Latent-Space Probabilistic Reconstruction of Unsteady Oblique Detonations from Wall-Pressure Measurements

Abstract

Reconstruction of temperature fields in oblique-detonation engine (ODE) combustors is critical for state monitoring and control, yet direct sensing remains challenging. This study extends latent-space probabilistic reconstruction to confined reacting oblique-detonation wave (ODW) temperature fields from sparse wall-pressure measurements. A nonlinear autoencoder with coordinate-attention modules and a gradient-weighted loss compresses the fields into a three-dimensional latent representation, whereas an auxiliary branch preserves specific-impulse information. A probabilistic pressure-to-latent mapping infers latent variables for field reconstruction and specific-impulse estimation, and it supports separate quantification of aleatoric and epistemic uncertainties. Using 60 hydrogen–air ODE cases and 2400 snapshots, the latent space distinguishes wave-system states across Mach numbers and periodic inflow disturbances. Reconstructions recover the ODW front, reflected shocks, and recirculation zones, with structural similarity indices of approximately 0.99 and peak signal-to-noise ratios above 42 dB. Measurement-space Gramian analysis identifies lower-wall sensors near ODW initiation, reflected structures, and separation-induced shocks as dominant constraints on latent-variable estimation. Aleatoric uncertainty concentrates near strong temperature gradients, whereas epistemic uncertainty increases under out-of-distribution conditions and correlates with reconstruction error. The combined uncertainty reaches 94.5% pooled pointwise empirical coverage within the [Formula: see text] interval over the test set, supporting uncertainty-aware temperature-field reconstruction from sparse wall-pressure measurements.