Published resources such as technical literature and patent documents are extremely useful in engineering design and form an important input to methods such as TRIZ. Often, design engineers will investigate these resources when working on new design problems. Aside from getting technical information and even direct design solutions, they may find the design principles used in each patent document a useful design stimulus. Unfortunately, patents are not classified based on such “design useful” characterizations. Using unsupervised clustering and Latent Dirichlet Allocation, this paper investigates four hypotheses using engineering patents in informing TRIZ based design. It first investigates the optimal number of TRIZ topics present in a corpus. Using this information, it attempts to map the TRIZ methods to the individual patents using unsupervised machine learning. Both rejected and accepted patents are then tested to determine if an autoencoder can successfully differentiate between the two, just from the text of the document. The autoencoder reconstruction errors of “Vehicle Brake Control” patents are also examined for possible correlation between reconstruction error and patent citation count. Finally, by combining the TRIZ clustering and the trained autoencoder, we show that high reconstruction error patents may be harder to assign to TRIZ methods than low reconstruction error patents.