NEURAL OPERATORS × FUSION ENGINEERING
A tokamak,
in response.
Explore MAST-U structural displacement. Choose a prediction or vary a parameter uncertainty box to see lower and upper displacement enclosures.
WHAT YOU ARE SEEING
One machine.
Two parameters.
A continuous response.
This is a steady structural response, with displacement shown on a simplified cutaway of the tokamak. Changing a slider evaluates a compact approximation of the trained neural operator directly in your browser. The colours update as you explore; no simulation server is needed.
The offline export encodes all 29.2 million native source nodes before evaluating the neural operator on the display mesh. The browser approximates those display predictions. This is a surrogate prediction of structural displacement, rather than a transient plasma simulation.
Approximation accuracy & provenance
The validated export report will appear when the model has loaded.
Reported errors compare the browser approximation with the trained network at separate test parameters and display vertices. They are empirical measurements, not a certified error bound over the entire parameter domain or an error bound against MOOSE.
Read the validation report ↗From a prediction
to an enclosure.
Explore the parameter uncertainty interactive ↗ to vary an uncertainty box around α and Tw and view lower and upper displacement enclosures, their width, and the region of validity.
These enclosures target the original neural operator at fixed display vertices. They do not certify the compact approximation shown on this page or every point inside each rendered triangle.
—
WHAT THE BOUNDS COVER
An enclosure
of the original
neural operator.
Every displayed vertex has a quadratic Taylor model with interval coefficients and a rigorous remainder inherited from the full-native, outward-rounded calculation. Your browser encloses that model over the requested parameter rectangle.
The bounds target the mathematical neural operator with its stored weights, material routing and input-coordinate casts. They cover fixed display vertices, rather than every location inside each rendered triangle. Surface colours interpolate these vertex values.
The nominal view is a Taylor estimate. The retained remainder was computed for the parent box, so enclosure width can remain positive when input uncertainty is zero. It also includes approximation conservatism and arithmetic uncertainty.
Arithmetic, coverage & provenance
Browser arithmetic pads each binary64 primitive outward. Square-root endpoints are checked with enclosing products. Binary64 bounds are retained for numerical readouts; rendering uses conservative float32 endpoint conversions.
These are surrogate enclosures. They do not bound MOOSE prediction error, represent probabilistic confidence intervals, or certify every vendor FP32/TF32 execution of the network.
Loading provenance…
Read the uncertainty export report ↗