Expert reasoning in the natural sciences appears to make extensive use of a relatively small number of general principles and reasoning strategies, each associated with a larger number of more specific inference patterns. Using a dual declarative hierarchy to represent strategic and factual knowledge, a framework for a robust scientific reasoning engine is analyzed. It is argued that such an engine could provide: (1) the ability to reason from basic principles in the absence of directly applicable specific information; (2) principled knowledge acquisition by using existing general patterns to structure new information; and (3) congenial explanation and instruction in terms of general and familiar patterns of inference. Whereas the reasoning system discussed is yet to be built, it is believed that the analysis and design presented suggest a powerful new framework for building reasoning engines. In addition to its intrinsic artificial intelligence interest, this work is relevant to questions central to philosophy of science, to psychology, and to education. For example, can students be helped to learn science by explicit teaching of general inference patterns? A general reasoning system could support explanatory intelligent computer-assisted-instruction systematically acquainting students with powerful re-usable patterns of inference. (Author/JN)