CONDITIONAL RELIABILITY MAPS FOR LANGMUIR-PROBE INVERSION UNDER NOISE, HOT-ELECTRON TAILS, AND FORWARD-MODEL DISCREPANCY

Langmuir probe inversions can appear precise when the assumed electron distribution or collection law is wrong. This study maps where three inversion strategies remain reliable by separating noise, voltage sampling, hot-electron structure, and forward-model discrepancy in synthetic argon plasma. Conventional semi-log analysis, robust whole-curve single-Maxwellian fitting, and multistart two-Maxwellian fitting screened by an operational BIC-shaped penalized-residual score were applied to 2,896 traces. The benchmark comprised 1,440 factorial noise and hot-fraction cases, 220 matched-family Latin-hypercube cases, 360 voltage-resolution trials, 300 out-of-family challenges, and 576 temperature-ratio cases. Under the matched Maxwellian family, whole-curve fitting reduced median absolute temperature error from 10.40% to 1.83% at 5% noise and density error from 2.58% to 1.13%. Matched-family Latin-hypercube validation produced median errors of 0.56% in temperature, 0.22% in density, and 0.014 V in plasma potential. Under the out-of-family challenge, robust-fit errors increased to 2.30%, 12.39%, and 0.543 V, revealing parameter-specific sensitivity concealed by matched-family validation. At 1% noise and a 20% hot fraction, forcing one Maxwellian produced mean-temperature biases of 5.87%, 21.69%, and 52.02% for hot/cold temperature ratios of 2, 3.5, and 5. The gate selected the two-temperature model in 0/240 sampled Maxwellian trials (exact 95% Clopper-Pearson interval, 0-1.5%), while hot-tail selection varied with noise, fraction, and temperature separation. The resulting maps are conditional reliability bounds, not universal detection limits, showing that small matched-family errors do not establish model-independent metrological accuracy.