Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Improving Physical Parametrizations in Climate Models using Machine Learning


Published on

This talk was given at George Mason University on October 3, 2018.

Published in: Science
  • I'd advise you to use this service: ⇒ ⇐ The price of your order will depend on the deadline and type of paper (e.g. bachelor, undergraduate etc). The more time you have before the deadline - the less price of the order you will have. Thus, this service offers high-quality essays at the optimal price.
    Are you sure you want to  Yes  No
    Your message goes here
  • Over the past few days I've been telling you about a NEW 5-second "water hack" that crushes food cravings and melts 62lbs of fat... I know the video is long and you probably don't have time to watch the whole thing, so... I had it transcribed for you, so you can read it whenever you have a few minutes... ❤❤❤
    Are you sure you want to  Yes  No
    Your message goes here
  • Get Paid To Manage Facebook Fan Pages! Facebook Fan Page Workers Required - Start Immediately. ★★★
    Are you sure you want to  Yes  No
    Your message goes here
  • Be the first to like this

Improving Physical Parametrizations in Climate Models using Machine Learning

  1. 1. Improving Physical Parametrizations in Climate Models using Machine Learning Noah Brenowitz October 3, 2018 George Mason University
  2. 2. Acknowledgments Nathan Kutz (Applied Math) Chris Bretherton (Applied Math and Atmos. Sciences)
  3. 3. What is a weather model? If it is true, as any scientist believes, that subsequent states of the atmosphere develop from preceding ones according to physical laws, one will agree that the necessary and sufficient conditions for a rational solution of the problem of meteorological prediction are the following: 1. One has to know with sufficient accuracy the state of the atmosphere at a given time. 2. One has to know with sufficient accuracy the laws according to which one state of the atmosphere develops from another. - Bjerknes (1904) Vilhelm Bjerknes
  4. 4. High resolution models produce realistic clouds Giga-LES (Khairoutdinov et al. 2009) 100x100 km with 100 m grid
  5. 5. Can only solve at coarse resolution Image from NOAA
  6. 6. Coarse resolution equations Apparent heating (K/day) Apparent moistening (g/kg/day) SW+ LW radiation, latent heating, etc
  7. 7. How we usually build parametrizations High resolution simulation or observations
  8. 8. Climate models have biases in mean state CMIP5 (models) GPCP (observations) Hwang and Frierson (2013)
  9. 9. …and in variability (e.g. MJO) Observations Models Jiang, X. et al. (2015)
  10. 10. Parametrization is a function approximation problem Machine Learning Q1 Q2Q, T, U, V, … Q3
  11. 11. Machine learning builds black boxes • Many 1000s of parameters • Need a lot of data • Designed to be trained not interpreted • Examples: Decision trees, neural networks, support vector machines Easy to tune/train Easy to interpret Many parameters Few parameters
  12. 12. Optimum Parametrization Physical-based parametrizations might “orbit” reality
  13. 13. Training models with data
  14. 14. Training black boxes from data • Step 1: Training data • Step 2: Flexible model • Step 3: Supervision (what is error?) • Step 4: Train (minimize error) • Step 5: Test on new data
  15. 15. Emulation of existing parametrizations Original model Neural network Radiation parametrizations
  16. 16. Limited area cloud resolving models from Muller and Held (2012)
  17. 17. Neural networks can diagnose cloud fields from: Krasnopalsky, et al. (2013)
  18. 18. Global Cloud Resolving Models DYAMOND project • Inter-comparison of 8 GCRMs • 40 day simulations • 1 – 5 km resolution • 3 - 6 hourly outputs for 3D fields No plans to output tendency information!
  19. 19. But no previous prognostic tests with CRM data
  20. 20. Single Column Model Prognostic tests of CRM-trained neural network parametrization
  21. 21. Near-global aqua-planet (NG-Aqua) simulation generated by the System for Atmospheric Modeling (Δ𝑥 = 4km)
  22. 22. Coarse-grain data to 160 km boxes Training regionTesting region Coarse-graining A B C
  23. 23. Machine learning inputs 𝜙𝑖 𝑛 Preprocessing: concatenate center/scale(𝑥𝑖, 𝑦𝑖)
  24. 24. Training Approach 1 1. Use finite differences to compute residual tendencies 2. Train neural network: q, s, SHF, LHF, TOA Q1, Q2 Neural Network
  25. 25. The diagnostic performance is good! Neural Network Q1 (finite diff.) 𝑅2 ≈ .50
  26. 26. What about prognostic performance? Single column dynamics
  27. 27. Uh oh…temperature = 1035 K after 1 day Time (d) p (hPa) Is this why most past studies only show diagnostic results?
  28. 28. Is fitting the tendencies the right approach? • Assumes that model dynamics are continuous in time • But they are not (Donahue and Caldwell, 2018) • Assumes moist physics tendencies are available and accurate • Not true for DYAMOND outputs • Not true for observations • Does not ensure good predictions over many time steps
  29. 29. Fitting the approximate Q1 and Q2 is equivalent to minimizing one-step error
  30. 30. …but that does not ensure longer term performance
  31. 31. The scheme is now stable Simulated time series at x=1000 km, y=5198 km
  32. 32. Matches NG-Aqua better than CAM Community Atmosphere Model Version 5 (CAM5) Single Column Mode (default physics, no chemistry) Humidity Anomaly (from true zonal mean ) (g/kg)
  33. 33. Temperature Anomaly (K)
  34. 34. Implementation in a GCM Weather forecasting tests
  35. 35. Coarse Resolution Atmospheric Model Coarse resolution model (cSAM) • System for Atmospheric Modeling (SAM) • 160 km resolution • ”Dry” anelastic dynamics • Advection of water vapor • Virtual temperature effect on buoyancy • Damping + diffusion • Meridionally varying Coriolis force • Double precision important for mass conservation! Neural network • 3 layers of 256 neurons each • Trained with full global NGAqua data
  36. 36. Estimating large scale forcing for training • SAM has advection and diffusion • To compute known forcing using SAM: 1. Initialize SAM with data at time t: x(t) 2. Evolve forward for 10 minutes 3. Sam Forcing = (x(t+10 min) – x(t))/10 min • Could also account for radiation and other model physics
  37. 37. 10 Day simulation with NN + SAM at 160 km
  38. 38. NN improves the forecast accuracy
  39. 39. ITCZ narrows in the simulation
  40. 40. Zonal mean of vertical velocity narrow
  41. 41. Potential Cause: Little vertical momentum mixing in tropics Zonal Momentum
  42. 42. Solution: parametrize momentum source?
  43. 43. Another Problem: Loss of stochasticity Net Precipitation at 1 day
  44. 44. One solution: Stochastic Parametrization
  45. 45. Another possible solution: Data Assimilation • Filter the unresolvable scales in the training data using Digital Filter Initialization (Lynch 1997) • Defines model error (Kaas, et. al. 1998; Rodwell and Palmer, 2007) From: Rodwell and Palmer (2007)
  46. 46. Potential algorithm: Assimilate Data Train neural network Trained neural network + coarse resolution model Analyzed initial conditions + large-scale tendencies
  47. 47. Conclusions and Future Directions • Achievements • Neural network parametrization for unresolved physics • Numerically stable in single column and spatially extended mode • Only trained from coarse resolution snapshots of a GCRM • To do list • Predict momentum source (Q3) • More realistic training data (DYAMOND will be great resource) • Stochastic parametrization • Data Assimilation
  48. 48. References • Brenowitz, N. D. & Bretherton, C. S. Prognostic Validation of a Neural Network Unified Physics Parameterization. Geophys. Res. Lett. 17, 2493 (2018). • S. Rasp, M. S. Pritchard, P. Gentine, Deep learning to represent subgrid processes in climate models. Proc. Natl. Acad. Sci. U. S. A. 115, 9684–9689 (2018). Noah Brenowitz ( Contact me!
  49. 49. Neural networks are a popular machine learning model
  50. 50. Sometimes negative
  51. 51. Vertically integrated error is small
  52. 52. Lower bias in time-mean fields
  53. 53. Super-parametrized models graphic: Krueger and Bogenschutz
  54. 54. Mass conserving initial conditions • 4km SAM and 160km cSAM are on a staggered grid • Zonal velocity and meridional velocity are averages along the interfaces (40 grid points) • All other variables: averaged over full 160 km box (40^2 grid points)
  55. 55. Parametrized source on the equator Moistening Heating
  56. 56. Super-parametrized CAM (SPCAM) Inputs: u, v, w, q, T, SHF, LHF Outputs: Q1, Q2, TOA Radiative flux, Precip Used neural network. The diagnosed tendencies and precipitation match SPCAM, but they show no prognostic tests. …they do show nice prognostic results in a very recent manuscript.
  57. 57. Sensitivity to Hyper parameters
  58. 58. Split-in-time time stepping
  59. 59. The primitive equations Total Derivative Momentum conservation (zonal) Momentum conservation (meridional) Hydrostatic Balance: Vertical velocity Mass conservation
  60. 60. Increasing training window size decreases 64- step error
  61. 61. Precipitation matches patterns match truth
  62. 62. We want the scheme to make good predi
  63. 63. Is using more layers another way to break the stability deadlock? arXiv (2018)
  64. 64. This really nice work, but there are some issues • Still uses true tendencies as inputs • Not clear it will work with GCRM outputs • Takes 8 hours to train (vs 5-20 minutes for our approach)
  65. 65. Example: Decision Trees Moist boundary layer? no Yes Stable temperature profile? Yes No Strong convection Not much convection Not much convection
  66. 66. …but can be hard to interpret Very similar to a lookup table…non-continuous outputs