Failure-averse Active Learning for Physics-constrained Systems