Convergence tools ==== In case of problems, we have several tools at our disposal: * sub-division of the time step, * change the matrices, * piloting, * linear search, * change the finite element formulation. The sub-division of the time step ---- In general, the smaller the time step, the less non-linear the problem is, and therefore easier to solve. Sub-dividing the time step is therefore an essential tool that makes it possible to overcome the most common difficulties. Subdivision is enabled by default in DEFI_LIST_INST [:external:ref:`U4.34.03 `]. We recommend that you always activate it. Furthermore, the smaller the time step, the smaller the time discretization error. The right choice of matrices ---- In Newton's algorithm, it is possible to use different matrices in prediction and correction and to update them more or less often. These choices are specified in STAT_NON_LINE under the NEWTON [:external:ref:`U4.51.03 `] keyword. The most versatile and robust choice is to use the tangent matrix that is updated at each Newtons' iteration and to choose an elastic prediction. PREDICTION =' ELASTIQUE ' MATRICE =' TANGENTE '(by default) REAC_ITER =1 REAC_INCR =1 (by default) In the event of a convergence problem to pass a delicate moment, it may be interesting to switch from the tangent matrix to the discharge matrix if the time step becomes too small (i.e. when the time step in question has been undercut several times in succession). The value of the time step below which the discharge matrix is taken is given by the keyword PAS_MINI_ELAS. The refresh rate of the discharge matrix is set by REAC_ITER_ELAS. For a precise definition of the discharge matrix, see the keyword factor NEWTON in [:external:ref:`U4.51.03 `]). Piloting ---- Load control is a continuation method for Newton's method. In particular, control makes it possible to calculate the response of a structure that would present instabilities, both geometric (buckling) and material (softening). Its use is limited to simulations for which time does not play a physical role, which excludes *a priori* dynamic, or viscous or thermo-mechanical problems. Linear search ---- Newton's method provides an increment of the unknowns, but this increment is only valid in the vicinity of initialization. The idea of linear research is to use the direction of the increment, but with control over the length of advance in that direction. The progress step is then chosen by minimizing a functional one. In particular, this makes it possible to avoid certain differences in Newton's algorithm. However, activating linear search is "expensive." It is recommended that you only activate it when needed. In some cases, the choice of the partitioner can have an influence on the linear search. We therefore recommend either increasing ITER_GLOB_MAXI or changing the partitioner. Partitioner SCOTCH tends to behave better than partitioner METIS in some cases (ssns115b test case). The right choice of the finite element formulation ---- Not all finite elements are the same. Some items behave better than others in certain situations [:external:ref:`U2.01.10 `]. The choice of finite elements is made in operator AFFE_MODELE.