Robust and informed Deep Learning for bad data with applications in earthquake engineering

Advances in Deep Learning bring further investigation into credibility and robustness, especially for safety-critical applications. Fully data-driven models may impose a performance ceiling due to data quality. Data problems pervasive in engineering and environmental studies hinder from learning effectively the true mechanism and being robust to uncertain inputs. There arises needs for characterisation, propagation, and quantification of uncertainty embedded in the bad data through a probabilistic computational pipeline. This thesis develops novel Deep-Learning based frameworks that (i) characterise bad data, allowing for robustness; (ii) incorporate physical domain knowledge, allowing for better-informed modelling. In this context, research works in this thesis are composed of two parts associated with unique data problems: (i) estimation of spectral representations of stochastic processes in the presence of missing data; (ii) uncertainty-aware detection and separation of concurrent seismic events. Addressing the first, an Augmented-Learning framework incorporating prior domain knowledge of physical processes is developed. Bayesian deep learning models are trained to learn the underlying stochastic process, probabilistically capturing temporal dynamics, from the physics-based pre-simulated data. Models are characterised by the knowledge-elicited probability distribution of model parameters, whereby uncertainties over the learned models, reconstructions and spectral representations are all quantified. This framework provides a robust solution to the general arbitrary missing data pattern in a non-stationary setting, even under significant incompleteness. As an extension, another Bayesian updating framework delving into additional means to further inform the modelling under missing data is presented. Posterior knowledge of the model representations for the underlying process is updated on physics-informed prior elicitation, in light of even incomplete empirical observation. Besides, a post-processing component applying extra physical constraint to the reconstruction is incorporated. This framework mitigates the data-model dependence dilemma due to the paucity of ground motion records in data scarce region, serving as a complementary approach to the current stochastic ground-motion models. The second data problem involves development of a Bayesian ConvNet based framework (named DBNet) that detects and separates unwanted or multiple concurrent seismic events, facilitating the automated and uncertainty-aware preprocessing procedure in dealing with continuous data streams under a significant increase of seismological data.