spatial research · v0.4.2

A neural pipeline that simulates adaptation.

SPARC (Spatial Research) is a spatial neural model with built-in causal inference. The pipeline ingests your data, learns local relationships, and simulates how a place adapts under interventions — with posterior uncertainty on every pixel.

neural · causal 13 domain templates CUDA · GPU-accelerated
sparc · bootv0.4.2 · 000%
 
n=1,247 · missing 12
ΔAAT_z = −0.258
Pct_Canopy (−) · σ 0.154
ensemble · posterior
χ² = 0.93
8 chains · R̂ < 1.01
philadelphia · scenario
+12.4 °C
UHI peak · jul 2025
causal · validated
24 → 41
nodes → directed edges
what it does

Three things, done with rigor.

An engineer's notebook for spatial science. Every step is auditable, every coefficient is interrogable, every claim has a confidence interval.

01 · neural model

Locally-weighted neural ensembles.

Every prediction has a local model. Coefficients vary smoothly across space; uncertainty travels with them. Spatial heterogeneity is a feature, not noise.

physics-constrained
02 · causal inference

Causal, not just correlational.

The pipeline learns under directed causal structure. Edges have direction, signs are physically grounded, and counterfactuals respect the graph.

monotone constraints
03 · adaptation engine

Simulate "what-if" — with error bars.

Increase canopy 12%. Lower impervious 8%. Get a posterior mean, σ, and a pixel-wise uncertainty raster — never a single number.

uncertainty quantified
live · interactive

The pipeline is the product.

A simplified urban-heat run on Philadelphia census tracts. Every stage is reproducible from project.yml.

pipeline · 7 stages · the product
running · stage 03
posterior trace · ΔAAT_z
live · 8 chains
μ = −0.258
chain 1 chain 2 chain 3 chain 4
adaptation scenario
philadelphia · 1.2km grid
Baseline
+11.8 °C
σ 0.42
+12% canopy
+9.4 °C
σ 0.51
Δ predicted mean−2.4 °C
13 templates · 1 blank slate

Domain-agnostic, not domain-naive.

All templates

UHI

n=24
Urban Heat Island

Coastal

n=18
Erosion + sea level

Drought

n=21
Soil moisture · ET

Geotechnical

n=16
Bearing capacity

Groundwater

n=19
Aquifer drawdown

Wildfire

n=27
Fuel · weather · ROS

Air Quality

n=22
PM₂.₅ + dispersion

Stormwater

n=17
Runoff · flooding
how it works

Setup → Causal → Pipeline → Adapt.

Four stages, hand-numbered. The same flow whether you're modeling drought across a watershed or noise around a runway.

01Setup

Bring your project

Drop a project.yml, point at your raster + vector data, set CRS and time window. Or fork one of 13 templates.

02Causal

Set the structure

SPARC proposes a causal graph from the correlogram; you accept, reject, or rewrite. Physics constraints (signs, monotones) are first-class.

03Pipeline

Run the neural ensemble

Locally-weighted models train in parallel across the grid. Posterior traces stream live. χ² and σ gate every stage.

04Adapt

Simulate and report

Compare adaptation scenarios on a single map. Export pixel-wise uncertainty rasters and a fully reproducible audit trail.

join the waitlist

The desktop app is in private beta.
Researchers go first.

Tell us your domain and your CRS. We'll send a build with the right template pre-loaded and a sidecar binary that talks to your GPU.