When deploying machine learning estimators in science and engineering (SAE) domains, it is critical to avoid failed estimations that can have disastrous consequences, e.g., in aero engine design.

However, most existing meta-learning approaches only can learn from tasks without any constraint.