⚡ WebGPU Inference

Browser-side AI4S surrogate inference. Two modes: GP kernel demo (Matérn/RBF) or 10 real trained MLP checkpoints loaded from data/small_models. Automatic CPU fallback.

Real model data: data/small_models/*.npz holds 10 MLP surrogates trained on SwarmLabs scenarios (Forrester/Branin/Hartmann6 + bio/chem/ML/physics). Each checkpoint is ~2KB JSON with float32 weights — loaded live, no server roundtrip. Forward pass runs entirely on WebGPU compute shader.

🖥️ Environment Detection

WebGPU Support
Detecting…
Backend
—
Device
—
Fallback
—

🔬 GP Model Configuration

Kernel
Matérn-ν=1.5
Length Scale ℓ
0.300
Signal Variance σ²
1.000
Smoothness ν
1.5
Noise ε
0.03
Training Points
16

📊 Inference Performance

0
Grid Points
—
Wall Time (ms)
—
Mpts/s
—
Compute Path

🎨 Predictive Surface

GP mode: heatmap shows μ(x,y) = K(x*,X)·α (blue dots = training points). MLP mode: 1D models render as a line plot; 2D+ render as a heatmap (high-D shows a 2D slice).