SA_autoencoder

An autoencoder to perform Organic Aerosol Source apportionment from Aerosol Mass Spectrometry measurements using literature derived reference mass spectra

Fixed-Profile Source-Based Autoencoder

This repository contains a minimal Source-Based Autoencoder (SourceAE) pipeline for source apportionment using known (fixed) profiles. The model trains only the encoder while keeping the source profiles fixed.

---

Overview

The workflow:

1. Load measurements (Measurement)
2. Load known source profiles (F_fixed)
3. Initialize encoder from the known profiles
4. Train encoder only
5. Export results and diagnostic plots via Measurement

The reconstruction follows:

X_hat = relu(X @ W^T) @ F^T

where:

• F — fixed known profiles (not trained)
• W — encoder weights (trained)
• G = relu(X @ W^T) — source contributions

---

How to Run

From the project root, run:

python train_sourceae_fixed.py \
  --input "INPUT-PATH" \
  --output /results \
  --fixed_profiles "KNOWN PROFILES PATH" \
  --fixed_labels Names of the sources ex(HOA BBOA "less oxidized" "more oxidized") \
  --epochs 1000 \
  --lr 1e-2

---

Command-Line Arguments

Required

Argument	Description
--input	Path to measurement data (.xlsx or .csv)
--output	Output prefix/directory used by Measurement
--fixed_profiles	Excel file containing known source profiles

Optional

Argument	Default	Description
--fixed_labels	HOA CCOA BBOA	Names of profiles to load
--epochs	500	Number of training epochs
--lr	1e-2	Learning rate
--random_fixed_profiles	False	Randomly select profiles from library

---

Outputs

After training, the pipeline automatically generates:

Excel

• Estimated F (profiles)
• Estimated G (contributions)
• Additional diagnostics

Plots

• Source profiles
• Time series of contributions
• Scaled residuals
• Uncertainty visualization
• Ground truth comparison (if available)

All outputs are written using the Measurement utilities.

---

Notes

• Only fixed-profile SourceAE is implemented.
• No free-profile learning is performed.
• The encoder is initialized analytically from the known profiles.
• The model enforces non-negative contributions via ReLU.

---

---

Excel errors

Install:

pip install openpyxl

NaN loss

Try:

• lowering learning rate (--lr 1e-3)
• reducing epochs
• checking input uncertainties