In this work, we focus on the automatic identification of fine-grained problem-solution structure in scientific argumentation. We operationalise the task of finding problem formulations within scientific text in a supervised setting, using a newly-created hand-curated corpus from the domain of computational linguistics. In terms of linguistic features for their detection, we distinguish features from within the statement, and features representing the surrounding context. Results from a classification task on our corpus show that the task of identifying problem statements is tractable using a mixture of features, whereby features modelling the rhetorical context are particularly successful. Overall, our experiment shows promise for future work in identifying scientific problem-solution structure in a more global way.