Digitizing the Sense of Smell

Sight and sound were digitized decades ago. We have cameras to capture images and microphones to record audio, but the sense of smell has remained stubbornly analog. Until now. A wave of deep-tech startups is currently using machine learning to map molecular structures to specific odor perceptions. This technology is not just about creating digital perfumes; it is about diagnosing diseases, ensuring food safety, and revolutionizing how computers understand the physical world.

The Science of Digital Olfaction

For years, scientists struggled to predict how a molecule would smell simply by looking at its structure. In vision, we know that a specific wavelength of light equals a specific color (red is around 700 nanometers). Smell does not work like that. Two molecules can look almost identical structurally but smell completely different. Conversely, two molecules with very different structures can smell exactly the same.

To solve this, researchers are using Graph Neural Networks (GNNs). This is a type of deep learning model designed to analyze data that can be represented as graphs, such as molecular bonds.

The Principal Odor Map

The breakthrough came when researchers, many of whom originated from Google Research, created a “Principal Odor Map” (POM). They trained an AI model on a dataset of over 5,000 known molecular structures and their corresponding odor descriptors (like “creamy,” “grassy,” or “meaty”).

By analyzing the complex geometry of atoms and bonds, the AI learned to place molecules on a map based on how they smell rather than just how they look. This map allows computers to predict the scent of a never-before-smelled molecule with human-level accuracy.

Key Startups Leading the Charge

Several companies are currently racing to commercialize this technology. They are moving beyond academic theory to build hardware and software that can “smell.”

Osmo

Osmo is perhaps the most prominent player in this space. Spun out of Google Research in early 2023 and led by neuroscientist Alex Wiltschko, Osmo launched with $60 million in Series A funding.

Osmo is using the Principal Odor Map to design new fragrance ingredients. Their goal is to create molecules that smell identical to rare or endangered natural ingredients (like sandalwood or certain musks) but are made from renewable, safe, and biodegradable sources. They are effectively using AI to search through billions of potential molecular combinations to find the perfect scent.

Aryballe

Based in Grenoble, France, Aryballe takes a hardware-first approach. They combine biochemical sensors with machine learning. Their device, the NeOse, uses silicon photonics. It works by grafting peptides (short chains of amino acids) onto a silicon chip.

When odor molecules hit these peptides, they cause a chemical reaction that the chip reads as a “signature.” Aryballe’s software then compares this signature against a database to identify the smell. This is widely used in the automotive industry to detect lingering odors in rental cars and in the food industry to ensure batch consistency in raw materials.

Moodify

This Israeli startup focuses on “functional fragrances.” Moodify uses AI to design scents that have a specific physiological effect. They have developed a technology called “White Scent,” which works similarly to white noise.

Instead of masking a bad smell with a strong perfume (like spraying lavender over garbage), Moodify’s AI identifies the specific molecules causing the malodor. It then creates a precise “counter-scent” that blocks the human brain from perceiving the bad smell entirely. This has massive applications for waste management, public restrooms, and automotive interiors.

Real-World Applications

While the idea of sending a smell over the internet is the futuristic dream, the current applications are far more practical and immediately valuable.

Early Disease Detection

The human body emits Volatile Organic Compounds (VOCs) through breath and skin. These compounds change when we are sick. Historically, doctors relied on their own noses to detect sweet-smelling breath (diabetes) or fishy odors (liver failure).

Digital noses can detect these VOCs at parts-per-billion concentrations. Companies are developing breathalyzers that can screen for conditions like Parkinson’s disease, various cancers, and COVID-19 long before clinical symptoms appear.

Food Safety and Quality Control

Supply chains are currently testing digital noses to monitor freshness. A sensor inside a shipping container can detect the specific gas emitted by rotting fruit days before the spoilage is visible.

This reduces food waste significantly. If a distributor knows a shipment of strawberries is about to turn, they can route it to a local grocer immediately rather than sending it on a three-day cross-country trip.

Pest Control

Insects rely heavily on smell to find mates and food. Startups are using digital olfaction to analyze the pheromones of specific pests. By synthesizing these pheromones, they can create traps that attract only the target species (like mosquitoes or crop-destroying moths) without using toxic pesticides that harm bees or other beneficial insects.

The Challenges Ahead

Despite the rapid progress, digitizing smell remains difficult. The primary challenge is the sheer complexity of the real world. A cup of coffee does not emit a single molecule; it emits hundreds of different compounds that mix in the air.

Separating background noise (like pollution or perfume) from the target smell (like a gas leak) requires incredibly sophisticated filtering. Furthermore, “smell” is subjective. Cultural background and genetics influence how humans perceive odors. For example, cilantro smells like soap to a percentage of the population due to a genetic variant. Teaching an AI to account for these human variances is the next hurdle for the industry.

Frequently Asked Questions

Can we transmit smells over the internet yet? Not yet. While we can send the data for a smell, the receiving user needs a device capable of synthesizing chemicals to recreate it. “Smell synthesizers” exist as prototypes, but they are bulky, expensive, and limited to the chemical cartridges they hold.

How accurate are digital noses compared to dogs? Dogs are still the gold standard for sensitivity. A dog’s nose has roughly 300 million scent receptors, whereas human noses have about 6 million. However, digital noses are becoming more consistent. A dog gets tired or distracted; a sensor does not.

Is digital smell technology expensive? High-end industrial sensors like those from Aryballe are expensive and intended for corporate use. However, the goal for companies like Osmo and various sensor manufacturers is to miniaturize the technology so it can eventually be integrated into smartphones or smart home devices.

What is the “Principal Odor Map”? It is a machine learning model that organizes molecules based on their scent properties. It allows scientists to predict how a molecule will smell based on its atomic structure, similar to how a color wheel organizes light frequencies.