Laibaff7/cda-hmoc-classifier

HMOC Binary Classifier (5-model ensemble)

A 1D-CNN ensemble that flags High-Mass Organic Cation (HMOC) spectra inside large pools of CDA (Cosmic Dust Analyzer) mass spectra. It is an expert triage model: it does not guarantee a clean label, but it ranks the top candidates with a high density of true HMOCs for manual review.

Task & data

• Formulation: anomaly/retrieval binary classification — HMOC (class 1) vs. clean, non-HMOC background (class 0).
• Imbalance: only 83 labeled HMOC rows vs. 2,600 curated clean negatives (31:1). ? and Noise classes were dropped to keep the negative class high-quality. No SMOTE / augmentation.
• Input: each raw spectrum is truncated or zero-padded to 1000 bins, then min-max normalised to [0, 1] per spectrum.

Architecture

Shallow 1D CNN (deliberately small to avoid overfitting 83 positives):

Conv1d(1->32, k=7, p=3) + BN + ReLU
Conv1d(32->64, k=5, p=2) + BN + ReLU
Conv1d(64->128, k=3, p=1) + BN + ReLU
AdaptiveAvgPool1d(1)
Linear(128->64) + ReLU + Dropout(0.4)
Linear(64->1)  -> sigmoid

Ensemble

Five independent models, one per StratifiedKFold fold, trained with seeds [42, 100, 2026, 777, 888]. At inference the mean of the 5 sigmoid outputs is the final HMOC probability (lower variance than any single model).

Training

• Loss: Focal Loss (alpha=0.85, gamma=2.0) to focus on hard/rare positives.
• Optimizer: AdamW (lr=1e-3, weight_decay=1e-4).
• Batches: WeightedRandomSampler by inverse class frequency.
• Early stopping on validation AUPRC (patience=15).

Metrics (per seed, validation fold)

Seed	Fold	Precision	Recall	F1	AUPRC	Best threshold
42	0	0.435	0.625	0.513	0.508	0.759
100	1	0.400	0.706	0.511	0.359	0.446
2026	2	0.571	0.471	0.516	0.455	0.680
777	3	0.636	0.438	0.519	0.484	0.826
888	4	0.700	0.438	0.538	0.567	0.768

AUPRC baseline for a random guess here is ~0.03, so 0.36–0.57 is well above chance.

Usage

import numpy as np
from model import predict_ensemble, SEEDS, default_weights_dir

d = default_weights_dir()
paths = [f"{d}/hmoc_model_seed_{s}.pth" for s in SEEDS]

raw_spectra = [...]  # list of 1-D intensity arrays (any length)
probs = predict_ensemble(paths, raw_spectra)   # -> np.ndarray of HMOC probabilities

# rank candidates
order = np.argsort(probs)[::-1]
top = [(raw_spectra[i], float(probs[i])) for i in order[:100]]

Single-model variant:

from model import load_model
net = load_model(f"{d}/hmoc_model_seed_888.pth")

Files

Path	Contents
model.py	Self-contained architecture, preprocessing and ensemble inference
config.json	Architecture, training hyperparameters, per-seed metrics/thresholds
hmoc_model_seed_{42,100,2026,777,888}.pth	model.state_dict() for each seed
scripts/train_hmoc_ensemble.py	5-fold training of the ensemble
scripts/generate_metrics.py	Per-fold validation metrics
scripts/infer_70k.py	Scores the ~70k unlabelled pool with the ensemble
predictions/top_359_hmoc_predictions.csv	Top-ranked HMOC candidates (sclk, confidence %)
docs/METHODOLOGY.md	Full write-up of data curation, architecture, loss and evaluation
docs/hmoc_metrics.md	Metrics table

The scripts expect the training / unlabelled parquet files under data/; the data itself is not included. The full 70k-row candidate table (1.2 GB, includes spectra) is also left out; rerun scripts/infer_70k.py to regenerate it.

License

No open-source license is provided for these weights; all rights reserved. Do not reuse, redistribute, or build upon this model without the author's permission.