🔥 Qwen-0.6B Microgravity Flammability Regime Classifier

A fine-tuned Qwen-0.6B sequence classification model using LoRA (Low-Rank Adaptation) for predicting spacecraft microgravity flame-spread regimes under varying environmental atmospheres.

Developed as part of the NASA Space Apps Challenge · Flame in Freefall (Project: ShunnoShikha AI / Flaming Fire Fall).


🛰️ Model Overview

In microgravity ($0\text{ g}$), buoyant natural convection is absent. Flame spread is governed by forced airflow velocity, oxygen concentration, fuel thickness, and cabin pressure, dividing combustion behavior into three critical regimes:

  1. Sustained Burn (Class 0): Steady or growing flame propagation.
  2. Radiative Extinction (Class 1): Low flow / thick fuel leading to excessive radiative heat loss that quenches the flame.
  3. Blowoff (Class 2): High forced airflow exceeding chemical reaction kinetics, blowing the flame off.

Key Performance Metrics

  • Base Model: Qwen/Qwen3-0.6B
  • Test Accuracy: 99.71%
  • Macro F1-Score: 0.9968
  • Target Classes: Sustained Burn, Radiative Extinction, Blowoff

📊 Classification Report

Regime Precision Recall F1-Score Support
Sustained Burn 0.988 1.000 0.994 85
Radiative Extinction 1.000 0.993 0.996 137
Blowoff 1.000 1.000 1.000 119
Overall Macro Avg 0.996 0.998 0.997 341

💻 How to Use in Python

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel

MODEL_ID = "simanto848/qwen-microgravity-flammability"
BASE_MODEL = "Qwen/Qwen3-0.6B"

# 1. Load tokenizer and base model
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
base_model = AutoModelForSequenceClassification.from_pretrained(
    BASE_MODEL,
    num_labels=3,
    id2label={0: "Sustained Burn", 1: "Radiative Extinction", 2: "Blowoff"},
    label2id={"Sustained Burn": 0, "Radiative Extinction": 1, "Blowoff": 2}
)

# 2. Load LoRA adapter
model = PeftModel.from_pretrained(base_model, MODEL_ID)
model.eval()

# 3. Format mission prompt
prompt = (
    "NASA Microgravity Combustion Mission. Material: PMMA. "
    "Fuel Thickness: 2.5 mm. Forced Airflow Velocity: 5.0 cm/s. "
    "Oxygen Concentration: 21.0% O2. Cabin Pressure: 14.7 psi. "
    "Damköhler Reaction Proxy: 0.45. Radiative Loss Index: 1.00. "
    "Gravity: Microgravity (0 g, absence of natural buoyant updraft). "
    "What is the flame-spread flammability regime?"
)

inputs = tokenizer(prompt, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
    pred_id = torch.argmax(logits, dim=-1).item()
    probs = torch.softmax(logits, dim=-1)[0].tolist()

id2label = {0: "Sustained Burn", 1: "Radiative Extinction", 2: "Blowoff"}
print(f"Predicted Regime: {id2label[pred_id]}")
print(f"Probabilities: {probs}")

🔬 Dataset & Grounding

  • Grounded in NASA Physical Science Informatics (PSI) flight records:
    • BASS / BASS-II (Burning and Suppression of Solids aboard the ISS)
    • Saffire (Spacecraft Fire Safety large-scale tests on Cygnus vehicles)
    • FLEX (Flame Extinction Experiment)
  • Physics-reconciled labels ensure consistency with empirical Damköhler and Radiative loss extinction boundaries.
Downloads last month
21
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for simanto848/qwen-microgravity-flammability

Finetuned
Qwen/Qwen3-0.6B
Adapter
(765)
this model

Evaluation results

  • Test Accuracy on NASA BASS-II & Saffire Microgravity Combustion Dataset
    self-reported
    0.997
  • Macro F1 on NASA BASS-II & Saffire Microgravity Combustion Dataset
    self-reported
    0.997