Open software for Astronomical Data Analysis

@ NASA Goddard

Transcript

1. OPEN
   
   
   SOFTWARE
   
   
   FOR
   
   
   ASTRONOMICAL
   
   
   DATA ANALYSIS
   by Dan Foreman-Mackey
   

   View Slide

2. View Slide

3. open software for astrophysics
   0
   

   View Slide

4. credit: Adrian Price-Whelan
   / /
   data: SAO/NASA ADS
   

   View Slide

5. View Slide

6. many fundamental software packages
   
   
   have a shockingly small number of
   
   
   maintainers.
   

   View Slide

7. 7
   credit: Adrian Price-Whelan
   

   View Slide

8. * astronomical software can be
   very high impact
   
   
   * we should think about career
   trajectories & mechanisms for
   supporting this work
   

   View Slide

9. View Slide

10. case study: gaussian processes
    1
    

    View Slide

11. °0.6
    °0.3
    0.0
    0.3
    0.6
    raw [ppt]
    0 5 10 15 20 25
    time [days]
    °0.30
    °0.15
    0.00
    de-trended [ppt]
    N = 1000
    reference: DFM+ (2017)
    

    View Slide

12. °0.6
    °0.3
    0.0
    0.3
    0.6
    raw [ppt]
    0 5 10 15 20 25
    time [days]
    °0.30
    °0.15
    0.00
    de-trended [ppt]
    N = 1000
    reference: DFM+ (2017)
    

    View Slide

13. reference: Aigrain & DFM (2022)
    

    View Slide

14. reference: Aigrain & DFM (2022)
    

    View Slide

15. reference: Aigrain & DFM (2022)
    ignoring
    
    
    correlated
    
    
    noise
    accounting
    
    
    for
    
    
    correlated
    
    
    noise
    

    View Slide

16. reference: Aigrain & DFM (2022)
    

    View Slide

17. a Gaussian Process is a
    
    
    drop
    -
    in replacement for
    
    
    chi
    -
    squared
    

    View Slide

18. more details:
    
    
    Aigrain & Foreman-Mackey (2023)
    
    
    arXiv:2209.08940
    

    View Slide

19. View Slide

20. 7
    [1] model building
    
    
    [2] computational cost
    

    View Slide

21. reference: Luger, DFM, Hedges (2021)
    

    View Slide

22. [2] computational cost
    

    View Slide

23. 7
    [1] bigger/better computers
    
    
    [2] exploit matrix structure
    
    
    [3] approximate linear algebra
    
    
    [4] etc.
    

    View Slide

24. View Slide

25. View Slide

26. View Slide

27. View Slide

28. °0.6
    °0.3
    0.0
    0.3
    0.6
    raw [ppt]
    0 5 10 15 20 25
    time [days]
    °0.30
    °0.15
    0.00
    de-trended [ppt]
    N = 1000
    reference: DFM+ (2017)
    

    View Slide

29. reference: Gordon, Agol, DFM (2020) / tinygp.readthedocs.io
    

    View Slide

30. * a Gaussian Process is a drop
    -
    in
    replacement for chi squared
    
    
    * model building & computational
    cost are (solvable!) challenges
    
    
    * you should check out tinygp!
    

    View Slide

31. case study: probabilistic inference
    2
    

    View Slide

32. have:
    
    
    physics
    = >
    data
    

    View Slide

33. want:
    
    
    data
    = >
    physics
    

    View Slide

34. 7
    [1] physical models
    
    
    [2] legacy code
    

    View Slide

35. View Slide

36. number of parameters
    patience required
    a few tenish not outrageously many
    reference: DFM (priv. comm.)
    

    View Slide

37. number of parameters
    patience required
    emcee
    a few tenish not outrageously many
    reference: DFM (priv. comm.)
    

    View Slide

38. number of parameters
    patience required
    emcee
    a few tenish not outrageously many
    how things should be
    reference: DFM (priv. comm.)
    

    View Slide

39. View Slide

40. View Slide

41. View Slide

42. View Slide

43. 3.0 3.5 4.0 4.5 5.0
    Wavelength [micron]
    2.05
    2.10
    2.15
    2.20
    2.25
    2.30
    Transit Depth [%]
    Alderson et al. 2023
    Joint Fit (N = 50)
    reference: Soichiro Hattori, Ruth Angus, DFM,
    . . .
    (in prep)
    WASP-39b / NIRSpec
    

    View Slide

44. reference: Soichiro Hattori, Ruth Angus, DFM,
    . . .
    (in prep)
    showing 23 of the
    
    
    404 parameters
    
    
    (8 per channel + 4 shared)
    

    View Slide

45. how?
    

    View Slide

46. d(physics
    = >
    data) / dphysics
    

    View Slide

47. automatic differentiation
    aka “backpropagation”
    

    View Slide

48. View Slide

49. 7
    [1] physical models
    
    
    [2] legacy code
    

    View Slide

50. 7
    [1] domain
    -
    specif
    i
    c libraries
    
    
    [2] emulation
    

    View Slide

51. View Slide

52. * gradient
    -
    based inference using
    autodiff can improve eff
    i
    ciency
    
    
    * there are practical challenges
    with these methods in astro
    
    
    * of interest: domain
    -
    specif
    i
    c
    libraries & emulation
    

    View Slide

53. aside: JAX
    3
    

    View Slide

54. View Slide

55. import numpy as np
    
    
    def linear_least_squares(x, y)
    :
    
    A = np.vander(x, 2)
    
    
    return np.linalg.lstsq(A, y)[0]
    

    View Slide

56. import jax.numpy as jnp
    
    
    def linear_least_squares(x, y)
    :
    
    A = jnp.vander(x, 2)
    
    
    return jnp.linalg.lstsq(A, y)[0]
    

    View Slide

57. View Slide

58. open research practices
    4
    

    View Slide

59. View Slide

60. View Slide

61. View Slide

62. View Slide

63. View Slide

64. View Slide

65. View Slide

66. open software is foundational to
    astrophysics research
    
    
    there are opportunities at the
    interface of astro & applied f
    i
    elds
    
    
    there are ways you can participate
    & benef
    i
    t right away
    

    View Slide

67. 7
    I want to chat about…
    
    
    [1] your data analysis problems
    
    
    [2] building astronomical software
    
    
    [3] writing documentation & tutorials
    

    View Slide

68. get in touch!
    
    
    dfm.io
    
    
    github.com/dfm
    

    View Slide