Building Flexible Machine Learning Models for Scientific Computing at Scale