Aeroforms: Modeling Olfactory Cartographies Of Japan
Interview with Zinab Eisa & Esha Morakhiya
Zinab Eisa and Esha Morakhiya, graduate students in the Master in Design Engineering program at Harvard University, will present Aeroforms at CDFAM Tokyo, a computational model of the scents of Japan built on the 100 Scents of Japan catalog compiled by a Japanese ministry, public environmental data, and literature-derived odor thresholds.
Ahead of their presentation, they answered six questions on why they chose Japan, how they treated a cultural archive as a dataset, where the modeled results diverged from remembered scent, and what they hope engineers take from modeling a phenomenon with no established tooling or ground truth.
I can still distinctly remember the smell of stepping out of my stopover hotel near Narita airport for the first time over twenty years ago, stumbling across a shrine in a cedar forest and what I swear was a 7-23 rather than a 7-Eleven, all of it bathed in jet fuel. So I was very interested when your proposal came through, and I am curious, what made you both decide to explore the scents of Japan?
We came to olfactory perception in a fairly general way at first, just curious about smell as a design material, and in the course of that wandering we found a curated list of scents compiled by one of Japan’s ministries.
In it were a hundred smells nominated one by one, deemed worthy of preservation as part of the national landscape. What struck us wasn’t just the list itself, but the fact that a government thought this was worth doing at all, that scent could be treated as infrastructure for identity, in the same register as a landmark or a dialect.
Pulling that thread, we found that a huge amount of the foundational research on odor thresholds (the actual numbers that tell you how faint a smell can be and still be perceptible) comes out of Japanese studies. Japan also turned out to have an unusually rich, unusually open body of public data such as satellite readings, pollution monitoring, land-use records, and ministry reports.
Without that openness, this project simply would not have been buildable. Also, sitting underneath all of it was the bigger history of Japan’s climb as a global center of industry and manufacturing, which itself is a story of enormous change. We wanted to know how a national “smell-identity” might have shifted underneath that transformation, and whether it was possible to actually model that shift.
Can you each tell us a little about your backgrounds, what brought you to the Design Engineering program at Harvard, and what the Aeroforms presentation will cover at CDFAM?
Esha’s background is in design while Zinab’s is in computer science and biology. We came to the Master in Design Engineering program for the same underlying reason, from opposite directions, wanting to learn how to interface with a complementary skill set. MDE gave us a place where science and design are expected to sit and inform each other, and where being new to a space (as we both were, to olfaction specifically) isn’t a liability but the whole point.
At CDFAM we want to talk about why this kind of work matters, and more practically, about how you might begin to model something that has no established precedent.
Olfaction doesn’t have the tooling, datasets, or shared modeling conventions that vision or sound do. So a large part of the talk is about methods under uncertainty, the decisions we made, the assumptions we had to be explicit about, and the places we hit the current limits of the science and the data.
We also want to be honest about the open opportunities in this space, and about the specific limitations we ran into while building Aeroforms.
What role does the “100 Scents of Japan” catalog play in the project, and how did you approach working with a cultural archive as a data source?
The catalog is the spine of the whole project. Aeroforms exists because that curation exists. We built our model directly on top of the 100 Scents list, using it as ground truth for what a culturally remembered, nationally significant “smell of Japan” actually consists of. That gave us something rare in this space, a real human-curated reference point to model toward and against.
From there we could generate comparable, independently derived scent data and put the two in conversation, asking where lived, cultural perception of place holds up against what the physical and environmental record actually supports.
This is particularly meaningful across a period of very rapid industrial growth. At the same time, treating a cultural archive as a dataset meant respecting that it wasn’t collected as data in the first place.
We tried to build our methodology to be legible about where our numbers agree with that memory and where they don’t, rather than flattening the two into one thing.
Analyzing and synthesizing scent is notoriously difficult. How did you approach it, and what software and data sources did you use?
We tried to be as meticulous and as transparent as the current state of the field allows. Wherever direct measurement was possible, we used it. Wherever it wasn’t (and at this scale, for most of what we were modeling, the sensing technology simply doesn’t exist yet), we derived it computationally, grounded in molecular science and existing literature.
As one example, to represent the volatile organic compound (VOC) profile of Japan’s apple orchards and the land given over to that kind of agriculture, we went through the literature on VOCs associated with apple cultivation and farming practice generally, and built our representation from that.
Every VOC we included was weighted using literature-derived odor threshold values, so that how strong, heavy, or diffuse a compound actually is in the world carries into the model rather than treating every molecule as equally present.
To translate molecular structure into perceived character (what a given VOC might actually smell like), we drew on Google Research’s Principal Odor Map, which links molecular structure to olfactory perception.
Layered on top of all of that was as much public data as we could responsibly bring in, including records from multiple Japanese ministries, satellite imagery, and pollution monitoring, so that the model wasn’t just chemically plausible but grounded in the actual environmental record of where things are grown, built, and emitted.
Throughout, our goal was to make this the most heavily documented, data-driven attempt at this kind of scent modeling that we could produce, and to be explicit at every step about which numbers were measured and which were derived.
Where did the modeled results and the catalog agree or diverge, and what did you make of the differences?
The clearest divergence showed up around pollution and man-made scent sources.
Our modeled, data-driven picture of certain places carried a much stronger industrial and emissions signature than the curated list did, which makes sense, since the 100 Scents catalog is built from memory and cultural identity rather than sensor readings.
That gap became the most interesting part of the project to us, because in some cases the “memory” being preserved is of something that no longer physically exists. A lavender field that isn’t there anymore is a smell the model simply cannot compute in the present-day data, because its source is gone.
That raises a real question about scent as an archival medium. What does it mean to conserve or remember a place through smell, decades or centuries after the thing that produced it has disappeared?
It’s admittedly speculative, but we found ourselves genuinely wondering what it would take to reconstruct the streets of ancient Athens (or any place we can no longer visit) through scent, and whether work like this is a step toward being able to do that.
There’s also a more grounded, engineering-facing side to the divergence.
Rapid industrialization comes at a direct environmental cost, and that cost shows up as odor. Japan happens to regulate this well, with legal thresholds for exactly these kinds of emissions, but this isn’t true everywhere. That matters for the cultural-identity question we started with, but it also matters in a much more immediate way, for quality of life, for the simple fact of breathing fresh air rather than pollution, for all of us as a species.
The gap between what a place remembers itself smelling like and what it actually, chemically smells like now is in part a story about environmental cost, and we think that’s worth engineers’ attention as much as designers’ or cultural historians’.
Scent is an unusual subject for the CDFAM audience, typically hard-nosed (pun intended) engineers and software developers. What do you hope engineers take away from this work, and what are you hoping to learn from the other presenters and attendees?
We hope engineers leave thinking of scent the way they already think of sound or light, as a real, physical, quantifiable signal rather than an aesthetic afterthought.
It’s easy to treat olfaction as too subjective to model rigorously (we’ve heard versions of that skepticism ourselves), but the underlying chemistry, threshold science, and structure-to-perception mapping are legitimate, tractable engineering problems. We’d like this talk to make the case that olfaction deserves the same technical seriousness as any other sensory channel in computational design.
The other thing we hope lands is more about method than about smell specifically. What does it actually look like to model something with no established precedent, no standard pipeline, and no ground truth you can fully trust? That kind of work forces a certain comfort with ambiguity, being explicit about assumptions, being honest about where measurement ends and inference begins. We think that’s a transferable skill for anyone doing computational or generative work in an under-modeled domain, whatever it happens to be.
As for what we’re hoping to learn, we’re excited to see how other presenters have approached modeling other elusive or subjective phenomena, sound, haptics, material feel, or really anything that resists easy quantification.
We suspect a lot of the hard problems are analogous with each other even if the sensory channel is different. We’re also hoping for technical pushback on our own pipeline from people who build and fabricate, exposure to datasets or tools we haven’t found yet, and honestly, just the company of a community that takes computational, data-driven approaches to design as seriously as we’re trying to.
Aeroforms is one of many presentations at CDFAM Tokyo working at the edge of what computational design can currently represent, from architected materials to sensory perception. If you are exploring problems that resist easy quantification, or want to compare methods with others working under similar uncertainty, join us in Tokyo on October 8 & 9 to connect with the engineers, researchers, and software developers shaping this field.
If you cannot join us in Tokyo this year, register for updates on future CDFAM events around the world and check out the CDFAM Archives and Index of all previous recorded presentations from leading experts in computational design at all scales and physical dimensions.
