Large-scale simulations, e.g. fluid-structure interactions and aeroacoustics noise generation, require important computing power, visualization systems and high-end storage capacity. Because 3D multi-physics simulations also run long processes on large datasets, an important issue is the robustness of the computing systems involved, i.e., the ability to resume the inadvertantly aborted computations. A new approach is presented here to handle application failures. It is based on extensions of bracketing checkpoints usually implemented in database and transactional systems. An assymetric scheme is devised to reduce the number of checkpoints required to safely restart aborted applications when unexpected failures occur. The tasks are controled by a workflow graph than can be deployed on various distributed platforms and high-performance infrastructures. An automated bracketing process inserts in the workflow graph checkpoints that are placed at critical execution points in the graph. The checkpoints are inserted using a heuristic process based on a evolving set of rules. Preliminary tests show that the number of checkpoints, hence the overhead incurred by the checkpointing mechanism, can be significantly reduced to enhance the application performance while supporting its resilience.