Add configuration file parameter [LineSearch][fail_if_not_converged]

This parameter determines, whether the line search is cancelled once the
state system cannot be solved at a (trial) iterate. If this is True, then
the line search is cancelled. Otherwise, the step size is "halved" and a
new iterate is generated.

CHANGELOG.rst

@@ -10,6 +10,12 @@ of the maintenance releases, please take a look at
 2.1.0 (in development)
 ----------------------
 
+* New configuration file parameters:
+
+  * Section LineSearch
+
+    * :ini:`fail_if_not_converged` determines, whether the line search is cancelled once the state system cannot be solved or if a new iterate is tried instead.
+
 
 
 2.0.0 (May 16, 2023)

cashocs/_optimization/line_search/armijo_line_search.py

@@ -24,6 +24,7 @@
 import fenics
 from typing_extensions import TYPE_CHECKING
 
+from cashocs import _exceptions
 from cashocs import _loggers
 from cashocs._optimization.line_search import line_search
 
@@ -124,9 +125,9 @@ def search(
             )
 
             current_function_value = solver.objective_value
-
-            self.state_problem.has_solution = False
-            objective_step = self.cost_functional.evaluate()
+            objective_step = self._compute_objective_at_new_iterate(
+                current_function_value
+            )
 
             decrease_measure = self._compute_decrease_measure(search_direction)
 
@@ -176,3 +177,24 @@ def _compute_decrease_measure(
             return self.decrease_measure_w_o_step * self.stepsize
         else:
             return float("inf")
+
+    def _compute_objective_at_new_iterate(self, current_function_value: float) -> float:
+        """Computes the objective value for the new (trial) iterate.
+
+        Args:
+            current_function_value: The current function value.
+
+        Returns:
+            The value of the cost functional at the new iterate.
+
+        """
+        self.state_problem.has_solution = False
+        try:
+            objective_step = self.cost_functional.evaluate()
+        except (_exceptions.PETScKSPError, _exceptions.NotConvergedError) as error:
+            if self.config.getboolean("LineSearch", "fail_if_not_converged"):
+                raise error
+            else:
+                objective_step = 2.0 * current_function_value
+
+        return objective_step

cashocs/io/config.py

@@ -224,6 +224,9 @@ def __init__(self, config_file: Optional[str] = None) -> None:
                     "attributes": ["less_than_one", "positive"],
                     "larger_than": ("LineSearch", "factor_low"),
                 },
+                "fail_if_not_converged": {
+                    "type": "bool",
+                },
             },
             "AlgoLBFGS": {
                 "bfgs_memory_size": {
@@ -560,6 +563,7 @@ def __init__(self, config_file: Optional[str] = None) -> None:
 polynomial_model = cubic
 factor_high = 0.5
 factor_low = 0.1
+fail_if_not_converged = False
 
 [ShapeGradient]
 lambda_lame = 0.0

docs/source/user/demos/optimal_control/doc_config.rst

@@ -337,6 +337,14 @@ For the polynomial models, we also have a safeguarding procedure, which ensures
 
 and the values specified here are also the default values for these parameters.
 
+Finally, we have the parameter
+
+.. code-block:: ini
+
+    fail_if_not_converged = False
+
+which determines, whether the line search is terminated if the state system cannot be solved at the current iterate. If this is :ini:`fail_if_not_converged = True`, then an exception is raised. Otherwise, the iterate is counted as having too high of a function value and the stepsize is "halved" and a new iterate is formed.
+
 .. _config_ocp_algolbfgs:
 
 Section AlgoLBFGS
@@ -688,6 +696,9 @@ in the following.
       - Safeguard for stepsize, upper bound
     * - :ini:`factor_low = 0.1`
       - Safeguard for stepsize, lower bound
+    * - :ini:`fail_if_not_converged = False`
+      - if this is :ini:`True`, then the line search fails if the state system can not be solved at the new iterate
+
 
 [AlgoLBFGS]
 ***********

docs/source/user/demos/shape_optimization/doc_config.rst

@@ -356,6 +356,14 @@ For the polynomial models, we also have a safeguarding procedure, which ensures
 
 and the values specified here are also the default values for these parameters.
 
+Finally, we have the parameter
+
+.. code-block:: ini
+
+    fail_if_not_converged = False
+
+which determines, whether the line search is terminated if the state system cannot be solved at the current iterate. If this is :ini:`fail_if_not_converged = True`, then an exception is raised. Otherwise, the iterate is counted as having too high of a function value and the stepsize is "halved" and a new iterate is formed.
+
 .. _config_shape_algolbfgs:
 
 Section AlgoLBFGS
@@ -1128,7 +1136,6 @@ in the following.
       - :ini:`picard_verbose = True` enables verbose output of Picard iteration
 
 
-
 [OptimizationRoutine]
 *********************
 
@@ -1178,7 +1185,8 @@ in the following.
       - Safeguard for stepsize, upper bound
     * - :ini:`factor_low = 0.1`
       - Safeguard for stepsize, lower bound
-
+    * - :ini:`fail_if_not_converged = False`
+      - if this is :ini:`True`, then the line search fails if the state system can not be solved at the new iterate
 
 [AlgoLBFGS]
 ***********