Ensembling geophysical models with Bayesian Neural Networks
Authors: Ushnish Sengupta, Matt Amos, Scott Hosking, Carl Edward Rasmussen, Matthew Juniper, Paul Young
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Applied to the prediction of total column ozone from an ensemble of 15 chemistry-climate models, we find that the Bayesian neural network ensemble (Bay NNE) outperforms existing methods for ensembling physical models, achieving a 49.4% reduction in RMSE for temporal extrapolation, and a 67.4% reduction in RMSE for polar data voids, compared to a weighted mean. Uncertainty is also well-characterized, with 91.9% of the data points in our extrapolation validation dataset lying within 2 standard deviations and 98.9% within 3 standard deviations. |
| Researcher Affiliation | Academia | Ushnish Sengupta University of Cambridge Cambridge, UK us271@cam.ac.uk Matt Amos Lancaster University Lancaster, UK m.amos1@lancaster.ac.uk J. Scott Hosking British Antarctic Survey Cambridge, UK Carl Edward Rasmussen University of Cambridge Cambridge, UK Matthew P. Juniper University of Cambridge Cambridge, UK Paul J. Young Lancaster University Lancaster, UK |
| Pseudocode | Yes | An outline of the algorithm used to train our Bay NNE is provided below. Algorithm 1: Algorithm for initialising and training the Bay NNE |
| Open Source Code | Yes | The code and pretrained models accompanying this paper are hosted in a Github repository https://github.com/Ushnish-Sengupta/Model-Ensembler. |
| Open Datasets | Yes | The dataset of total column ozone observations [6] and chemistry-climate model predictions from 1980 to 2010 [28] are processed, combined and made available as a resource (https://osf.io/ynax2/download) for future studies in geophysical model ensembling. |
| Dataset Splits | Yes | Overall, the Bay NNE is trained on 77% (1.8 million datapoints), tested on 4% (85,000 datapoints) and validated on 19% (440,000 datapoints) of the available data. |
| Hardware Specification | No | The paper mentions 'research credits provided by Google Cloud' but does not specify any particular hardware details such as GPU models, CPU types, or detailed computer specifications used for the experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies or libraries used in the experimental setup, only mentioning 'TensorFlow' in a citation context. |
| Experiment Setup | No | The paper specifies details like the number of neural network ensemble members (65) and hidden layer nodes (500) for the ozone experiment, and mentions training with ADAM until convergence, but it does not provide concrete hyperparameter values such as learning rate, batch size, or number of epochs in the main text. |