New UQ-SHRED framework adds uncertainty quantification to sparse sensor reconstruction using distributional neural networks

UQ-SHRED extends the SHRED architecture to handle uncertainty estimation in high-dimensional spatiotemporal field reconstruction from sparse sensor data

Uses engression (neural network-based distributional regression) to model predictive distributions of spatial states conditioned on sensor history

Addresses limitations of original SHRED in complex, data-scarce, high-frequency, and stochastic systems by providing valid uncertainty estimates

Training employs stochastic noise injection into sensor inputs and energy score loss optimization for improved uncertainty modeling