Laser-plasma accelerators (LPAs) promise very large accelerating fields in a compact device. But routine operation is hard because electron-beam properties drift during a run and the physical cause is often invisible to standard diagnostics. This paper proposes a practical way to diagnose those drifts by inferring three hidden interaction‑point variables from routine beam measurements and by deciding which subsystem likely caused the change.

The authors write the LPA as a latent state‑space model. The hidden state contains three physical quantities: the normalized laser amplitude a0 (how strong the laser is at the focus), the normalized plasma electron density ñe, and the residual pulse chirp C (a remaining frequency sweep in the laser pulse). An emission model maps those latent variables to routine shot‑by‑shot beam observables such as centroid energy, relative energy spread and total charge. A separate transition model describes how the latent state can drift between shots. An extended Kalman filter then estimates the three hidden variables from the sequence of measurements.

Keeping the emission model separate from the transition model lets the method answer two questions in turn. First, the emission model tells which latent variable moved by showing how changes in a0, ñe or C would change the diagnostics. Second, competing transition models (based on hardware and environmental drivers) are compared to assign what actually moved the variable. The paper implements a toy emission model that uses known scalings from the three‑dimensional blow‑out regime and tests the detection and attribution steps in synthetic sessions.

The approach matters because current tools either re‑optimize settings without saying why performance changed, or find correlations that can shift over time. A diagnostic that reports “the energy fell because the delivered amplitude fell” gives operators a specific subsystem to check and can reduce hours lost to manual troubleshooting. The construction needs only a set of physical latent variables, an emission model and plausible hardware‑derived transition models, so it can transfer to other drift‑prone accelerator subsystems.