Generation & Evaluation

Knowledge Encoding

• Knowledge encoding implements computable data structures to represent domain prior knowledge, including structure, dynamics, and physical constraints, of entities in spatio-temporal data.
• Generative models are built to explicitly leverage this encoded knowledge, enabling principled and controllable synthetic data generation.

Relation Modeling

• Relation modeling aims to identify how ocean variables are connected across time, space, and resolution. In SST analysis, this helps explain how past temperature changes, extreme high- and low-temperature events, and coarse-resolution ocean fields relate to future or high-resolution SST patterns.
• To model these relationships, we use EVL-based SST forecasting to capture temporal extreme events and GAN-based SST downscaling to learn spatial relationships between coarse- and high-resolution ocean fields.

Data Assimilation

• Data assimilation combines numerical model outputs and real observations to estimate a more reliable ocean state. It is especially useful when observations are sparse, noisy, or incomplete, because physical constraints can guide the reconstruction of missing or uncertain variables.
• To address this, we combine 3D-Var data assimilation with deep learning to generate spatiotemporal ocean data that reflects both observational evidence and physically informed constraints.

Statistical Similarity

Quantitatively evaluate distributional similarity and inter-variable correlations between synthetic and real data.
SoftwareEvalCorrSyn — Evaluation of multivariate distribution similarity in synthetic data
EvalSynLongD — Statistical similarity of real and synthetic longitudinal data (linear mixed-effects model, Cohen's d)
EvalDistSim — 1D data distribution similarity via KS test, KL divergence, and Jensen-Shannon divergence

Objective Relevance

The objective of the CNN-based synthetic data evaluation is to assess whether synthetic brain imaging data, generated to address the limitations of real neuroimaging data such as data scarcity, privacy concerns, and disease-group imbalance, possess sufficient biological, anatomical, and clinical validity to be used in real research settings and AI model training.

Specifically, this evaluation aims to determine:

• Whether synthetic images are visually and structurally similar to real brain images.
• Whether CNN-extracted features are comparable between real and synthetic data.
• Whether synthetic data preserve disease-related differences between patient and control groups across disease progression.