.. _ch:networks:integrators: *************** ODE Integrators *************** Available integrators ===================== We use a high-order implicit ODE solver for integrating the reaction system. A few alternatives, including first order implicit and explicit integrators are also provided. Internally, the integrators use different data structures to store the integration progress, and each integrator needs to provide a routine to convert from the integrator’s internal representation to the ``burn_t`` type required by the ``actual_rhs`` and ``actual_jac`` routine. .. index:: INTEGRATOR_DIR .. note:: The integrator is chosen at compile-time using the ``INTEGRATOR_DIR`` variable in the makefile. Presently, allowed integrators are: * ``BackwardEuler``: an implicit first-order accurate backward-Euler method. An error estimate is done by taking 2 half steps and comparing to a single full step. This error is then used to control the timestep by using the local truncation error scaling. Optionally, the user can disable error estimation and force a single-step backward-Euler integration by setting `integrator.do_single_step = 1`. * ``ForwardEuler``: an explicit first-order forward-Euler method. This is meant for testing purposes only. No Jacobian is needed. * ``QSS``: the quasi-steady-state method of :cite:`mott_qss` (see also :cite:`guidry_qss`). This uses a second-order predictor-corrector method, and is designed specifically for handling coupled ODE systems for chemical and nuclear reactions. However, this integrator has difficulty near NSE, so we don't recommend its use in production for nuclear astrophysics. .. index:: integrator.use_circle_theorem * ``RKC``: a stabilized explicit Runge-Kutta-Chebyshev integrator based on :cite:`sommeijer_rkc_1998`. This does not require a Jacobian, but does need to estimate the spectral radius of the system, which is done internally. This works for moderately stiff problems. The spectral radius is estimated by default using the power method, built into RKC. Alternately, by setting ``integrator.use_circle_theorem=1``, the `Gershgorin circle theorem <https://en.wikipedia.org/wiki/Gershgorin_circle_theorem>`_ is used instead. .. index:: integrator.use_jacobian_caching * ``VODE``: the VODE :cite:`vode` integration package. We ported this integrator to C++ and removed the non-stiff integration code paths. .. note:: The VODE integrator uses Jacobian caching when run on a CPU by default. This can be disabled at runtime by setting ``integrator.use_jacobian_caching = 0``. On GPUs, we disable Jacobian caching due to the increased memory needs. Jacobian caching on GPUs can be enabled by explicitly setting the build parameter ``USE_JACOBIAN_CACHING=TRUE``. We recommend that you use the VODE solver, as it is the most robust. .. index:: integrator.scale_system .. note:: The runtime parameter ``integrator.scale_system`` will scale the internal energy that the integrator sees by the initial value of :math:`e` to make the system :math:`\mathcal{O}(1)`. The value of ``atol_enuc`` will likewise be scaled. This works for both Strang and simplified-SDC. For the ``RKC`` integrator, this is enabled by default. For most integrators this algebraic change should not affect the output to more than roundoff, but the option is included to allow for some different integration approaches in the future. This option currently does not work with the ForwardEuler or QSS integrators. Timestep selection ================== All of the integrators will select the timestep internally to meet the tolerances. There are 2 controls that affect timestepping: * ``integrator.ode_max_dt`` : sets the maximum allowed timestep * ``integrator.ode_max_steps`` : sets the maximum number of steps the integrator is allowed to take. If it exceeds this, then it will return an error. Linear algebra ============== All implicit integrators use the LINPACK LU decomposition routines. For the templated networks (``aprox13``, ``aprox19``, ...) the implementation is done using ``consexpr`` loops over the equations and no pivoting is allowed. .. index:: integrator.linalg_do_pivoting For the other networks (usually pynucastro networks), the implementation is provided in ``Microphysics/util/linpack.H`` and is templated on the number of equations. Pivoting can be disabled by setting ``integrator.linalg_do_pivoting=0``. Integration errors ================== .. important:: The integrator will not abort if it encounters trouble. Instead it will set ``burn_t burn_state.success = false`` on exit. It is up to the application code to handle the failure. The ``burn_t`` ``error_code`` field will provide an error code that can be used to interpret the failure. The current codes are: +-------+----------------------------------------------------------+ | code | meaning | +=======+==========================================================+ | 1 | success | +-------+----------------------------------------------------------+ | -1 | invalid inputs | +-------+----------------------------------------------------------+ | -2 | underflow in computing $\Delta t$ | +-------+----------------------------------------------------------+ | -3 | spectral radius estimation did not converge | +-------+----------------------------------------------------------+ | -4 | too many steps needed | +-------+----------------------------------------------------------+ | -5 | unable to meet the accuracy demanded by the tolerances | +-------+----------------------------------------------------------+ | -6 | non-convergence in the corrector iteration | +-------+----------------------------------------------------------+ | -7 | LU decomposition failed | +-------+----------------------------------------------------------+ | -100 | entered NSE | +-------+----------------------------------------------------------+ Tolerances ========== Tolerances dictate how accurate the ODE solver must be while solving equations during a simulation. Typically, the smaller the tolerance is, the more accurate the results will be. However, if the tolerance is too small, the code may run for too long, the ODE solver will never converge, or it might require at timestep that underflows. .. index:: integrator.rtol_spec, integrator.rtol_enuc, integrator.atol_spec, integrator.atol_enuc There are separate tolerances for the mass fractions and the energy, and there are both relative and absolute tolerances which act together. The tolerances are: * ``integrator.rtol_spec`` : the relative tolerance for the species (mass fractions when running with Strang and partial densities when running with SDC). * ``integrator.rtol_enuc`` : the relative tolerance on the energy (specific internal energy when running with Strang, internal energy density when running with SDC). * ``integrator.atol_spec`` : the absolute tolerance for the species (this is always interpreted in terms of mass fraction and the appropriate density weighting will be added for SDC). * ``integrator.atol_enuc`` : the absolute tolerance for energy -- this is generally not interesting, since the energy is so large and therefore best served via a relative tolerance. The tolerances are combined, e.g. for species, as: .. math:: \epsilon_{\mathrm{total}, k} = \epsilon_\mathrm{abs} + \epsilon_\mathrm{rel} |X_k| so if the mass fraction, $X_k$, is very small, then the absolute tolerance will set the error that the integrator tries to achieve. If the mass fraction is large, $\mathcal{O}(1)$, then the relative tolerance dominates. Some suggestions when setting tolerances: .. index:: integrator.X_reject_buffer * If a burn does not converge with one type of Jacobian (analytic or numerical) then it may do better with the other type. This can be automated via the ``integrator.use_burn_retry`` mechanism described above. * Sometimes a burn completes better if the absolute tolerances are made even smaller -- this will require the integrator to track trace species better which can help with equilibrium better. * The VODE integrator has additional logic meant to ensure that species don't change too much per timestep. This is controlled by ``integrator.X_reject_buffer``. If a species $k$, has a mass fraction $X_k > \mbox{X_reject_buffer} \cdot \mbox{atol_spec}$ then we reject a VODE timestep if the mass fraction changes by more than a factor of 4 in a single VODE timestep and we try again. This is all done internally to VODE. Making ``X_reject_buffer`` larger will allow it to ignore more trace species. Below is a comparison of how the tolerances affect the nucleosynthesis. This is run using ``burn_cell`` and the ``aprox13`` network. Four separate runs were done, using tolerances of $10^{-3}$, $10^{-5}$, $10^{-8}$, and $10^{-12}$ (all 4 tolerance parameters were set to the same value). The run with the tightest tolerances ($10^{-12}$) was taken as the reference and relative errors were computed with respect to it. The scripts for this are in ``Microphysics/unit_test/burn_cell/compare_tolerances/``. .. _fig:tolerances: .. figure:: tolerance-compare.png :alt: Relative error in mass fractions :width: 100% Relative error of runs with varying tolerances as compared to a run with an ODE tolerance of :math:`10^{-12}`. We see that using a tolerance of $10^{-5}$ generally gives reasonable mass fractions. Looser than this can produce large errors. Controlling Species $\sum_k X_k = 1$ ==================================== .. index:: integrator.renormalize_abundances, integrator.SMALL_X_SAFE, integrator.do_species_clip The ODE integrators don't know about the constraint that $$\sum_k X_k = 1$$ so this is only going to be preserved to the level that the integrator tolerances allow. There are a few parameters that help enforce this constraint on the intermediate states during the integration. * ``integrator.renormalize_abundances`` : this controls whether we renormalize the abundances so that the mass fractions sum to one during a burn. This has the positive benefit that in some cases it can prevent the integrator from going off to infinity or otherwise go crazy; a possible negative benefit is that it may slow down convergence because it interferes with the integration scheme. Regardless of whether you enable this, we will always ensure that the mass fractions stay positive and larger than some floor ``small_x``. This option is disabled by default. * ``integrator.SMALL_X_SAFE`` : this is the floor on the mass fractions. The default is ``1.e-30``. * ``integrator.do_species_clip`` : this enforces that the mass fractions all in $[\mathtt{SMALL\_X\_SAFE}, 1.0]$. This is enabled by default. Retry Mechanism =============== .. index:: integrator.ode_max_steps Integration can fail for a number of reasons. Some of the errors you may see are: 1. Not enough steps allowed (``integrator.ode_max_steps``) 2. The timestep selected by the integrator is too small (comparable to roundoff) 3. The final abundances do not sum to 1. There can be a number of reasons for these failures, including: * The Jacobian is not accurate enough This can lead to issues 1 or 2 above * The integrator is not appropriate for the thermodynamic conditions For example, the RKC integrator may be working too hard, leading to issue 1. * The tolerances you are requesting are too tight This can lead to issues 1 or 2 above * The tolerances (in particular, ``integrator.atol_spec``) are too loose This can lead to issue 3 above * The evolution is entering NSE This can lead to issue 1. The ``integrator()`` function that calls the actual integrator drive for the choice of integrator allows for a retry if a burn failure was detected. This is enabled by setting :: integrator.use_burn_retry = 1 This will call the same integrator again, restarting from the initial conditions but with a different choice of tolerances and Jacobian. The runtime parameters that come into play when doing the retry are: * ``retry_swap_jacobian`` : do we swap that Jacobian type for the retry (i.e. use the numerical Jacobian if we try the analytic Jacobian for the first attempt) * ``retry_rtol_spec`` : relative tolerance for the species on retry * ``retry_rtol_enuc`` : relative tolerance for the energy on retry * ``retry_atol_spec`` : absolute tolerance for the species on retry * ``retry_atol_enuc`` : absolute tolerance for the energy on retry .. note:: If you set any of the retry tolerances to be less than $0$, then the original (non-retry) tolerance is used on retry. The default value for all of the retry tolerances is $-1$, which means the same tolerances are used on retry unless you override them at runtime. .. tip:: Sometimes a simulation runs best if you set ``integrator.ode_max_steps`` to a small value (like ``10000``) and start with the analytic Jacobian (``integrator.jacobian = 1``) and then use the retry mechanism to swap the Jacobian on any zones that fail.