Photo credit: www.sciencedaily.com
Researchers at the Institute of Science Tokyo (Science Tokyo) have tackled the complexities of fragrance design by creating an advanced AI model capable of automating the formulation of new scents based on user-defined scent descriptors. This model employs mass spectrometry profiles of essential oils combined with corresponding odor descriptors to produce essential oil blends for innovative fragrances. This advancement marks a significant shift in the fragrance industry, moving past traditional trial-and-error methods to enable rapid and scalable production of scents.
The development of new fragrances is essential in sectors such as perfumery, food, and home products, where scent greatly affects consumer experience. Traditionally, the creation of fragrances has been a lengthy and intricate process, heavily reliant on the expertise of specialized perfumers. This labor-intensive approach often involves numerous attempts before achieving the desired olfactory outcome.
To streamline this process, a team of researchers led by Professor Takamichi Nakamoto at Science Tokyo has introduced a model known as Odor Generative Diffusion (OGDiffusion). This AI-driven model utilizes generative diffusion networks—a machine learning technique that constructs new content by reversing a noise process guided by existing datasets. While generative diffusion networks have been successfully used to produce images and text, this research team has adapted the technology specifically for fragrance creation. Their results were published in IEEE Access on March 27, 2025.
The OGDiffusion system analyzes the chemical profiles of 166 essential oils, each labeled with nine distinct odor descriptors, such as “citrus” or “woody.” When users input their desired scent characteristics, the AI generates a corresponding mass spectrum of chemical profiles that aligns with those descriptors. Subsequently, it employs a mathematical approach known as non-negative least squares to determine the essential oil combinations needed to reproduce the specified scent.
“Our diffusion network identifies patterns within the mass spectrometry data of essential oils to automatically create new fragrance profiles in a fully data-driven manner, ensuring high-quality results. By removing human intervention and molecular synthesis from the equation, we offer a swift, general, and efficient fragrance generation method,” Nakamoto explains.
While previous AI fragrance generation models exist, they often depend on proprietary datasets and still require the guidance of experts. The principal advantage of the OGDiffusion model is its capability to fully automate the creation of new scents. Additionally, because it produces fragrances based on essential oil formulations, the final scents can be easily recreated.
The research team also conducted sensory evaluation tests to assess whether AI-generated fragrances accurately matched the intended scent profiles. In a double-blind study, 14 participants were tasked with associating AI-created fragrances with the correct descriptors, such as “citrusy” or “floral.” The results revealed that participants consistently identified the correct scents, suggesting that the system generates fragrances that align with user expectations. In a separate test, participants successfully distinguished between two fragrances: one created to showcase a specific odor descriptor and the other without it, proving the model’s ability to produce distinct and recognizable scent profiles.
Nakamoto’s groundbreaking model is a pioneering step toward an era where AI revolutionizes scent design. “This methodology signifies a major leap in aroma innovation,” he notes. He adds, “By automating the generation of mass spectra for desired olfactory profiles, the OGDiffusion network provides a faster and more scalable approach to fragrance development. Furthermore, even individuals without extensive knowledge could create tailored scents for digital applications.”
In conclusion, this pioneering approach facilitates more efficient and adaptable scent design, offering extensive potential for various industries. By harnessing AI for scent generation, the OGDiffusion model highlights the possibility of computers becoming adept at creative tasks traditionally reserved for human expertise.
Source
www.sciencedaily.com