CoFFe: Collaboratively Fused Features for Learning Multi-physical System on a Unified Manifold

Welcome to the CoFFe project repository!

CoFFe stands for Collaboratively Fused Feature, a method designed for nonlinear data dimensionality reduction of multi-physics coupled systems.

CoFFe offers the following key contributions:

- High Efficiency and Flexibility: The model eliminates the need to process complete FOM data in each epoch and imposes no restrictions on input dimensions, computational domain shapes, or the discretization methods. The encoder utilizes sparse sampling to progressively cover the entire computational domain during training.

- Sparse Recognition Capability: The pre-trained model reliably identifies system states at a given condition from measurements, accommodating different sensor configurations in terms of number and location.

- Fast Inference: In downstream tasks, the fine-tuning process converges quickly with varying types, numbers, and positions of measurements, enabling the retrieval of the unified system feature using only sparse data from partial physical field(s).

- Strong Extensibility: By capturing invariant features of multi-physical systems, the model is easily adaptable to downstream tasks requiring sparse and partial observations, such as parameter inversion, sensor arrangement optimization, and few-shot prediction of previously unobserved variables.