Linear substructure synthesis via Lyapunov stable penalty methods (Q1197287)

As well-known, one of the most prevalent methods to analyse large and complex structures with a good accuracy and avoiding too costly computations, is to use substructuring and modal synthesis methods. These methods comprise two main previous steps, namely: reduced discretized representations of the substructures, and then a boundary interconnection procedure. Here Kurdila and his co-authors tackle only the connection problem, and this in the case of dynamic linear structures. The authors first recall the methods based on the use of Lagrange multipliers to account for linear constraints, and some ways to eliminate the multipliers easier by representing the degrees of freedom in a basis of the kernel defined by the constraint matrix, or by direct elimination. They observe that these methods usually give spurious modes associated to small constraint violations. Hence they propose a new penalty method by introduction of penalized versions of the kinetic and potential energies added to the given ones, which are quadratic functions of the global constraint matrix and which depend on a small parameter \(\varepsilon\) \((\varepsilon\to 0)\). This way gives in fact a regularization of the regarded constraint. The method may be extended to the damping energy. The control method analogy provides a methodology which allows to show by two theorems the minimization of the constraint work, the Lyapunov stability, and also the asymptotic behaviour with error bounds as \(\varepsilon\to 0\). Moreover an abstract angle is introduced which measures the projections of the eigenvectors onto an hyperplane tangential to the constraint surface and gives the way to pick out the spurious modes. Two examples are treated with that so-called penalty method on two systems over 100 degrees of freedom which show in particular the separation of spurious modes from the structural ones. Besides these interesting results, the article calls for some comments: (i) the method has to be proved to keep actually the same efficiency for nonlinear problems; (ii) a posteriori error bounds due to the reduced representation of each substructure remains an open important problem.