Dr Gleb Shishaev’s  PhD thesis has been done and dusted – minor correction completed and approved by the examiners.
The PhD Thesis —  History Matching and Uncertainty Quantification of Reservoir Performance with Generative Deep Learning and Graph Convolutions was examined by Prof Denis Voskov (external, TU Delft), Prof Ahmed ElSheikh (internal) and found worthy a PhD degree.
This was a long and windy road PhD for Gleb who complete a full-time PhD  effectively being a distance learning student visiting Edinburgh for supervision sessions. His PhD studies were disrupted by Covid pandemic and lockdowns. Never-the-less Gleb demonstrated a great dedication to deliver a quality and novel work on history matching with Graph Variational Autoencoder (GVAE) supported by his supervisors Prof Vasily Demyanov, Dr Dan Arnold.
The thesis introduces a novel approach to assisted history matching via a latent space generated by (GVAE) that aims to handle uncertainty across multiple geological model scenarios to generate multiple history matched models. The new GVAE method is compared to conventional VAE and Wasserstein VAE on a synthetic case study demonstrates GVAE is superior in reproducing geological realism via the latent space constructed through through learning on a prior ensemble of models that cover multiple geological scenarios.

Key advantages of history matching with Graph Variational Auto-Encoders include:

• More coherent representation of geological structures with graph description of unstructured grids rather than lattice-based neural networks.
• More consistent conditioning to reservoir static and dynamic well data in balanced history-matching loop, when both static and dynamic data jointly condition the model rather then in a sequential conditioning in convention geomodelling.
• Better control of geological realism in describing prior geological uncertainty comprising multiple model scenarios with the latent space and subsequent control of geological realism during the model update in the search for history matched models.

Furthermore, the thesis implement GVAE history matching to Brugge benchmark case. This case study was presented by Gleb (recorded) as an e-poster at ECMOR 2022: History Matching and Uncertainty Quantification Of Reservoir Performance With Generative Deep Learning And Graph Convolutions.