PlasmaBench: Setting the Stage for AlphaCRM

The state vector of low temperature plasma may be the most poorly measured quantity in science, yet precision plasma process control underlies the 100B+ semiconductor fabrication industry and many others. Today, we steer processes by knobs (power/pressure/flows) rather than state (radicals, $n_e$, $T_e$, sheath/IED), because state diagnostics rarely achieve better than $\pm$30\% accuracy. Many processes use optical emission spectroscopy (OES) for endpoint detection, and photons are ideal nonintrusive probes, but inverting OES into state requires thousands of (often unknown) collisional–radiative pathways. Even with perfect collisional radiative models (CRM), inversion is ill-posed with multiple state capable of near-identical spectra. We propose PlasmaBench: the first dataset pairing calibrated OES with gold-standard Thomson scattering (TS) state labels ($n_e$, $T_e$/EEDF), plasma impedance probe (PIP) auxiliary state labels, and in situ process drift monitoring. PlasmaBench establishes supervised benchmarks for forward CRM validation while enabling future attempts at backwards OES-to-state Bayesian inversion using PIP state priors.