This is exactly the problem I've been obsessing over. The challenge is that olfaction isn't like vision. you're not detecting photons at discrete wavelengths, you're dealing with ~400 olfactory receptor types responding to millions of possible volatile molecules in combinatorial ways.

MOX sensors (like the SnO2 in this paper) have been around for decades but hit a fundamental ceiling—they require specific coatings to bind to specific VOCs. Want to detect a new substance? You're changing hardware.

The more promising path, IMO, is carbon nanotube (CNT) sensors that actually mimic how our nose works. Instead of measuring bulk resistance changes, you functionalize CNT arrays to respond to specific molecular binding events—much closer to how olfactory receptors operate. detection of new substances becomes a software/ML problem rather than a hardware redesign. That's how biology does it—your nose doesn't grow new receptors, your brain learns new patterns.

Full disclosure: I'm building in this space (https://nosy.network) Nosy is using CNT paired with transformer models to create what we call a "Large Essence Model" (LEM). LEM "GPT for smell" processes scent information similar to how LLMs process text.