Emulation of greenhouse?gas sensitivities using variational autoencoders

Flux inversion is the process by which sources and sinks of a gas are identified from observations mole fraction. The often involves running Lagrangian particle dispersion model (LPDM) to generate simulations movement over domain interest. LPDM must be run backward in time for every measurement, this can computationally prohibitive. To address problem, here we develop novel spatio-temporal emulator sensitivities that built using convolutional variational autoencoder (CVAE, two-piece neural network capable condensing reconstructing images). With encoder segment CVAE, obtain approximate (variational) posterior distributions latent variables low-dimensional space. We then use Gaussian on space emulate new at prediction locations points. Emulated passed through decoder CVAE yield emulated sensitivities. show our CVAE-based outperforms more traditional empirical orthogonal functions it used with different LPDMs. conclude emulation-based approach reliably reduce computing needed outputs high-resolution flux inversions.