FLINT

Fortran Library for numerical INTegration of differential equations

Author

Bharat Mahajan

Copyright

Copyright 2021 Bharat Mahajan

The initial code was written by Bharat Mahajan at Odyssey Space Research LLC, Houston, TX as part of the work under contract no. 80JSC017D0001 with NASA-Johnson Space Center. FLINT source code is licensed under the Apache License, Version 2.0 (the "License") found in LICENSE file contained in this distribution.

The coefficients for DOP853 method were derived by Ernest Hairer. His original codes are available at http://www.unige.ch/~hairer/software.html. The coefficients for Verner65E and Verner98R methods were derived by Jim Verner, and are available at http://people.math.sfu.ca/~jverner/.

Introduction

FLINT is a modern object-oriented Fortran library that provides four adaptive step-size explicit Runge-Kutta (ERK) methods of order 5, 6, 8, and 9 along with dense-output and multiple event-detection support for each of the methods. The code is written such that any other ERK method can be implemented by including its coefficients with minimum changes required in the code. The DOP853 integrator is the default method chosen, and its implementation is hand-optimized specific to its coefficients. For other integrators, a generic routine for step-integration is implemented. This generic routine supports both FSAL and non-FSAL methods. Dense output is supported with delayed interpolation. When interpolation is enabled, FLINT computes the interpolation coefficients during the integration and stores them in internal memory. Thereafter, the interpolation method can be used any number of times to find the solution values at any user-specified grid within the initial and final conditions. Interpolation is much faster than integration, as the coefficients are all precomputed during the integration. Multiple event detection is supported for each integrator along with many features such as event root-finding, event step-size, event actions. In a nutshell, the features are:

• Modern object-oriented, thread-safe, and optimized Fortran code
• 4 Adaptive-step (fixed step-size also supported) ERK integrators: DOP54, DOP853, Verner98R, Verner65E
• Any other ERK method can be implemented by including their coefficients
• Dense output with delayed interpolation (integrate once, interpolate as many times)
• Multiple event-detection as well as finding location of events using root-finding (Brent's algorithm) with static and dynamic event masking
• Ability to set a maximum delay (referred to here as event step-size) after which events are guaranteed to be detected
• Ability to restart the integration or change solution on the detection of events
• Stiffness detection

Performance benchmark against Julia's Differential Equation package

The latest FLINT code (compiled using Intel Fortran compiler) is tested against Julia's DifferentialEquations package (https://diffeq.sciml.ai/stable/) and FLINT appears to be faster with and without event detection in the most cases as shown in the following screenshots. The Julia test code along with results are provided in the tests folder on FLINT's GitHub repository https://github.com/princemahajan/FLINT.

• Three-Body problem propagation

• Lorenz equations integration

Installation

FLINT is tested with ifort (18.0.2) compiler from Intel Parallel Studio XE Composer for Windows 2016 integrated with Microsoft Visual Studio Community version 2017. Some testing is also done with MinGW-W64 gfortran (gcc 8.1.0). Doxyfile is provided for generating extensive API documentation using Doxygen. FLINT has no dependency on any other library. The CMakeLists file is provided to generate Visual Studio projects or makefiles on Windows and Linux using cmake. Additionally, it generates cmake config files to easily link FLINT using the find_package() command. The steps to link FLINT in cmake-based projects are:

• In cmake GUI or command-line, set FLINT_INSTALL_LIB_DIR to the desired directory, where the compiled library, modules, and cmake config files will be installed.
• In cmake GUI or command-line, set FLINT_INSTALL_BIN_DIR to the desired directory, where the compiled test executables of FLINT will be installed.
• In your project CMakeLists.txt, insert

      find_package(FLINT REQUIRED 0.9 CONFIG 
          PATHS "<SAME_PATH_AS_IN_FLINT_INSTALL_LIB_DIR>")
      target_link_libraries(<YOUR_TARGET_NAME> FLINT::FLINT)

How to Use

See the test program files, test.f90 and DiffEq.f90, in tests folder for a comparatively complex example problem that uses multiple events. Following are the steps in brief for a simpler problem:

1. Create a differential equation system class by providing differential equation function, events function (if any), and parameters (if any).

    use FLINT
    implicit none

    type, extends(DiffEqSys) :: TBSys
        real(WP) :: GM = 398600.436233_WP  ! parameter constant
    contains
        procedure :: F => TwoBodyDE     ! Differential equations
        procedure :: G => SampleEventTB ! events, use only if needed
    end type TBSys

    contains

    function TwoBodyDE(me, X, Y, Params)
        implicit none
        class(TBSys), intent(in) :: me !< Differential Equation object
        real(WP), intent(in) :: X
        real(WP), intent(in), dimension(:) :: Y
        real(WP), intent(in), dimension(:), optional :: Params
        real(WP), dimension(size(Y)) :: TwoBodyDE

        TwoBodyDE(1:3) = Y(4:6)
        TwoBodyDE(4:6) = -me%GM/(norm2(Y(1:3))**3)*Y(1:3) ! Two-body orbit diffeq
    end function TwoBodyDE

    subroutine SampleEventTB(me, X, Y, EvalEvents, Value, Direction, LocEvent, LocEventAction)
        implicit none
        class(TBSys), intent(in) :: me !< Differential Equation object            
        real(WP), intent(in) :: X
        real(WP), dimension(:), intent(inout) :: Y
        integer, dimension(:), intent(in) :: EvalEvents
        real(WP), dimension(:), intent(out) :: Value
        integer, dimension(:), intent(out) :: Direction
        integer, intent(in), optional :: LocEvent
        integer(kind(FLINT_EVENTACTION_CONTINUE)), intent(out), optional :: LocEventAction

        if (EvalEvents(1)==1) Value(1) = Y(2) ! detect y-crossing
        if (EvalEvents(2)==1) Value(2) = Y(1) ! detect x-crossing    

        Direction(1) = 1  ! detect -ve to +ve transitions for y coordinate
        Direction(2) = -1 ! detect +ve to -ve transitions for x coordinate

        ! Set actions for each event if located
        if (present(LocEvent) .AND. present(LocEventAction)) then
            if (LocEvent == 1) then
                LocEventAction = IOR(FLINT_EVENTACTION_CONTINUE, &
                    FLINT_EVENTACTION_MASK)
            else if (LocEvent == 2) then
                ! Mask the event-2 after first trigger            
                LocEventAction = IOR(FLINT_EVENTACTION_CONTINUE, &
                        FLINT_EVENTACTION_MASK)
            end if   
        end if
    end subroutine SampleEventTB  

2. Initialize the differential equation and ERK class objects for using Runge-Kutta integrators.

        use FLINT

        type(ERK_class) ::  erk
        type(TBSys)     :: diffeq
        
        diffeq%n = 6        ! Number of 1st-order differential equations
        diffeq%m = 2        ! Number of events
        
        ! Initialize the ERK object for 10000 max steps, DOP54 method with abs. tol. 1e-12, rel. tol. 1e-9
        ! and turn on interpolation and events detection    
        call erk%Init(diffeq, 10000, Method=ERK_DOP54, ATol=[1e-12], RTol=[1e-9],&
        InterpOn=.TRUE.,EventsOn=.TRUE.)

3. Call the Integrate subroutine for performing the integration if init was successful. Note if interpolation is enabled, then the IntStepsOn option for computing the states at integrator's natural step-size must not be set to True.

    integer :: stiffstatus    
    real(WP) :: x0, xf    
    real(WP), dimension(6) :: y0, yf
    real(WP), allocatable, dimension(:,:) :: EventData ! allocated by Integrate
    
    stiffstatus = 1 ! detect stiffness and terminate if equations are stiff    
    x0 = 0.0    
    y0 = [6400.0_wp,0.0_wp,0.0_WP, 0.0_WP,7.69202528825512_WP,7.69202528825512_WP]
    xf = 161131.68239305308_WP      
    
    ! Call Intgerate with final solution in yf, no initial step-size given, events-related
    ! data (event-id, x value, y state) in EventData, and all events are detected
    if (erk%status == FLINT_SUCCESS) then    
        call erk%Integrate(x0, y0, xf, yf, StepSz=0.0, IntStepsOn=.FALSE.,&
        EventStates=EventData, EventMask = [.TRUE.,.TRUE.],StiffTest=stiffstatus)
    end if    

4. Call the Interpolate function for computing solution on the desired grid of x values. The last parameter must be specified as True if user wants FLINT to keep the internal storage for calling Interpolate again. Otherwise, the internal storage is deleted and the user must integrate the equations again before calling Interpolate.

    real(WP), dimension(:), allocatable :: Xarr1, Xarr2
    real(WP), dimension(6,:) :: Yarr1, Yarr2
    
    Xarr1 = [(x0 + (xf-x0)/10*i, i=0,9)]    ! grid-1 with 10 points
    Xarr2 = [(x0 + (xf-x0)/1000*i, i=0,999)]    ! grid-2 with 1000 points
    allocate(Yarr1(6,size(Xarr1)))    ! allocate solution storage
    allocate(Yarr2(6,size(Xarr2)))        
    
    if (erk%status == FLINT_SUCCESS) then
        ! interpolate and keep the internal storage for further calls
        call erk%Interpolate(Xarr1, Yarr1, .TRUE.)
        ! After this interpolation, delete the internal storage (default).
        call erk%Interpolate(Xarr2, Yarr2, .FALSE.) 
    end if    
    
    ! print the solutions
    print *, 'Solution at grid-1'    
    print *, Yarr1
    print *, 'Solution at grid-2'        
    print *, Yarr2
    ! print the event data
    print *, EventData(1,:)     ! Time at which events occured
    print *, EventData(2:7,:)   ! corresponding position and velocity states
    print *, Eventdata(8,:)     ! Event-ID number

For all the FLINT status codes and options supported by Init, Integrate, and Interpolate procedures along with the interfaces for user-supplied functions, see the FLINT_base module in FLINT_base.f90 file.

Testing

The following tests were performed using the very first verion of FLINT and they are not updated for the latest version. In the tests, an orbit is propagated for 4 orbital periods and the integration is repeated 5000 times. The tables give the total time for 5000 propagations and all other testing parameters are given for each integration. The relative tolerance is taken as 1e-11 and absolute tolerance as 1e-14 in all the cases except when explicitly mentioned otherwise. Interpolation is used to capture 1000 points per orbit period. Note that the delayed interpolation feature of FLINT is not used in these results.

The initial conditions in Cartesian coodinates used are as follows:

• Two-Body circular Earth orbit (Units: km, sec)
  • GM: 398600.436233
  • Y0 = [6400.0,0.0,0.0, 0.0,5.58037857139,5.58037857139]
• Two-Body elliptic Earth orbit (Units: km, sec)
  • GM: 398600.436233
  • Y0 = [6400.0,0.0,0.0,0.0,7.69202528825512,7.69202528825512]
• Arenstorf orbit in rotating frame (Units: nondimensional)
  • mass-ratio: 0.012277471
  • time-period: 17.0652165601579625588917206249
  • Y0 = [0.994, 0.0, 0.0_wp, -2.00158510637908252240537862224]

Note that JDOP853 and JDDEABM are modern fortran implementations of DOP853 and DDEABM by Jacob Williams, and are available at https://github.com/jacobwilliams. The plots are generated using https://github.com/jacobwilliams/pyplot-fortran.

• Two-Body circular Earth orbit

• Two-Body circular Earth orbit with interpolation

• Two-Body elliptic Earth orbit

• Two-Body elliptic Earth orbit with interpolation

• Arenstorf orbit

• Arenstorf orbit with interpolation

For relative tol=1e-9 and abs tol=1e-12, the Arenstorf orbit propagated for 4 periods by JDOP853, JDDEABM and FLINT's Verner98R are shown below. Verner98R diverges much slower than DOP853 and DDEABM for this orbit and tolerance values.

The following figure shows the multiple event detection capability of FLINT, in which the X-crossings in decreasing and Y-crossings in increasing direction are detected and reported to the user.

References

• Hairer, Ernst, Norsett, Syvert P., Wanner, Gerhard, Solving Ordinary Differential Equations I, Springer-Verlag Berlin Heidelberg, 2nd Ed., 1993.