)           │                               │
├──────────────────────┴──────────────────────────────┤
│           AWS Bedrock  ·  S3  ·  IAM                │
└─────────────────────────────────────────────────────┘

40,000 synthetic thermal scenarios — Taylor658/deep-space-optical-chip-thermal-dataset

• 4 instruments — Spectrometer, Laser Communication Terminal, Waveguide Sensor Array, Photonic Signal Processor
• 3 strategy types — Passive, Active, Hybrid

# Clone
git clone https://github.com/ATaylorAerospace/Thermal-Agent.git
cd Thermal-Agent

# Install
pip install -r requirements.txt

# Configure
cp .env.example .env
# → Edit .env with your AWS credentials and S3 bucket

# Prepare dataset & upload to S3
python src/data_prep.py --output_dir data/ --upload_to_s3

# Launch fine-tuning job
python src/bedrock_finetune.py --config config/bedrock_config.yaml --action start

# Run the interactive app
streamlit run app/streamlit_app.py

from src.simulator import ThermalDriftSimulator

sim = ThermalDriftSimulator()

# Evaluate Indium Phosphide on a Jovian mission
result = sim.evaluate("Indium Phosphide", "Jovian System")

print(f"Δn = {result['delta_n']:.6f}")       # Δn = 0.061200
print(f"Strain = {result['strain']:.2e}")     # Strain = 8.28e-04
print(f"Risk: {result['risk']}")              # Risk: Critical
print(f"Strategy: {result['recommended_strategy_hint']}")  # Strategy: Hybrid

from src.simulator import ThermalDriftSimulator

sim = ThermalDriftSimulator()

for material in sim.get_all_materials():
    r = sim.evaluate(material, "Outer Solar System")
    print(f"{material:20s}  Δn={r['delta_n']:.6f}  ε={r['strain']:.2e}  → {r['risk']}")

Silicon               Δn=0.044640  ε=6.24e-04  → Critical
Silicon Nitride       Δn=0.005880  ε=1.92e-04  → Moderate
Polymer               Δn=0.026400  ε=5.28e-04  → Critical
Indium Phosphide      Δn=0.081600  ε=1.10e-03  → Critical

from src.strategy_classifier import StrategyClassifier

clf = StrategyClassifier()
clf.load("results/strategy_classifier.pkl")

# Predict with calibrated probabilities
proba = clf.predict_proba(
    material="Silicon",
    instrument="Spectrometer",
    environment="Mars Transit",
    thermal_effect="Spectral Drift",
)
print(proba)
# {'Active': 0.12, 'Hybrid': 0.61, 'Passive': 0.27}

from src.inference import BedrockInferenceClient

client = BedrockInferenceClient(model_id="your-fine-tuned-model-arn")

prompt = client.build_thermal_prompt(
    instrument="Laser Communication Terminal",
    material="Indium Phosphide",
    environment="Outer Solar System",
    thermal_effect="Waveguide Misalignment",
)

# Stream the response
for token in client.stream_invoke(prompt):
    print(token, end="", flush=True)

# One command — prepare data, launch fine-tuning, poll until complete
bash scripts/run_pipeline.sh

Thermal-Agent/
├── .github/
│   └── workflows/
│       └── ci.yml                   # GitHub Actions CI — pytest on 3.10–3.12
├── app/
│   └── streamlit_app.py             # Interactive two-tab demo
├── config/
│   └── bedrock_config.yaml          # Bedrock hyperparameters & S3 paths
├── docs/
│   └── thermals.png                 # Hero banner image
├── notebooks/
│   ├── 01_eda.ipynb                 # Exploratory data analysis
│   └── 02_bedrock_fine_tuning.ipynb # End-to-end fine-tuning walkthrough
├── results/                         # Model artifacts & evaluation outputs
├── scripts/
│   └── run_pipeline.sh              # Full pipeline runner
├── src/
│   ├── __init__.py                  # Public API exports
│   ├── bedrock_finetune.py          # Bedrock job lifecycle manager
│   ├── data_prep.py                 # HuggingFace → Bedrock JSONL → S3
│   ├── inference.py                 # Base + fine-tuned model inference
│   ├── simulator.py                 # Physics-based thermal drift engine
│   └── strategy_classifier.py       # XGBoost Passive/Active/Hybrid
├── tests/
│   ├── __init__.py                  # Test package init
│   ├── test_classifier.py           # Classifier tests
│   └── test_simulator.py            # Physics simulator tests
├── .env.example                     # AWS credential template
├── .gitignore
├── conftest.py                      # Pytest path configuration
├── CONTRIBUTING.md                  # Development setup & PR guidelines
├── README.md
└── requirements.txt

Computes refractive index shift (Δn = dn/dT × ΔT) and mechanical strain (ε = α × ΔT) for any material-environment pair. Classifies risk as Low → Moderate → High → Critical and maps to a strategy hint. Fully type-annotated for IDE support.

Loads the HuggingFace dataset, converts to Bedrock-compatible {"prompt": ..., "completion": ...} JSONL, performs stratified train/validation splitting, and uploads to S3.

Full lifecycle management — start, monitor, cancel, and list fine-tuning jobs on Amazon Titan Text Express.

Synchronous and streaming inference against base and fine-tuned Bedrock models. Both invoke() and stream_invoke() accept configurable max_tokens and temperature parameters. Includes a structured prompt builder matching the training format and a side-by-side model comparison utility.

XGBoost classifier predicting Passive / Active / Hybrid strategies with calibrated probability estimates. Includes fitted-state validation to prevent silent failures. Fast fallback when Bedrock is unavailable.

Tests run automatically via GitHub Actions CI on every push and pull request against Python 3.10, 3.11, and 3.12.

# Run all tests locally
pytest tests/ -v

# Run simulator tests only
pytest tests/test_simulator.py -v

# Run classifier tests only
pytest tests/test_classifier.py -v

See CONTRIBUTING.md for development setup, testing instructions, and pull request guidelines.

This work is licensed under CC BY 4.0.

Copyright (c) 2026 A Taylor

Have questions, ideas, or want to collaborate? Reach out directly: