A pattern database is a lookup table that stores an exact evaluation function for a relaxed search problem, which provides an admissible heuristic for the original search problem. In general, the larger the pattern database, the more accurate the heuristic function. We consider how to build large pattern databases that are stored in external memory, such as disk, and how to use an external-memory pattern database efficiently in heuristic search. To limit the number of slow disk I/O operations needed to construct and query an external-memory pattern database, we adapt an approach to external-memory graph search called structured duplicate detection that localizes memory references by leveraging an abstraction of the state space. We present results that show this approach increases the scalability of heuristic search by allowing larger and more accurate pattern database heuristics.