Gab Abramowitz_The e-MAST data-model interface

282 views

Published on

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
282
On SlideShare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
0
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide
  • Structured talk mostly to speak to non-modelling folk
  • Models incorporate our understanding of how natural systems workThe types of models that can benefit from TERN-type data
  • Take, for example, a continental scale simulation of Australian C cycle over a decade.Equifinality = underconstrainedmodelling system
  • Reasons why the modelling community get lazy
  • This is not all entirely complete
  • This is not all entirely complete
  • This is not all entirely complete
  • This is not all entirely complete
  • This is not all entirely complete
  • This is not all entirely complete
  • This is not all entirely complete
  • Gab Abramowitz_The e-MAST data-model interface

    1. 1. The e-MAST data-model interfaceGab AbramowitzClimate Change Research Centre, UNSWARC Centre of Excellence for Climate System Science
    2. 2. Outline• Why care about models?• Why model evaluation is complicated• How models and observations interact• How model-based experiments quantify uncertainty• Why multiple data streams are particularly important• Protocol for the Analysis of the Land Surface (PALS) – the eMAST data portal – and what it’s trying to achieve
    3. 3. Why care about models?• Land surface models; hydrological models; ecosystem models.• They drive: climate projections, weather forecasts, water resources assessments, impacts assessments for natural systems• Many modelling areas have a long history of development before high quality observations became available – legacy code is still a real issue.• Observations play a key role in reducing uncertainty in model predictions, yet communication channels between modelling and observational groups are often poor
    4. 4. Model evaluation is complicated• Diagnostic model evaluation is contingent parameters upon the quality of the observations used in all the four green categories. input MODEL Outputs• Model outputs may cover a wide range of s systems – e.g. carbon, water, energy states• Uncertainty in these is rarely low enough to tightly constrain simulations• Leads to “equifinality” – several different combinations of inputs / initial states / parameters give equally good results.• This makes identification of the “best” model structure very difficult NASA LIS
    5. 5. How models and observations interact• Simple comparison of observations with predictions (obs ≈ model output)• Parameter estimation (obs ≈ model output)• Direct restriction of parameter ranges (obs ≈ model parameters)• Comparison with empirical approaches (obs ≈ model input / output / params)• Data assimilation generating reanalysis products (obs ≈ model output / state) All can give diagnostic information about model structure
    6. 6. How models and observations don’t interact• Long pathways to data access / uncertainty in availability• Physical inconsistencies in data (i.e. quality control)• Data formatting and a lack of standardisation (file formats, standards within formats, time and space sampling) • e.g. CMIP5 database is anticipated to be ~50PB• Example – Fluxnet and the land surface modelling community
    7. 7. Gauging uncertainty in model predictions• Uncertainty in predictions is typically estimated by sampling uncertainty in parameters parameters, inputs, and/or initial states input MODEL Outputs• Multiple streams of data mean that s uncertainty ranges can be better states constrained => more reliable predictions• Examples: meteorology, C fluxes, water and heat fluxes, biomass, carbon pool sizes, physical soil properties, vegetation characteristics, soil moisture and temperature, streamflow, N, P etc NASA LIS
    8. 8. Multiple data streams – an example Model parameter estimation based on each data type separately, and in combination (error bars are uncertainty from propagated parameter uncertainty – 1σ): See Vanessa’s talk tomorrow at 11:10am Prior estimate Eddy fluxes Streamflow LitterfallHaverd et al, BGD, 2012 Eddy fluxes + Litterfall Streamflow + Litterfall Streamflow + Eddy fluxes Eddy fluxes + Litterfall + Streamflow 0 1 2 3 4 -1 NPP (GtC y )
    9. 9. How the eMAST data portal addresses this issue• The Protocol for the Analysis of the Land Surface (PALS) is a web application for evaluating land surface/ecosystem/hydrology models
    10. 10. How the eMAST data portal addresses this issue• The Protocol for the Analysis of the Land Surface (PALS) is a web application for evaluating land surface/ecosystem/hydrology models• PALS hosts Experiments:  All data sets required to drive/force a model for an experiment are downloadable in a standardised netcdf format, easily read by most LS modelling groups internationally
    11. 11. How the eMAST data portal addresses this issue• The Protocol for the Analysis of the Land Surface (PALS) is a web application for evaluating land surface/ecosystem/hydrology models• PALS hosts Experiments:  All data sets required to drive/force a model for an experiment are downloadable in a standardised netcdf format, easily read by most LS modelling groups internationally  Users upload their model simulations for an experiment in a way that maximises reproducibility – ancillary files (log and parameter files, run scripts etc)
    12. 12. How the eMAST data portal addresses this issue• The Protocol for the Analysis of the Land Surface (PALS) is a web application for evaluating land surface/ecosystem/hydrology models• PALS hosts Experiments:  All data sets required to drive/force a model for an experiment are downloadable in a standardised netcdf format, easily read by most LS modelling groups internationally  Users upload their model simulations for an experiment in a way that maximises reproducibility – ancillary files (log and parameter files, run scripts etc)  PALS automatically runs analysis of the model output, comparing with a range of observational data sources (where available), other models and standard benchmarks
    13. 13. What PALS aims to achieve• Bridge the observation and modelling communities • Avoid data formatting issues and quality control issues • Expose modelling communities to the nature of uncertainties within observational data • Give data collection communities access to model simulations that utilise their data• Provide international exposure for a broad collection of ecosystem field data• Introduce standard reference benchmarks for modelling communities• Provide a fast, free, and comprehensive model evaluation facility for model developers• Provide detailed diagnostic information that can identify model weaknesses across the international modelling community
    14. 14. PALS – progress so far• Structure for site-based experiments (primarily based on flux tower sites) is in place and being used by around 140 researchers internationally.• Met Office (UK) coordinating an experiment between 6+ land surface modelling groups using what’s already in place (presentation from Martin Best at AMS, Austin, Jan 2013):
    15. 15. PALS – progress so far• Structure for site-based experiments (primarily based on flux tower sites) is in place and being used by around 140 researchers.• Involvement with GEWEX Global Land Atmosphere System Study (GLASS) panel; OzEWEX; CABLE LSM management group; Met Office training…• Generic Experiments structure being built now: • Single-site; multiple-site; catchment-based; regional; global • Field ecosystem, hydrological and satellite-based data streams • First inclusion likely continental-scale C budget – Vanessa’s talk 11:10am tomorrow• Early engagement from hydrological research community• PALS adopted as offline benchmarking environment for CABLE (Australian community land surface model)
    16. 16. The last slide• Models that produce climate, weather, hydrological, ecological projections are ripe to benefit from integration of ecological data streams• “Benefit” can mean diagnostic model evaluation AND reduction in projection uncertainty - concurrent multiple data streams especially useful• To date the modelling and observation communities have not communicated as well as they might – we’d like to try to help• The PALS web application tries to address this by serving standardised model experiments that utilise multiple data streams• PALS already has around 140 users, active international experiments and is used as a model development tool by a number of modelling groups• PALS is expanding to include distributed experiments (up to global scale) for a range of model types Gab Abramowitz: gabriel@unsw.edu.au

    ×