Open · Apache-2.0 · 35 built of 35
The physics that proves a fusion machine — turbulence, magnets, disruptions, materials, control — is expensive to compute. KODEX is our suite of fast AI/ML surrogates that reproduce that physics at inference speed, each benchmarked against the real code it stands in for, each honest about its uncertainty and its limits.
A two-layer spine wraps every surrogate: KHALO attaches calibrated uncertainty, KGATE a fail-closed trust boundary. Each predictive member answers one contract — predict(x) → (y, uncertainty, in_domain) — so any code plugs into the same tools. Two capstones sit on top: KFORGE chains the fleet into a design search, KPILOT orchestrates it in natural language.
Every code — from turbulence to magnets to control — exposes the same call. You always get the prediction, an honest error bar, and a flag for whether the input is inside the region the model is trusted on.
from kronos_ml import KYRO
model = KYRO() # loads the trained surrogate
y, sigma, in_domain = model.predict(x) # y = turbulent heat flux Q
if not in_domain: # out of its trusted region
fall_back_to_full_physics() # KYRO abstains, never extrapolatesSwap KYRO for any of the 30. The toolkit's uncertainty and study machinery consumes any KODEX surrogate unchanged.
# install
pip install kronos-ml
# use any surrogate — same contract
from kronos_ml import KOIL
y, sigma, in_domain = KOIL().predict(x) # magnet stress field + quench flag