1 / 38

Results with Neural Network Approximation for the NCAR CAM Long Wave Radiation Parameterization

Results with Neural Network Approximation for the NCAR CAM Long Wave Radiation Parameterization. Vladimir Krasnopolsky , University of Maryland & NOAA/NCEP and Michael Fox-Rabinovitz, University of Maryland. Acknowledgments: William Collins, Phillip Rasch, Joseph Tribbia, Dmitry Chalikov.

abena
Télécharger la présentation

Results with Neural Network Approximation for the NCAR CAM Long Wave Radiation Parameterization

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Results with Neural Network Approximation for the NCAR CAM Long Wave Radiation Parameterization Vladimir Krasnopolsky, University of Maryland & NOAA/NCEP and Michael Fox-Rabinovitz, University of Maryland Acknowledgments: William Collins, Phillip Rasch, Joseph Tribbia, Dmitry Chalikov NNs for NCAR LWR

  2. OUTLINE • INTRODUCTION: Motivation for the study: new paradigm – combining deterministic GCM & statistical MLT(Machine Learning Techniques) • APPROACH: “NeuroPhysics” - NN Approximations for Model Physics • NCAR CAM-2 LWR: • Accuracy and Performance of NN Approximation • Comparison of CAM-2 Climate Simulations: two 10 Year Parallel Runs with the original LWR and its NN approximation • CONCLUSIONS & DISCUSSION NNs for NCAR LWR

  3. Distribution of Total Model Calculation TimePhysics vs. Dynamics (in %) CAM-2 (T42 x L26):  3 x 3 NNs for NCAR LWR

  4. Motivations for Applying NN to LWR • Calculations of model physics take about 70% of total time in CAM-2 • Calculations of radiation, LW & SW, takes about 60% of model physics calculations • All current improvements of CAM (and other GCMs) such as increasing resolution result in further increase of calculation time for physics including LWR (to 90+% for physics) • Reexamination of model physics calculations seems to be a timely and urgent problem! NNs for NCAR LWR

  5. Generic Solution – “NeuroPhysics” Fast and Accurate NN Approximation for Parameterizations of PhysicsLearning from Data Parameterization GCM NN Approximation F FNN Y X Training Set …, {Xi, Yi}, … Xi Dphys NN Approximation FNN Y X NNs for NCAR LWR

  6. Major Advantages of NNs Relevant for Approximating Model Physics: • NNs are very generic, accurate and convenient mathematical (statistical) models which are able to approximate model physics as complicated nonlinear input/output relationships (continuous mappings ). • NNs are robust with respect to random noise and fault- tolerant. • NNs are analytically differentiable (training, error and sensitivity analyses): almost free Jacobian! • NNs are simple and fast. Training (only one for a model version!) is time consuming, application is NOT! • NNs are well-suited for parallel processing NNs for NCAR LWR

  7. NNs for NCAR CAM-2 Long Wave Radiation Parameterization NNs for NCAR LWR

  8. NN for CAM-2 Physics CAM-2 Long Wave Radiation • Long Wave Radiative Transfer: • Absorptivity & Emissivity (optical properties): NNs for NCAR LWR

  9. Neural Network for NCAR LW Radiation NN characteristics • 220 Inputs: • Profiles: temperature; humidity; ozone, methane, cfc11, cfc12, & N2Omixing ratios, pressure, cloudiness, emissivity • Relevant surface characteristics: surface pressure, upward LW flux on a surface - flwupcgs • 33 Outputs: • Profile of heating rates (26) • 7 LW radiation fluxes: flns, flnt, flut, flnsc, flntc, flutc, flwds • Hidden Layer:One layer with 90 neurons (may go up to 150-200) • Training: • Training Data Set: Subset of about 100,000 instantaneous profiles simulated by CAM-2 for the 1-st year • Training time: about 10 days (SGI workstation) • Training iterations: 4,653 • Validation on Independent Data: • Validation Data Set (independent data): about 100,000 instantaneous profiles simulated by CAM-2 for the 2-nd year NNs for NCAR LWR

  10. NN Approximation Accuracy and Performance vs. Original Parameterization Comparisons with ECMWF NN1) 1)ECMWF NN approximation consists of the battery of 40 NNs Operational at ECMWF since October 2003 NNs for NCAR LWR

  11. Errors and Variability Profiles Bias and RMSE profiles in  Bias and RMSE Profiles in K/day Bias = 3.  10-5 K/day RMSE = 0.38 K/day PRMSE = 0.29 K/day PRMSE = 0.25 K/day Bias - dashed RMSE - Solid NNs for NCAR LWR

  12. Errors and Variability Profiles (NASA) Bias and RMSE Profiles in K/day Bias and RMSE profiles in  NNs for NCAR LWR

  13. NN Approximation Accuracy: Typical P  626 mb Bias in NN HR K/d HR in K/d HR in K/d RMSE in Error Distribution Error in K/d HR in K/d NNs for NCAR LWR

  14. Zonal Mean – Approximation Bias -0.1 -0.05 0. 0.05 0.1 K/day NNs for NCAR LWR

  15. Individual Profiles Black – Original Parameterization Red – NN with 90 neurons Blue – NN with 150 neurons PRMSE = 0.47 & 0.34 K/day PRMSE = 0.16 & 0.08 K/day PRMSE = 0.18 & 0.10 K/day NNs for NCAR LWR

  16. Individual Profiles Black – Original Parameterization Red – NN with 90 neurons Blue – NN with 150 neurons PRMSE = 0.14 & 0.07 K/day PRMSE = 0.05 & 0.04 K/day PRMSE = 0.11 & 0.06 K/day NNs for NCAR LWR

  17. NCAR CAM-2: 10 YEAR EXPERIMENTS • CONTROL: the standard NCAR CAM version (available from the CCSM web site) with the original Long-Wave Radiation (LWR) (e.g. Collins, JAS, v. 58, pp. 3224-3242, 2001) • LWR/NN: the NCAR CAM version with NN approximation of the LWR (Krasnopolsky, Fox-Rabinovitz, and Chalikov, to be submitted in March 2004; and Fox-Rabinovitz, Krasnopolsky, and Chalikov, to be submitted in April 2004) NNs for NCAR LWR

  18. CONSERVATION PROPERTIES(Global Annual Means) NNs for NCAR LWR

  19. Zonal MeanVertical Distributions and Differences Between the Experiments NNs for NCAR LWR

  20. NCAR CAM-2 Zonal Mean Heating Rates10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b), contour = 0.1 K/day all in K/day NNs for NCAR LWR

  21. NCAR CAM-2 Zonal Mean U10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b), contour 1 m/sec all in m/sec NNs for NCAR LWR

  22. NCAR CAM-2 Zonal Mean V10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b), contour 0.1 m/sec all in m/sec NNs for NCAR LWR

  23. NCAR CAM-2 Zonal Mean Temperature10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b), contour 1K all in K NNs for NCAR LWR

  24. NCAR CAM-2 Zonal RMSE for Heating Rates10 Year Average – Original LWR Parameterization - NN Approximation - RMS Difference (a) – (b), contour 0.1K/day all in K/day NNs for NCAR LWR

  25. NCAR CAM-2 Zonal RMSE for Temperature10 Year Average – Original LWR Parameterization - NN Approximation - RMS Difference (a) – (b), contour 1K all in K NNs for NCAR LWR

  26. Horizontal Distributions of Model Diagnostics NNs for NCAR LWR

  27. NCAR CAM-2 Surface Pressure 10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b), all in hPa NNs for NCAR LWR

  28. NCAR CAM-2 Total Cloudiness10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b), all in fractions NNs for NCAR LWR

  29. NCAR CAM-2 Total Precipitation10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b), all in mm/day NNs for NCAR LWR

  30. NCAR CAM-2 FLNT(OLR)10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b) all in W/m2 NNs for NCAR LWR

  31. NCAR CAM-2 LWR Heating Rates (near 850 hPa)10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b) all in K/day NNs for NCAR LWR

  32. Horizontal Distributions of Model Prognostics NNs for NCAR LWR

  33. NCAR CAM-2 Temperature (near 850 hPa) 10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b) all in K NNs for NCAR LWR

  34. NCAR CAM-2 Temperature (near 200 hPa) 10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b) all in K NNs for NCAR LWR

  35. NCAR CAM-2 U(near 850 hPa) 10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b) all in m/sec NNs for NCAR LWR

  36. NCAR CAM-2 U(near 200 hPa) 10 Year Average – Original LWR Parameterization - NN Approximation - Difference (a) – (b) all in m/sec NNs for NCAR LWR

  37. CONCLUSIONS: • The proof of concept: Application of MLT/ NN for fast and accurate approximation of model physics has been successfully demonstrated for the NCAR CAM LWR parameterization. • NN approximation of the NCAR CAM LWR is 80 times faster and very close to the original LWR parameterization • The simulated diagnostic and prognostic fields are very close for the parallel NCAR CAM climate runs with NN approximation and the original LWR parameterization • The conservation properties are very well preserved • A solid scientific foundation is laid for development of MLT/NN approximations for other NCAR CAM physics components or a complete set of MLT/”Neuro-Physics”. Such a focused effort will result in a complete reexamination of the model physics calculations. NNs for NCAR LWR

  38. OTHER POTENTIAL FUTURE DEVELOPMENTS • The effort can be extended in the future to reexamination of calculations for: other CCSM components; and chemical and physical components of chemistry transport models • MLT/NN can be applied for future development of new more sophisticated and time consuming parameterizations and even superparameterizations • MLT/SVM approximations should be explored and compared to those of NN • A feasibility study on MLT approximation of model dynamics will be conducted • Developing MLT approximations for other time-consuming “bottleneck” procedures (e.g., solvers, iterations, inversions, transformations, etc.) NNs for NCAR LWR

More Related