EEG Emotion Recognition From AI-Generated Biodigital Architecture Images

arXiv:2607.24808 · q-bio.NC, cs.AI, cs.HC · Submitted 2026-08-16 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "EEG Emotion Recognition From AI-Generated Biodigital Architecture Images".

Jane: The paper was written by Hongye Yang and Eva Guttmann-Flury from Beijing Institute of Architectural Design Co., Ltd. and Shanghai Jiao Tong University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Alright, welcome back to the show, everybody. We’ve got a fascinating one today, and I’m already buzzing about it. The paper is called "EEG Emotion Recognition From AI-Generated Biodigital Architecture Images."

Jane: And Tom, I have to say, just that title is a mouthful, but it’s so exciting. We’re talking about reading people’s brainwaves while they look at buildings that were designed by an AI. It’s like science fiction becoming real.

Tom: Exactly, Jane. And for our listeners who might be tuning in, let’s break down what “biodigital architecture” even means. It’s basically buildings that are inspired by nature and biology, but designed and built using digital tools and even genetic materials.

Jane: Right, so think of buildings that grow, or that have facades made of living materials, like moss or clay bricks that are three dee printed. It’s a really new field, and there aren’t many real examples out there yet.

Tom: So the researchers had a clever idea. Since we don’t have a ton of real biodigital buildings to test, they used an AI to generate images of them. They made six hundred images, and then they asked over three hundred people to rate them on an emotional scale.

Jane: And they narrowed it down to sixty images that really provoked strong feelings—either awe, disgust, or just a calm sense of contentment. Then they brought in fifty-two volunteers to look at those images while wearing an EEG cap to measure their brain activity.

Lu: You know, what’s really interesting to me, Jane, is that they didn’t just guess which brain areas to look at. They used an existing public dataset to figure out which electrodes to place on the scalp. That’s a really rigorous way to start, rather than just putting a cap on and hoping for the best.

Tom: Oh, absolutely, Lu. And that’s a great point to bring up. They even calculated the minimum number of subjects they needed based on that prior data. It’s that kind of methodological care that makes the results so much more trustworthy.

Jane: And the results are wild. They found that the gamma brainwave band, which is associated with high-level processing and attention, could predict the “awe” emotion with over seventy-seven percent accuracy. That’s huge.

Meng: I’m curious about the practical side of that, though. Seventy-seven percent accuracy is good, but what kind of EEG hardware were they using? Because if this is a cheap, portable headset, that’s one thing. If it’s a clinical-grade system, that changes the story for real-world applications.

Tom: Great question, Meng. They used an OpenBCI Cyton board, which is actually a pretty affordable, open-source system. It’s not the top-of-the-line medical equipment, but it’s accessible. And they had to deal with some noise from using a mobile phone hotspot for the connection, which is a real-world constraint.

Jane: And despite that noise, they still got those strong results. That tells us that the emotional signals are robust enough to be picked up even with a consumer-grade device. That’s a big deal for making this kind of research more accessible.

Lu: It also suggests that we could potentially use this in a design feedback loop. Imagine an architect sketching a concept, generating a few AI variations, and then testing them on a small group of people to see which one evokes the most positive emotional response before breaking ground.

Tom: That’s exactly where I think this is heading. And we’re just scratching the surface here. We’ve got the big picture, but I want to get into the nitty-gritty of how they actually processed those brain signals and what they found about the images themselves. That’s coming up next.

Summary: Jane: Welcome back. So we’ve set the stage with this paper, "EEG Emotion Recognition From AI-Generated Biodigital Architecture Images." Tom, I think we need to dig into the actual summary of what they did and what they found, because the details are really clever.

Tom: Absolutely, Jane. So after they collected all that EEG data, they had to clean it up. Brain signals are messy. They used a special algorithm to remove blinks and drift, which is that slow, wandering baseline in the signal. They even had to throw out a few subjects because their electrodes weren’t working properly.

Jane: Right, quality control is so important. Then they took that clean data and ran it through a neural network called EEGNet. It’s a compact model designed specifically for brain-computer interfaces, and it’s perfect for this kind of classification task.

Lu: And what they did next is what really caught my eye. They didn’t just look at the whole brain signal. They filtered it into the classic frequency bands—delta, theta, alpha, beta, and gamma—and then classified the emotions within each band separately.

Tom: And the results were pretty clear. The gamma band, which is the fastest brainwave, gave them the best accuracy for detecting “awe,” at around seventy-seven percent. The delta band, which is the slowest, was the next best at around seventy percent.

Meng: So the fast and the slow bands are the most informative. That’s interesting. It makes me wonder if the gamma is picking up on the visual processing of the complex textures, while the delta might be tied to a more deep-seated, emotional or motivational state.

Jane: That’s a really intuitive way to think about it, Meng. And they didn’t stop at just classifying the emotions. They wanted to know *why* people felt the way they did. So they used a technique called SHAP values to figure out which features of the images were driving the model’s decisions.

Tom: And this is where it gets really practical for architects. They found that the presence of greenery, like trees and plants, was the biggest factor in whether an image was rated as awe-inspiring. That makes sense, right? We’re drawn to nature.

Lu: But they also found some more subtle things. Images that looked “sticky” or damp were consistently disliked. And they found that images with a lot of visual complexity, or what they call “granularity,” were preferred over ones that were too uniform.

Jane: So it’s not just about adding a tree. It’s about the texture and the materiality of the building itself. A facade that looks dry and clean, with a rich, varied texture, is going to be received much better than something that looks wet or monotonous.

Tom: And that’s a direct, actionable insight for designers. It gives them a target to aim for. And it shows that this method can be used to get objective feedback on design choices that are usually just a matter of subjective opinion.

Meng: I’m still thinking about the hardware, though. They used a low-cost EEG system, and they still got these clear results. That suggests the barrier to entry for this kind of research is dropping fast.

Jane: Exactly, Meng. And that’s what makes this paper so exciting. It’s not just a lab experiment. It’s a proof of concept for a workflow that could be used by design firms everywhere. But we need to talk about the improvements they suggest, because that’s where the future of this research really lies.

Improvements: Tom: We’re back, and we’re diving deeper into "EEG Emotion Recognition From AI-Generated Biodigital Architecture Images." Jane, we’ve talked about the results, but the authors are also pretty honest about the limitations and what needs to improve.

Jane: Yeah, and I appreciate that honesty, Tom. They point out that their EEG setup, that affordable OpenBCI board, is susceptible to electromagnetic interference. They even had to use a phone’s mobile hotspot for the network connection, which probably added more noise to the signal.

Lu: But they see that as a trade-off, not a dead end. The fact that they could still extract meaningful patterns from noisy data is actually a testament to the strength of the underlying neural signals. The improvements they suggest are about making the whole pipeline more robust.

Tom: Right. One of the big suggestions is to expand the image dataset. They only used sixty images, which is a good start, but to really train a robust model, you’d want hundreds or thousands of images covering a wider range of biodigital styles and environmental contexts.

Meng: And I bet they’d also want to test different lighting conditions and spatial arrangements. The emotional response to a building isn’t just about the building itself; it’s about how it sits in the environment, how light plays across its surfaces.

Jane: That’s a great point, Meng. They also mention that they want to look at how cultural and demographic factors play into these emotional responses. What’s awe-inspiring to someone in one culture might not be the same for someone in another.

Lu: And that’s where I think the real potential is. If we can build a model that understands these cultural nuances, we can create truly personalized environments. Imagine a system that can adapt a building’s facade or interior lighting based on the emotional needs of the people inside it.

Tom: Whoa, that’s a big vision, Lu. A building that reacts to your mood. That’s the kind of thing that could redefine how we think about architecture. But there’s a more immediate improvement too, right?

Jane: Yes, and it’s about the methodology. They used a sample size estimation based on a different dataset, which is smart, but they acknowledge that the EEG electrode density and the experimental environment in their study were different. So future work needs to refine that estimation process for specific setups.

Meng: So it’s about making the whole process more reliable and reproducible. That’s what will really push this from a cool experiment to a standard tool in the architect’s toolbox.

Tom: And that’s the bridge to the conclusion. We’ve seen the results, we’ve seen the limitations, and now we can see the path forward. Let’s wrap this up and think about what this all means for the future of our built environment.

Conclusion: Jane: Well, Tom, we’ve had a fantastic time unpacking "EEG Emotion Recognition From AI-Generated Biodigital Architecture Images." It’s one of those papers that feels like it’s opening a new door.

Tom: It really does, Jane. To sum it up, they used AI to generate images of futuristic, nature-inspired buildings, and then used EEG to objectively measure how people felt about them. They found that gamma and delta brainwaves are the best predictors of awe, and that greenery and complex textures are the most appealing features.

Lu: And the negative side is just as important. They found that anything that looks damp or sticky triggers a strong negative reaction, which makes sense from an evolutionary standpoint. We’re hardwired to avoid things that might be unhealthy.

Meng: From an engineering perspective, the most exciting part is that they did this with affordable, off-the-shelf hardware. That means this isn’t just a lab curiosity. It’s a technique that could be adopted by design studios and even city planners.

Jane: And that’s the real takeaway for me. This gives architects a way to test their ideas before they’re built, to get objective feedback on what will make people feel calm, inspired, or even uncomfortable. It’s a way to design spaces that truly support human well-being.

Tom: And it ties into sustainability too. By understanding what people find beautiful and engaging, we can design buildings that people will love and want to preserve for a long time, which is the most sustainable thing you can do.

Lu: And the future work they outline is just as promising. Expanding the dataset, incorporating cultural factors, and refining the methodology will only make this tool more powerful. I can see a future where buildings are designed in a feedback loop with the people who will use them.

Meng: It’s a shift from designing for people to designing *with* people, using their own brain activity as the guide.

Tom: That’s a beautiful way to put it, Meng. So with that, we’re going to say goodbye to this paper. It’s been a real pleasure, and we’re excited to see where this research goes.

Jane: Absolutely. Thanks for joining us, everyone. We’ll be back soon with another fascinating paper, so stay tuned. Goodbye from all of us here.

Hongye Yang, Eva Guttmann-Flury

Beijing Institute of Architectural Design Co., Ltd. · Shanghai Jiao Tong University

q-bio.NC, cs.AI, cs.HC

Submitted: 2026-08-16

Updated: 2026-08-18

Comments: 12 pages, 3 figures; published in the proceedings of SIGraDi 2024

Journal ref: Proceedings of the 28th International Conference of the Iberoamerican Society of Digital Graphics (SIGraDi 2024): Biodigital Intelligent Systems, 2024, pp. 2443-2454

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 69/100

The gist: Summary This study investigates emotional responses to biodigital architecture—a design approach that fuses biological and digital technologies using biogenetic materials for self-organizing

Key concepts

Biodigital Architecture: Buildings inspired by nature and biology, designed using digital tools and even genetic materials. Examples include buildings with living facades made of moss or three-D printed clay bricks.
EEG Emotion Recognition: Measuring brain activity while people look at images to identify emotions. Researchers used EEG to measure brainwaves associated with feelings like awe, disgust, or contentment.
Gamma Brainwave Band: The fastest brainwave frequency measured in the study. It was found to be the best predictor for detecting 'awe' emotion with approximately seventy-seven percent accuracy.

Terminology

Accuracy — 14× in this paper · explained in 1 episode
SHAP values — 5× in this paper · explained in 2 episodes
Generalizability — 3× in this paper · explained in 6 episodes
alpha — 3× in this paper · explained in 1 episode
Activation Function — 2× in this paper · explained in 1 episode
Adam — 2× in this paper · explained in 2 episodes
Adaptive Blink Correction and De-Drifting — 2× in this paper
Clustering — 2× in this paper · explained in 2 episodes
Feature Extraction — 2× in this paper · explained in 2 episodes

Summary

Summary

This study investigates emotional responses to biodigital architecture—a design approach that fuses biological and digital technologies using biogenetic materials for self-organizing buildings—by analyzing electroencephalographic (EEG) data from participants viewing AI-generated images. The research addresses a gap in the literature: while biodigital architecture offers energy-efficient, low-carbon solutions and innovative design methodologies (e.g., the Pod Hotel, Genetic Barcelona Project, and Biodigital Barcelona Clay Bricks), research on its aesthetics and public acceptance remains limited. The authors note that this situation mirrors early modernist architecture, which was criticized in the 1980s for being cold and impersonal, and emphasize that urban designers need to integrate biodigital architecture into contemporary cities in a way acceptable to the public. To overcome the limitations of self-reported questionnaires, which can distort conclusions about emotional responses, the study uses EEG signals to directly measure cognitive states, offering greater objectivity and accuracy.

The methodology is detailed and emphasizes reproducibility. First, a pre-experiment involved 336 participants rating 600 AI-generated biodigital architecture images (created using ChatGPT-4 for descriptions and DALL-E for generation) on an online platform using emojis, from a crossed-out eyes face (strongest negative reaction, scored as 0) to a star-struck face (strongest positive reaction, scored as 4). Only participants who classified a minimum of 50 images were included, resulting in 259 participants who classified an average of 149 images each. From this, 60 images were selected that elicited strong emotional responses categorized as awe (high positive scores), disgust (high negative scores), or content (minimal score variability), mitigating central tendency bias.

Channel selection was guided by analysis of the imagined emotion study dataset (Onton & Makeig, 2022), using eLORETA source localization with the MNE library in Python. The analysis revealed that the occipital, parietal, and medial prefrontal cortices are implicated in emotional responses, leading to the selection of eight EEG recording sites: Fp1, Cz, CPz, CP1, CP2, Pz, O1, and O2. Sample size estimation was performed using covariance matrices from the same dataset, calculating Riemannian averages, medians, and distances to estimate Cohen's d (approximately 0.1 for content, 0.2 for disgust, and 0.15 for awe). Using the FittedDistribution Monte Carlo (FDMC) approach, conservative estimates required 2,585 trials (43 subjects), while balanced estimates suggested 730 trials (12 subjects).

The experiment involved 52 healthy volunteers (24 women, 28 men; mean age 30.62 ± 7.52), none of whom had prior exposure to AI-generated biodigital architecture images. Each trial began with a 4-second display of a randomly chosen image, followed by participants choosing an emoji representing their emotional response, and then a gray image displayed for 3 seconds to reset the visual stream. The questionnaire consisted of 60 images, with each session lasting about 10 to 15 minutes. EEG data were acquired using the Cyton Board V3-32 by OpenBCI with a WiFi Shield and a 10-20 wet electrode cap. Due to environmental constraints, mobile phone USB tethering provided network access, which may introduce some electromagnetic interference.

Preprocessing involved down-sampling to 250 Hz, spike artifact removal (identifying extreme values above a 200 μV threshold and replacing them with filtered values from a fourth-order 5 Hz low-pass Butterworth filter), low-pass filtering with a 4th-order 15 Hz Butterworth filter, and blink removal using the Adaptive Blink Correction and De-Drifting (ABCD) algorithm. Five subjects were excluded due to faulty electrodes (S04, S05, S25, and S28 for faulty FP1 electrodes, and S21 for faulty O1 and O2 electrodes). Drift curves were computed and removed from the raw data.

For processing, the EEGNet model—a specialized convolutional neural network for EEG-based Brain-Computer Interfaces—was applied. The model uses depthwise and separable convolutions, with three blocks: a temporal convolution layer for frequency-specific pattern capture, a depthwise convolution layer for spatial filtering, and a separable convolution layer for enhanced feature extraction. The output is processed through a fully connected layer with a sigmoid activation function, trained with binary cross-entropy loss and the Adam optimizer. Additionally, each image was segmented into a 4x4 grid, and K-means clustering identified the three primary colors in each segment. Pairwise Euclidean distances were computed to measure color similarity, and the most distinct colors were selected. Each image was then evaluated using ChatGPT-4 for six characteristics: dryness, greening, curvature, smoothness, brightness, granularity, and texture, each scored from 1 to 5.

Results showed that EEGNet applied to ABCD-cleaned data across five frequency bands (delta 0.5-4 Hz, theta 4-7.5 Hz, alpha 7.5-12.5 Hz, beta 12.5-30 Hz, gamma 30-45 Hz) yielded the highest classification accuracy for the awe emotion in the gamma band (77.07% ± 13.80%) and delta band (70.65% ± 10.96%) compared to other frequency bands. The analysis focused on awe because disgust elicited minimal neural responses, failing to exceed the 33% threshold across frequency bands, possibly due to the unfamiliar architectural shapes, and content serves as a reference class that does not represent awe or disgust, risking over-representation and bias.

Feature importance was assessed using SHAP values (SHapley Additive exPlanations) to improve machine learning model transparency. The presence of trees and plants was a primary factor in preferences for certain architectural designs. Image granularity also played a significant role; disliked images were mainly attributed to perceived dryness, with sticky images receiving negative responses. The research confirmed that images with overly uniform light and dark contrasts are less popular, consistent with previous studies indicating that the sensitivity of the suprachiasmatic nucleus (SCN) to light affects preferences, favoring buildings with strong light-dark contrasts.

The discussion summarizes that gamma and delta frequency bands showed the highest explanatory power in predicting participants' preferences. Awe was linked to the presence of trees and plants, consistent with prior research on brain activity in the monkey visual cortex when exposed to natural scenes. Due to people's preference for nature-inspired living environments, biodigital architecture with natural styles can enhance mental health. Although granularity and brightness had lower predictive power than greenery, they still influenced preferences, with non-uniform and well-lit images generally favored due to their visual complexity and illumination. Conversely, images with stickiness or a damp, unclean appearance were consistently disliked, likely due to evolutionary instincts associating such environments with health risks.

Limitations include the Cyton Board V3-32's susceptibility to electromagnetic interference, exacerbated by using a mobile phone's USB interface for network connectivity, introducing significant noise to EEG signals. The experimental setting lacks professional laboratory control, leading to variability from human activities, and subtle participant actions, muscle contractions, or verbalizations before emoji selection can create artifacts. Image processing is constrained by resolution and quality, potentially losing intricate color and shape features. The sample size analysis, based on a different dataset, may not account for differences in EEG electrode density or the experimental environment, potentially limiting the generalizability of results.

The conclusions state that integrating EEG data to understand emotional responses to biodigital architecture enables designers to align their creations more closely with user preferences. The use of EEGNet and SHAP values effectively identifies visual elements that evoke positive or negative emotions, such as greenery and visual complexity, offering architects actionable insights for enhancing aesthetic appeal. Notably, EEG power in the gamma (30-45 Hz) and delta band (0.5-4 Hz) are significant predictors of the awe emotion, underscoring their relevance in visual processing of natural environments and color discrimination. The study reveals that perceived dryness or stickiness leads to negative reactions, emphasizing the need for materials and designs that convey cleanliness and health. Conversely, higher granularity—referring to more detailed and textured surfaces—elicits more favorable responses, suggesting a preference for visually complex and rich designs. Incorporating granular textures in biodigital architecture can enhance the sensory experience, making spaces more dynamic and engaging. Additionally, this style also contributes to sustainability, improving humanity's capacity to cope with climate change. Future research can build upon these findings by broadening the image dataset to include a more diverse array of biodigital architectural styles and environmental contexts, exploring the impact of various lighting conditions and spatial arrangements, and examining the influence of cultural and demographic factors on emotional responses across different populations.

Improvements for AI systems

Based on the paper, here are the specific improvements I can make to AI systems and what the improved AI system can do:

1. Emotion-Aware Generative Architecture Model

Improvement: Integrate EEG-derived emotional feedback into the generative loop of text-to-image models (e.g., DALL-E, Stable Diffusion). Use the paper's finding that gamma-band (30–45 Hz) EEG activity predicts awe with 77.07% accuracy to create a reward signal for reinforcement learning. The model learns to generate biodigital architecture images that maximize gamma-band responses associated with positive awe, while penalizing features that trigger disgust (e.g., perceived dampness/stickiness).

What it can do: Automatically generate architectural concept images that are pre-validated to evoke positive emotional responses, reducing the need for costly human-in-the-loop iterations. It can also adjust greenery, granularity, and light-dark contrast in real time to optimize emotional appeal.

2. EEG-Based Preference Prediction Module

3. Feature-Importance-Driven Design Recommender

4. Adaptive Image-Selection Pipeline for Emotion Elicitation

5. Cross-Dataset Transfer Learning for Low-Density EEG

6. Real-Time Artifact-Robust Preprocessing Module

7. Predictive Model for Visual Complexity and Preference

Summary of Capabilities:

The improved AI system can:

Generate emotionally optimized biodigital architecture images.
Predict user emotional responses from low-cost EEG in real time.
Recommend specific design modifications (greenery, texture, lighting) to enhance positive affect.
Automatically curate high-impact stimuli for emotion research.
Transfer knowledge from high-density EEG to portable devices.
Operate reliably in noisy, real-world environments.
Screen designs for aesthetic appeal using image statistics alone.

These improvements directly leverage the paper's methodological rigor (channel selection, sample size estimation, artifact correction, EEGNet classification, SHAP analysis) to create practical, deployable AI tools for architects, urban designers, and emotion researchers.