Deep-learning-based digital twins (DDT) are a promising tool for data-driven system health management because they can be trained directly on operational data. A major challenge for efficient training however is that industrial datasets remain unlabeled. This is remedied by simulators that can generate specific run-to-failure trajectories of assets as training data, but extensive simulations are limited by their computational cost. Therefore, it remains difficult to train DDTs that generalize over a wide range of operational conditions. In this research, we propose a novel meta-learning framework that is able to efficiently generalize an arbitrary DDT using the output of a differentiable simulator. While previous generalization approaches are based on randomly-sampled data augmentations, we exploit the differentiability of the full pipeline to actively optimize the training data sampling by means of condition parameter's gradients. We use these gradients as an accurate tool to control the sampling distribution of the simulator, improving the representativeness, robustness, and training speed of the DDT. Moreover, this metalearning approach leads to a higher quality of generalization and makes the DDT more robust to perturbations in the conditional parameters.