An international research collaboration has developed a machine-learning model that can predict the perceived similarity of complex odor mixtures, potentially advancing computational fragrance and flavor development and the broader effort to make smell machine-readable.
“We’ve been interested for a long time in trying to digitize odors, to mathematically represent these in some way similar to what we have done with color vision and for hearing,” said co-author Joel Mainland, Ph.D., a member of the Monell Chemical Senses Center.
He added, "What we’re seeing is that ... if you know what the components smell like, you know what the mixture smells like—it’s basically just an average of the components, which I think is really surprising to a lot of people in the field."
Researchers from the center, Yale University, IBM Research and a broad international group of academic and industry-affiliated institutions have developed a computational model that can predict how similar or different complex odor mixtures will smell to people. The work, published Aug. 4 online and Aug. 11 in Proceedings of the National Academy of Sciences (PNAS), addresses one of the major unresolved problems in olfactory science: developing a quantitative framework for connecting the physical composition of a complex smell with human perception.
(In unrelated work, startup Patina is working to build a “Pantone for scent”—a universal system of primary scent building blocks that can be used to construct any fragrance or flavor on demand.)
The paper, “A semantic-based community model for high-fidelity tuning of olfactory mixture distances,” is led by Vahid Satarifard of Yale University's Human Nature Lab and includes Joel D. Mainland and colleagues at the Monell Chemical Senses Center, as well as researchers from Boston University, University of Michigan, KU Leuven, Université Côte d’Azur/CNRS, Oxford, KTH Royal Institute of Technology, Rockefeller University, Weizmann Institute of Science, UC Davis, University of Pennsylvania and IBM Research.
The result is significant because most smells encountered in the real world are not produced by individual molecules. They are mixtures of many molecules, often dozens or hundreds. While machine learning has made substantial progress predicting how individual odor molecules are perceived, predicting the perceptual properties of combinations has proved considerably more difficult.
Vision and hearing have long benefited from quantitative systems that allow researchers to describe physical stimuli and predict aspects of human perception. Those frameworks have enabled technologies ranging from digital imaging to audio recording and reproduction. Olfaction has not had an equivalent system.
The researchers describe the problem in terms of perceptual distance: if people smell two mixtures, how far apart do those mixtures seem? A distance of zero would mean the odors are perceived as indistinguishable, while a larger distance indicates increasingly different olfactory experiences.
Creating a reliable way to calculate that distance could become an important building block for computational olfaction. It provides a way to translate the inherently subjective experience of smell into a quantitative measure that algorithms can analyze and predict.
The research grew out of the second DREAM Olfactory Mixtures Prediction Challenge in 2025, a community effort that brought together international teams to develop models capable of predicting perceptual distances between odor mixtures. The researchers assembled a unified dataset, established common benchmarks and compared approaches developed independently by participating teams.
The challenge involved 26 teams, which collectively submitted 159 leaderboard predictions; 19 teams ultimately submitted predictions for the test set. Rather than selecting a single winning algorithm, the researchers combined the strongest approaches into a post-challenge ensemble.
The resulting model was tested against a hidden set of 46 odor-mixture pairs. It reduced root-mean-square error by approximately 33% to 0.08 compared with previous state-of-the-art approaches, while increasing Pearson correlation by 53% to 0.57.
The researchers then tested the model on an independently designed validation set containing 50 new mixture pairs. Performance remained strong, suggesting that the model was not simply memorizing the mixtures used to develop the original benchmark.
A further ensemble that retained only the models based on olfactory semantic features improved the Pearson correlation to 0.61 on the original test set and 0.54 on the independent validation set.
One of the study's most intriguing findings concerns the type of information the best-performing models used. The strongest models relied heavily on compact semantic representations of individual odor molecules, rather than relying exclusively on molecular structure.
That suggests that information encoded in linguistic descriptions of individual odorants can capture meaningful aspects of the perceptual space in which complex mixtures are experienced. In practical terms, descriptions of individual molecules may provide useful information about how those molecules will contribute to the perceived similarity or difference between mixtures.
The researchers say this finding suggests that mixture perception may not require a fundamentally different representational system from single-molecule olfaction.
That is an important conceptual advance. Previous computational work has established that machine learning can predict perceptual descriptors for individual odor molecules. The new research suggests that some of those representations can be extended into the much more complicated territory of mixtures.
The Monell Chemical Senses Center's involvement is particularly notable because of its longstanding focus on the biological and perceptual science of taste and smell. Mainland is an author on the paper, with an additional affiliation with the Department of Neuroscience at the University of Pennsylvania. Xuebo Song, Tiffany Yang and Robert Pellegrino are also affiliated with Monell.
Monell was also involved directly in the study's human validation work. Researchers recruited 16 participants at the Monell Chemical Senses Center, who evaluated the independently generated odor mixtures used to test the model. The study notes that the participants were 18 to 55 years old, with an average age of 32, and that the procedures were approved by the University of Pennsylvania institutional review board.
That human testing is important because the ultimate target of the model is perceptual similarity. A pair of formulations can be chemically quite different but smell similar to people. Conversely, relatively small chemical changes can sometimes produce substantial perceptual differences. The validation experiments therefore provide a human reference against which the machine-learning predictions can be evaluated.
The research has obvious implications for industries that work with complex odor systems. A fine fragrance, functional fragrance, flavor system or consumer-product odor is not simply the sum of its ingredients. Perception emerges from the combination, concentration and interaction of many odorants. That creates an enormous design space.
A computational system capable of reliably estimating perceptual distance could eventually help researchers navigate that space more efficiently. Potential applications include comparing formulations, identifying candidates that are perceptually close to a target, screening large numbers of possible combinations and understanding how formulation changes affect overall odor character.
The study itself does not demonstrate an AI system that can formulate a finished perfume or flavor. Nor does it claim that the model can predict every sensory attribute of a mixture. Rather, the work establishes a reproducible way to measure and predict the perceptual distance between complex smells.
For fragrance developers, for example, knowing that two formulas occupy similar positions in perceptual space could eventually become useful when evaluating alternatives to a formulation or exploring variations around a creative target. For flavor developers, a similar framework could eventually help compare flavor systems or identify formulations that occupy a desired region of sensory space.
Those applications remain future possibilities rather than direct capabilities demonstrated in the current study.
The researchers see the implications extending beyond formulation. A quantitative metric for olfactory perception could become part of the infrastructure required for digital olfaction—the broader effort to measure, model and ultimately reproduce smells using machines.
That could include more sophisticated electronic noses, computational odor analysis and systems that translate chemical measurements into perceptual descriptions.
The paper also points toward the longer-term possibility of transmitting and reproducing olfactory information, which the authors describe in the context of concepts such as “odor teleportation.” That remains speculative, but it illustrates the technological implications of establishing a standardized computational representation of smell.
The nearer-term opportunity is considerably more practical: creating a common numerical language for comparing odors.
That said, the findings should not be interpreted as meaning that AI has solved the complexity of smell. The researchers identify important limitations related to the availability and distribution of training data. Prediction performance was stronger in portions of perceptual space where more data were available, while performance declined toward the extremes of similarity and difference, where fewer examples existed.
There is also an inherent limitation in human sensory data. Any computational model attempting to predict human perception is ultimately constrained by how consistently humans themselves perceive and rate odors. This makes the creation of larger, better-standardized psychophysical datasets an important next step.
It also highlights why the community-based structure of the research may be as important as the model itself. The project brings together expertise in olfactory neuroscience, psychophysics, chemistry, machine learning, statistics and computational science. Its authors span 23 listed institutional affiliations, with the DREAM Olfactory Mixtures Prediction Consortium adding another large network of contributors.
The development comes at an interesting moment for computational fragrance and flavor. Much of the recent discussion around AI in these industries has focused on generating formulas, predicting ingredient properties or accelerating formulation. This study tackles a more fundamental problem: Can a machine understand whether two complex smells will be perceived as similar?
Generating candidate formulas is relatively straightforward if a system has enough chemical possibilities to explore. Determining whether those formulas will actually produce the desired perceptual experience is considerably harder. The PNAS research suggests that machine learning is beginning to establish a bridge between those two worlds.
For the fragrance and flavor industries, the immediate impact is likely to be in research rather than replacing creative or sensory expertise. Perfumers, flavorists and sensory scientists still bring contextual knowledge about balance, character, performance, application, cultural interpretation and consumer response that cannot be reduced to a single similarity score.
But a reliable computational measure of olfactory distance could become another tool in that process, particularly as datasets improve and models become better at representing the complexity of real-world mixtures.
