Physics-IntegratedVariationalAutoencodersfor RobustandInterpretableGenerativeModeling