JUPYTER / PYTHON / DATA-DRIVEN DESIGN

Earthquake: Data into Space

2025–26 / Bartlett School design project · Data-driven design

Earthquake: Data into Space

A design workflow connecting Jupyter, Python, Blender, and Processing

This earthquake study turns visual features, narrative signals, and geophysical records into computable design variables. Collection, cleaning, analysis, and fusion in Jupyter lead to Python-driven architectural fragments, ground fractures, and animated scenes in Blender, followed by particles, waves, and real-time feedback in Processing.

My contribution: Python data processing, design mappings, and spatial visualisation
Project stage: Multimodal analysis, procedural modelling, and real-time visualisation

From data to design

How can earthquake data become rules for spatial form and dynamic behaviour?

Statistics, photographs, and news each describe different aspects of earthquakes. This project converts those descriptions into computable features and a shared parameter interface, connecting analysis, procedural modelling, and real-time visualisation across tools.

240cleaned images

232news titles and summaries

240fusion design samples

20spatial instances

01 / Jupyter · Python
Collect and clean
Collect images, text, and seismic records; remove duplicates and structure the data.
02 / Python
Analyse and fuse
Extract features, standardise, cluster, and reduce dimensions into shared design parameters.
03 / Blender · Python
Rules and space
Translate parameters into fragments, ground displacement, fractures, waves, and dust.
04 / Processing
Behaviour and feedback
Read records sequentially to drive particles, propagating waves, and a data panel.

Complete multi-instance field of architectural fragments and luminous waves on black — The spatial field: shared rules generate distinct, comparable instances.

Images and visual features

icrawler collects images of collapsed buildings, urban rubble, and rescue operations. Invalid files and hash-based duplicates are removed. OpenCV and PIL extract average brightness, grayscale contrast, and Canny edge density from the remaining 240 images.

Complete original image-sample collage of rubble, architectural fragments, and rescue scenes — Image-sample collage: from visual references to computable features.

Image brightness histogram — Brightness distribution

Canny edge-density histogram — Edge-density distribution

Image damage-proxy distribution — Visual fragmentation proxy

Edge density, contrast, and darkness form damage_proxy, a design cue for fragmentation. It is an approximation based on visual features, rather than a measured structural-damage rating. Standardised features enter four K-means groups, with PCA used to inspect their distribution.

Counts of the four image clusters — Four cluster counts: 71 / 54 / 78 / 37.

PCA projection of clustered image features — PCA projection reveals relationships between visual features.

Inspect the original Python feature and clustering logic

df_img["darkness"] = 255 - df_img["brightness"]
df_img["damage_proxy"] = (
    df_img["edge_density"] * 0.5
    + (df_img["contrast"] / df_img["contrast"].max()) * 0.3
    + (df_img["darkness"] / df_img["darkness"].max()) * 0.2
)

features = df_img[["brightness", "contrast", "edge_density", "damage_proxy"]]
X = StandardScaler().fit_transform(features)
df_img["cluster"] = KMeans(n_clusters=4, random_state=42).fit_predict(X)

Introduce human presence

YOLOv8 records person_count and total_objects, adding signals for later fragment density and spatial complexity.

YOLOv8 person-count distribution — Detected people

YOLOv8 total-object distribution — Detected objects

Text and narrative signals

Earthquake-related headlines and summaries are combined, deduplicated, and cleaned into 232 records. Keywords such as damage, death, rescue, and collapse are extracted. TextBlob polarity distinguishes positive, neutral, and negative wording, turning narrative content into design variables.

Keyword frequencies in the earthquake news sample — News keyword frequencies

Positive, neutral, and negative text counts — Text polarity describes wording; it does not measure affected people’s emotions.

Count Vectorizer and TF-IDF representations are compared. TF-IDF features enter four K-means clusters, and PCA displays a two-dimensional projection. Keywords suggest instability and collapse, while clusters and polarity introduce variation in form and motion.

Four news-text cluster counts — Text groups

PCA projection of news TF-IDF features — Two-dimensional text-feature projection

Seismic records and data fusion

Date, magnitude, depth, latitude, and longitude come from the Significant Earthquakes, 1965–2016 dataset. Pandas removes invalid records and normalises variables with different units into a shared numerical interface.

Magnitude distribution of sampled seismic records — Magnitude distribution in the sample

Depth and magnitude of sampled records — Depth and magnitude: inspect the ranges and distribution of variables.

The fusion notebook adds disaster-text samples and satellite-tile features, comparing TF-IDF, Sentence-BERT, and image representations. Reduced features are standardised alongside numerical variables and emotion-intensity signals before five-group K-means clustering. The output contains 240 design samples.

Counts of the five fusion clusters — Five clusters after multimodal fusion

Three-dimensional preview of longitude, latitude, and magnitude — Longitude, latitude, and magnitude in three dimensions

Chaos, fear, and calm labels in the fusion sample — Semantic labels generated by rules or a model

Disaster-text keyword frequencies in the fusion workflow — Keyword distribution of the disaster-text sample

Scope of the fusion

Image and news analyses are separate branches. The final fusion pairs seismic records, disaster texts, and image features by sample without establishing a time-and-location match to the same event. Emotion and fragmentation signals explore design parameters in a data-driven generative experiment.

Python design mappings

Data translation makes the connection between features and design tools explicit. Each parameter has an input, a mapping, and a spatial effect. CSV acts as the interface between Jupyter, Blender, and Processing. Select a source sample below to inspect how magnitude and depth change its parameters.

DATA → PARAMETERS

Select a record to inspect its design parameters

240 design samples

Cluster 0Cluster 1Cluster 2Cluster 3Cluster 4

Source record

Magnitude: 6.1
Depth: 60.0 km
Geometry scale: 1.667 ×
Vertical offset: -0.087
Fragmentation: 0.600
Material group: 3

scale = 1 + magnitude_norm × 4
z = −depth_norm

Values come directly from the original final_fusion_data.csv. Clusters combine numerical, semantic, and image features. Sample pairings are design experiments, rather than verified matches to the same earthquake event.

Input variable	Mapping rule	Design effect
`magnitude_norm`	`1 + magnitude_norm × 4`	Geometry scale and shake amplitude
`depth_norm`	`z = −depth_norm`	Vertical displacement
`longitude_norm / latitude_norm`	`x / y`	Normalised spatial coordinates
`emotion_intensity`	`fragmentation / noise`	Fragmentation and disturbance intensity
`emotion_label`	`rotation_chaos`	Rotation and motion behaviour
`cluster_id`	`material_index`	Material or asset family

Inspect the original Python mapping and export logic

fusion_df['x'] = fusion_df['longitude_norm']
fusion_df['y'] = fusion_df['latitude_norm']
fusion_df['z'] = -fusion_df['depth_norm']
fusion_df['scale'] = 1 + fusion_df['magnitude_norm'] * 4
fusion_df['fragmentation'] = fusion_df['emotion_intensity']
fusion_df['rotation_chaos'] = np.where(
    fusion_df['emotion_label'].eq('panic'),
    fusion_df['emotion_intensity'],
    fusion_df['emotion_intensity'] * 0.4
)
fusion_df['material_index'] = fusion_df['cluster_id']
fusion_df.to_csv(OUTPUT_DIR / 'final_fusion_data.csv', index=False)

Download the numerical design samples ↗

Data-driven spatial generation

Python imports the data into Blender. Height extrusion and planar distribution first test the mappings; clustering density, structural fracture, emissive cracks, and epicentral waves follow. These five complete model views from the report show the progression from direct mapping to integrated spatial composition.

Direct extrusion

Test the relationship between values and height.

Spatial distribution

Spread individual objects into a comparable spatial field.

Clustering and density

Use grouping and density to organise local structures.

Fragmentation and instability

Introduce variation in orientation, fracture, and irregularity.

Integrated composition

Integrate height, density, fractures, and waves in one scene.

From one scene to multiple instances

After testing the rules, 20 parameter combinations form a grid for comparison. Each instance combines architectural fragments, displaced ground, fractures, waves, rubble, and dust. Data influences scene density, deformation, and motion intensity.

Complete grid render of 20 fragment instances and circular shockwaves — Compare different data conditions under shared rules.

Complete detail render of fragments, ground fractures, luminous rings, and dust — Images become surface textures; data parameters drive scale and disturbance.

Inspect Blender Python programming and scene development

Original Blender scene and Python script development record

From space to live behaviour

The exported Blender model and final_fusion_data.csv enter Processing. Every 25 frames, the program reads a record. Magnitude controls particle count, emission force, colour, and wave speed; depth and coordinates appear in the data panel. Automatic scene rotation turns static data into an observable temporal process.

Complete eight-frame Processing sequence with data panels, particles, and circular waves — Complete sequence: read data → release energy → expand waves → dissipate particles.

Capabilities demonstrated

Python structures disparate sources into analysable tables. Feature extraction, standardisation, clustering, and dimension reduction explain variation. Numerical values become reusable spatial rules, with CSV connecting analysis, modelling, animation, and real-time feedback.

Jupyter / Pandas / NumPy / OpenCV / scikit-learn / Blender Python / Processing