In this research, a high word-recognition accuracy was achieved using an e-Science friendly deep learning method on a highly multilingual data set. Deep learning requires large training sets. Therefore, we use an auxiliary data set in addition to the target data set which is derived from the collection Natuurkundige Commissie, years 1820-1850. The auxiliary historical data set is from another writer (van Oort). The method concerns a compact ensemble of Convolutional Bidirectional Long Short-Term Memory neural networks. A dual-state word-beam search combined with an adequate label-coding scheme is used for decoding the connectionist temporal classification layer. Our approach increased the recognition accuracy of the words that a recognizer has never seen, i.e., out-of-vocabulary (OOV) words with 3.5 percentage points. The use of extraneous training data increased the performance on in-vocabulary words by 1 pp. The network architectures in an ensemble are generated randomly and autonomously such that our system can be deployed in an e-Science server. The OOV capability allows scholars to search for words that did not exist in the original training set.