Nanopore PNP Neural-Operator Portal

Inputs

Surrogate model

V_app50.0 mV

C₀100 mol/m³

q_w (dimensionless wall charge)-0.0100

R (pore radius)5.0 nm

Debye length λ_D—

λ_D/R—

Wall charge σ—

Mean axial current—

Inference latency—

SG/FV reference—

ψ(r,z) dimensionless potential

c₊(r,z) / C₀

c₋(r,z) / C₀

ρ(r,z) = c₊−c₋

Heatmaps: vertical axis = z/L (0 = reservoir, 1 = reservoir), horizontal axis = r/R (0 = centre, 1 = charged wall). When comparison is on, the maps show the operator − SG/FV difference on a diverging scale.

Axial current profile I(z) in pA

Source: SG/FV direct solves (float64, 21×61) and ONNX operators exported from results/models/. Accuracy: DeepONet-reparam global relative L2 on held-out in-distribution states ψ 8.9×10⁻⁴, I 1.1×10⁻³. All inference runs client-side via onnxruntime-web.

Verified SG/FV benchmark states

One-click presets solved with the float64 finite-volume Scharfetter–Gummel reference solver. Toggle the comparison to overlay the operator error on the fields in the Operator inference tab.

Show error vs. SG/FV reference (selected preset)

Selecting a preset writes its parameters into the sliders and re-evaluates the operator; switch back to tab 1 to inspect the fields and the error overlay.

Live SG/FV reference solver

Runs the exact float64 Scharfetter–Gummel finite-volume reference in Python at the current slider parameters (~0.20 s). In the deployed portal the button calls the same-origin /api/solve_sg endpoint; locally it uses http://127.0.0.1:8765/api/solve_sg from scripts/serve_interactive.py.

The live solve is independent of the surrogate: it validates the operator by recomputing the benchmark current and the axial profile I(z) from first principles.