Food-tech company NotCo has developed a novel generative AI model, the Generative Aroma Transformer (GAT), capable of creating new flavor and fragrance formulations. The model, which the company presented on at the Food AI Summit last month, could be a potentially disruptive new tool that could impact a variety of consumer goods markets such as food, personal care, home care, and beauty industries.
The company’s Senior VP of Product, Aadit Patel, talked about the new model in a post on Linkedin today, highlighting how GAT can translate textual prompts into unique chemical formulations. “The system intakes your prompt, such as ‘an ocean scent on a breezy summer day on a tropical island’ to create a novel chemical formulation of that scent in one-shot.” From there, the model generates a corresponding fragrance formula. According to Patel, the model is built on a “natural language to chemical composition” framework, tokenizing molecules to create a system capable of understanding and generating novel combinations.
NotCo says early tests have been extremely positive. Their research indicates that GAT’s abilities rival those of human perfumers. In blind smell tests, fragrances created by GAT proved indistinguishable from those crafted by human experts. This finding signals a potentially significant shift in the industry, where only 600 certified perfumers exist globally.
How it Works
According to research presented at the Food AI Summit, the core of GAT’s functionality lies in its ability to understand and model complex interactions between volatile molecules. The model is trained on an extensive dataset of historical fragrance formulations and the molecular structures of volatile compounds. This training enables GAT to decipher the subtle relationships between different molecules and predict how they will interact to create specific aroma profiles.
The model employs a dual-system transformer network comprising an encoder and a decoder. The encoder processes the user’s prompt (with inputs such as top note (cherry candy), middle note (vanilla) and bottom note (cherry)), capturing the desired aroma profile. This is passed to the decoder, which generates a corresponding sequence of tokens representing the fragrance’s molecular structure.
GAT leverages the atomic structure of volatiles to generate novel formulations. Each molecule is represented as a graph, with atoms described by valence, degree, hydrogen count, hybridization, formal charge, and atomic number. These details are then translated into numerical representations and fed into a Graph Neural Network (GNN) model, which creates a unique vector representing each molecule. Similar vectors indicate similar molecules, allowing GAT to identify and utilize molecular structures with desired aromatic properties.
The potential impact of NotCo’s GAT is substantial. Developing new formulations for flavors and fragrances has traditionally been a time-consuming and resource-intensive process, often requiring weeks or months of expert work. If GAT can achieve the same outcome in mere seconds, it could significantly reduce flavor and fragrance development costs.
If you’d like to learn more about NotCo’s new generative AI tool for developing flavors and fragrances, NotCo’s head of machine learning, Francisco Clavero, and one of their key flavor and fragrance scientists, Cindy Sigler, will be our guests at the next Food AI Co-Lab on October 17th. You can register for this virtual event here.
Leave a Reply