Research question and contributions

Central research question

Can lightweight machine-learning models combined with resource-aware explanation strategies maintain useful predictive performance and interpretable behavior under chronological electronic-nose sensor drift, while satisfying the memory, latency, and energy constraints of TinyML-class microcontrollers?

This question is not yet answered. Chronological evaluation of classical models (Checkpoint 1–2) provides partial evidence toward the first half; the explainability, quantization, and hardware components required to answer the second half are not executed. See Results for what is currently supported, and Pipeline for what remains.

Why not “Pareto-aware explainability for TinyML”?

That framing understates the work. The intended contribution combines chronological e-nose drift evaluation, resource-aware explainability, explanation fidelity and stability under drift, quantized TinyML deployment, physical nRF52840/PPK2 measurement, reproducible evidence lineage, and an accuracy–trust–resource Pareto analysis — a multi-axis systems contribution, not a single explainability technique.

Contributions

Each contribution is labeled by its current evidence state, not its intended importance. A Planned label means genuinely not yet executed — see Pipeline for the specific blocking stage.