DataSoniPrint — 2D/3D data → 3D‑printable tactile relief

Turn scientific data, math formulas, or images into a watertight STL relief you can 3D‑print and feel. Built for inclusive teaching: a printed relief lets students — especially those with visual impairments — grasp a graph, map, or function by touch, not just by sight.

• Live app: https://datasoniprint.streamlit.app
• Project home & example gallery: https://www.physik.uzh.ch/~kaskoe/
• Licence: CC BY 4.0 — free to use/share/adapt with credit to “DataSoniPrint — B. Penning & K. König, University of Zurich.”

A “Tactile Data” micro‑innovation project by Prof. Dr. Björn Penning and Dipl.‑Des. M.Mus. Kaspar König, funded by ULF (Universitäre Lehrförderung), University of Zurich.

---

What it does

Three ways to make a relief, all ending in a printable STL:

1. 📊 From a data column — CSV / HDF5 / NetCDF / GRIB / ASDF. Pick X, Y, Z columns; the data is spread onto a 2D height grid.
2. 🔢 From a math formula — type sin(x)*cos(y), or pick from a curated equation gallery (Riemann ζ, Schrödinger states, gravity well, Mandelbrot, …). A complex mode renders f(z) with z = x + i·y (magnitude / real / imaginary / phase).
3. 🖼️ From an image — a plot/photo becomes a relief by brightness. Optional engraved text or braille axis labels (via OCR).

Plus: a live 3D preview, bed‑size/height controls, and a one‑click binary STL download.

---

Run it locally

pip install -r requirements.txt
# OCR labels (optional) also need the Tesseract binary:
#   Ubuntu/Debian: sudo apt-get install tesseract-ocr
#   macOS:         brew install tesseract
streamlit run app.py

Python 3.10+ recommended.

---

Under the hood

The core idea: an STL relief is a heightfield — one real height z over a 2D (x, y) grid. So every input mode is just a different way to produce a normalized z(x, y) grid, which is then turned into a solid, watertight mesh.

The code is split in two:

File	Role
app.py	Streamlit UI — widgets, session state, preview, download. No heavy math.
core_engine.py	Pure, testable processing pipeline. No Streamlit imports.

The pipeline

input ──► normalized height grid z(x,y) in [0,1] ──► scale to bed ──► watertight mesh ──► binary STL

1. Ingest & normalize

• load_file() dispatches by extension to CSV/HDF5/NetCDF/GRIB/ASDF loaders and returns {column: np.array}.
• normalize_to_01() rescales any array to [0, 1] (NaNs handled), so the height scale is controlled purely by the mm sliders later.

2. Build the 2D height grid (one of):

• Data: select_columns_for_axes() + spread_to_grid() project 1‑D column data onto a resolution × resolution grid (reshape or scatter‑interpolate).
• Formula: evaluate_formula() parses with SymPy and evaluates on a NumPy meshgrid. Complex functions go through _evaluate_complex() (mpmath, so zeta/gamma work) and are projected to a real height. mandelbrot_field() is a dedicated escape‑time generator. EQUATION_GALLERY + evaluate_gallery_item() hold the curated presets (formula, domain, height, pole‑clipping/log options).
• Image: build_image_relief() reads brightness into a height grid; detect_text_regions() (Tesseract OCR) finds labels; apply_labels() either engraves the glyphs (negative height) or embosses standard braille dots (text_to_braille_cells()), sized in real millimetres.

3. Scale to the print bed & mesh — scale_to_bed()

• Fits the grid onto a ≤200 mm bed, adds a 1 mm base, and builds the top surface, bottom plate, and all four side walls so the mesh is watertight.
• Fully vectorized with NumPy (a pure‑Python loop took minutes on a 2 M‑face relief). Outward normals are set analytically per face group (top +Z, bottom −Z, walls ±X/±Y) instead of the slow adjacency‑graph fix_normals().

4. Export — mesh_to_stl_bytes() writes a binary STL (≈5× smaller and faster than ASCII).

Design decisions that matter

• Memory safety (images): source images are capped to MAX_RELIEF_DIM = 800 px on the longest side in _load_grayscale(). A 5 MB JPG can be 12+ megapixels → without a cap it becomes tens of millions of mesh faces and OOM‑kills the worker. 800 px keeps the worst case ~2 M faces while giving ~8 px/mm of detail.
• True‑aspect bed: the bed is sized to the input so a non‑square relief isn't stretched into a square. Each mode sets sensible defaults: data fits the grid's aspect ratio to a 100 mm longest side at 10 mm height (a clear bas‑relief); formula sizes from the domain; image from the pixel aspect.
• Preview vs. print: the 3D preview is on demand (a Generate 3D preview button) and uses peak‑preserving downsampling so sharp features (Γ poles, sinc ripples) survive instead of being strided over. It also exaggerates the Z axis for legibility — thin reliefs would otherwise look flat — while the exported STL always uses the true millimetres shown on the sliders.
• Memory readout & maintenance: memory_usage_mb() powers the sidebar gauge. A hard OOM kill (SIGKILL) can't be reported by the process, so the gauge lets you watch the trend, and Python‑level MemoryErrors are caught with a clear message. The sidebar ⚙️ Maintenance panel adds Clear project & free memory (drops heavy objects + gc.collect()) and Restart app server for a guaranteed clean slate. Uploaded data files are parsed once and cached, not on every rerun, so slider changes don't re‑ingest a large file.
• Upload/limits: .streamlit/config.toml sets maxUploadSize/maxMessageSize. Note the hosted tier has a fixed RAM budget — a very large data file can still exceed memory while being parsed, even though the upload itself is allowed.

---

Project layout

app.py                     Streamlit UI (thin) + PROJECT_HOME/SOURCE links
core_engine.py             All processing (load, grid, formula, image, mesh, STL)
requirements.txt           Python deps
packages.txt               apt packages for Streamlit Cloud (tesseract-ocr)
.streamlit/config.toml     Upload/message size limits
site/                      Landing page (index.html) + example photos → UZH ~kaskoe
sample_data.csv            Example dataset

Deployment

• App: Streamlit Community Cloud, repo Kultkat/DataSoniPrint, branch streamlit-rebuild, main file app.py. Auto‑redeploys on push. packages.txt installs Tesseract so OCR labels work.
• Landing page: site/index.html (+ the example JPGs) is uploaded to the UZH personal web space served at www.physik.uzh.ch/~kaskoe.

---

Credits & licence

Developed at the University of Zurich (Physik‑Institut) as the “Tactile Data” project by Prof. Dr. Björn Penning and Dipl.‑Des. M.Mus. Kaspar König, funded by ULF. Released under CC BY 4.0 — please keep the attribution described above. See LICENSE.