Variational Data Assimilation with a Learned Inverse Observation Operator

Authors: Thomas Frerix, Dmitrii Kochkov, Jamie Smith, Daniel Cremers, Michael Brenner, Stephan Hoyer

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results for the Lorenz96 model and a two-dimensional turbulent fluid flow demonstrate that this procedure significantly improves forecast quality for chaotic systems.
Researcher Affiliation Collaboration 1Google Research 2Technical University of Munich 3Harvard University. Correspondence to: Thomas Frerix <thomas.frerix@tum.de>, Stephan Hoyer <shoyer@google.com>.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes 1https://github.com/googleinterns/ invobs-data-assimilation
Open Datasets No We train on a dataset of 32000 independent observation trajectories with batch size 8 for 500 epochs. (No access information provided for this dataset.)
Dataset Splits No The paper mentions a training dataset and test trajectories ('on a set of 100 test trajectories') but does not specify a separate validation split or its size.
Hardware Specification Yes All models can be trained and optimized on a single NVIDIA V100 GPU.
Software Dependencies No The paper mentions using JAX, Flax, and the Adam optimizer but does not provide specific version numbers for these software components.
Experiment Setup Yes We implement our models for the approximate inverse in JAX (Bradbury et al., 2018) and use Flax as neural network library, with the Adam optimizer (Kingma & Ba, 2015) and learning rate of 10 3 for training1... We train on a dataset of 32000 independent observation trajectories with batch size 8 for 500 epochs. We use L-BFGS (Nocedal & Wright, 2006) as an optimizer for assimilation, retaining a history of 10 vectors for the Hessian approximation.