.. _pitfalls: Frequent pitfalls encountered when doing Tkwant simulations =========================================================== Tkwant has been designed to be as simple to use and as automatic as possible. Yet, time-dependent simulations are not easy and require a working understanding of the underlying theory. One may encounter several problems that affect the accuracy or the validity of the simulations. Whenever possible, results should be validated with a known benchmark and the accuracy of the convergence (time-steps, manybody integral, absorbing boundary conditions) explicitly tested. In this section we list difficulties that are often encountered when performing simulations with Tkwant. We encourage users, especially inexperienced ones, to systematically check at least points 1 to 5 when starting a new type of simulation or moving to a new regime. The remaining points are less critical, but should also be checked for new problems, in case of doubt or persistent problems. As a general advise we recommend to check the standard output (stdout) for warning messages from Tkwant. The warnings in the log output are often a first hint to see where a calculation is getting stuck or what is missing. One can also increase the verbosity of the logged output with .. code-block:: python import tkwant import logging tkwant.logging.level = logging.INFO # possible levels (increasing verbosity): WARNING, INFO, DEBUG at the beginning of the Tkwant script. Further details are explained in the :ref:`logging` tutorial. Convergence of the manybody integral (1/2) ------------------------------------------ Problem ~~~~~~~ The manybody integral is poorly approximated (not enough sampling points and/or slowly converging integral) and the result inaccurate. Diagnostic ~~~~~~~~~~ Perform adaptive integral refinement of :math:`\texttt{tkwant.manybodyState()}` with the method: .. code-block:: python refine_intervals() at various stages of the calculation. If the results change, then the previous result were not converged. Without refinement, the numerical result is likely to be wrong, so we recommend to always use this method unless you have experience with numerical integration. Check the tutorial section `adaptive refinement and error estimate `_ for more details. Solution ~~~~~~~~ Try adding further calls to :math:`\texttt{refine_integrals()}` at different times in the simulation and check if the results change. Adding calls to this function always improve the accuracy but can be dear in computational time. At the initial time :math:`t=0` the manybody integral is refined by default, such that no additional call to :math:`\texttt{refine_integrals()}` is needed. Convergence of the many-body integral (2/2) ------------------------------------------- Problem ~~~~~~~ Adaptive refinement with :math:`\texttt{refine_intervals()}` blocks and endlessly adds points to refine the integral without convergence. Diagnostic ~~~~~~~~~~ In some problems, the many-body integral can converge very poorly due to a difficult integrant with e.g. very sharp resonances. A typical example is a very large quantum dot almost decoupled from its leads. Plotting the function that is integrated usually helps to identify the problem. For the integrator to work properly, the integrant should look resonnably smooth. The MPI parallelization makes it difficult to plot the integrand directly from the manybody solver, but Tkwant provides the auxiliary routine :math:`\texttt{tkwant.manybody.ManybodyIntegrand}` for diagnostic purpose. The following code snipped shows its use for the the density integrand which is plotted again momentum :math:`k` within the range :math:`kmin` and :math:`kmax` at a certain time snapshot. Note that one can only plot the contribution of a certain, lead and band to the manybody integral: .. code-block:: python import tkwant import kwant import numpy as np import matplotlib.pyplot as plt # user defined system definition syst = make_my_kwant_system().finalized() # define the momentum range in between which the integrand will be analyzed kmin = 0.01 kmax = np.pi/2 # specify the lead and band index interval = tkwant.manybody.Interval(lead=0, band=0, kmin=kmin, kmax=kmax) # define an operator density_operator = kwant.operator.Density(syst, sum=True) # define a time where to evaluate the integrand integrand = tkwant.manybody.ManybodyIntegrand(syst, interval, density_operator, time=100) # plot the integrand on some arbitrary points within [kmin, kmax] k = np.linspace(kmin, kmax) plt.plot(k, integrand.vecfunc(k)) Solution ~~~~~~~~ The difficulty often actually originates from the physical problem. The solution is often to sligthly change the problem to bring it to a simpler regime. For instance, in the case of a quantum dot, increasing the coupling to the leads will facilitates the convergence by making the resonances less narrow. In some contexts, increasing the Fermi energy in the lead can also help, see the discussion in section 10.1 of Ref. `[1] <#references>`__ (Figure 15). Presence of unincluded boundstates ---------------------------------- Problem ~~~~~~~ Manybody observables are wrong due to the negligence of existing boundstates. Diagnostic ~~~~~~~~~~ Check for the prensence of boundstates in your system with the function: .. code-block:: python boundstates_present() When this function returned :math:`\texttt{True}`, the system contains boundstates. These are not automatically handled by Tkwant but its up to the user to include them. Ignoring boundstates might lead to wrong results, at least for the density. Solution ~~~~~~~~ Read the tutorial :ref:`boundstates` for more information and to learn how to include them. Error in Hamiltonian construction or unphysical parameter regime ---------------------------------------------------------------- Problem ~~~~~~~ The initial Hamiltonian is wrong and/or the parameters are set to unphysical values. Diagnostic ~~~~~~~~~~ - Plot the Kwant system which defines the Hamiltonian via: .. code-block:: python kwant.plot() Check visually if all sites, couplings and leads appear as expected. - Plot the banstructure with the help of the `kwantSpectrum `_ package: .. code-block:: python import kwantspectrum as ks import kwant import numpy as np import matplotlib.pyplot as plt syst = my_kwant_system_with_leads().finalized() lead_index = 0 # change this number to the intended lead spec = ks.spectrum(syst.leads[lead_index]) momenta = np.linspace(-np.pi, np.pi, 500) for band in range(spec.nbands): plt.plot(momenta, spec(momenta, band), label='n=' + str(band)) Check that the energy dispersion is as expected and that there are available states below the chemical potential :math:`\mu`. Note that without defining the occupation explicitly (via :math:`\texttt{tkwant.manybody.lead_occupation()}`) Tkwant will assume :math:`\mu = 0` for all leads present in the system. Read the `manybody tutorial `_ for more details. - Plot the values of important matrix elements of the system. For instance a 2D plot of the onsite potential can help to see if it has been set properly. Solution ~~~~~~~~ Change the Hamiltonian to the physically correct one. Observables ----------- Problem ~~~~~~~ Error in the mapping between the result obtained from a Kwant operator and the lattice positions (not obvious for a system with more than one dimensions or orbitals). Diagnostic ~~~~~~~~~~ The result obtaind from a Kwant operator is packed in a one-dimensional array, check that you're using the correct ordering for that array. The mapping from a lattice position to the index in the array can be obtained from the class :math:`\texttt{siteId}` .. code-block:: python # system construction lat = kwant.lattice.square(a=1, norbs=2) syst = kwant.Builder() . . syst = syst.finalized() idx = tkwant.system.siteId(syst, lat) index = idx(i, j) # map integer lattice positions (i,j) to the array index where :math:`\texttt{syst}` refers to a finalized Kwant system, :math:`\texttt{lat}` the Kwant lattice and :math:`\texttt{i, j}` the integer lattice positions one is interested. The Kwant ordering of the sites can also be printed with: .. code-block:: python for site in syst.sites: print(site) Solution ~~~~~~~~ Don't assume the ordering of the results, use the above to know how the data as actually stored. Unsufficient accuracy in perturbation interpolation --------------------------------------------------- Problem ~~~~~~~ For abruptly varying time-dependent perturbations :math:`𝑊(𝑡)` of the Hamiltonian, the result becomes inaccurate. Diagnostic ~~~~~~~~~~ Tkwant interpolates the time-dependent perturbation :math:`𝑊(𝑡)` of the Hamiltonian using an adaptive cubic spline interpolation. If :math:`𝑊(𝑡)` becomes strongly varying however, this interpolation becomes inaccurate. We recommend to perform a second simulation with the interpolation turned off and to check if the result changes. Solution ~~~~~~~~ Turn of the interpolation, as explained in: - `Onebody time-dependent perturbation `_ - `Manybody time-dependent perturbation `_ Spurious reflections on the leads --------------------------------- Problem ~~~~~~~ The result shows spurious reflections by a lead. Diagnostic ~~~~~~~~~~ Tkwant provives automatic and adaptive boundary conditions to prevent reflections from the lead. So far, we have not observed any failure of the automatic tuning of the boundary conditions, but it might happen. To check that no reflection is occurring, one can perform a second simulation with a smaller reflection coefficient. Alternatively, one can also analyse the (static) lead reflection. Solution ~~~~~~~~ Make sure that there are no spurious reflections at the lead. The following tutorials explain how to perform the diagnostic and ensure the absence of spurious reflections. - `Onebody boundary conditions `_ - `Manybody boundary conditions `_ and to compare if the result changes. One can also analyse the lead reflection or tweak the boundary conditions as explained in the :ref:`boundary` tutorial. Unsufficient time-stepping accuracy ----------------------------------- Problem ~~~~~~~ The accuracy of the time integration of the Schrodinger equation is insufficient. Diagnostic ~~~~~~~~~~ Tkwant has automatic routines to adaptively integrate the onebody Schrödinger equation in time. So far, we have not observed any failure but it might happen. To check for the accuracy of the time-dependent Schrodinger solver, perform a second simulation with a higer accuracy of the adaptive algorithm and check for convergence. Solution ~~~~~~~~ Use a high enough accuracy. See, - `Onebody time integration `_ - `Manybody time integration `_ for the tutorials on this aspect of the numerical algorithm. Miscellaneous problems ---------------------- A list other known problems that may occur is given in the :ref:`faq` section. References ---------- [1] B. Gaury, J. Weston, M. Santin, M. Houzet, C. Groth, and X. Waintal, `Numerical simulations of time-resolved quantum electronics `_ Phys. Rep. **534**, 1 (2014).