Model Validation and Calibration with ML
Applied machine learning to systematically identify model-observation biases, speeding up the calibration cycle for a regional ocean model.
The Challenge
Traditional trial-and-error calibration of ocean model parameters against observations was slow and didn’t scale as the number of stations and variables grew.
The Approach
Trained a machine-learning model on the systematic bias between model output and observations across stations and conditions, using it to guide parameter adjustments and flag where structural (not just parametric) model error was likely.
The Result
Meaningfully shortened the calibration cycle and improved skill scores across validation stations.
Applied using MOSAIC, the model-observation comparison framework built to support exactly this kind of systematic calibration work.
Thumbnail: CTD rosette deployment, R/V Nancy Foster — public domain (USGS, Cassia Busch).