This study mined student interactions with visual representations as a means to automate assessment of learning in a complex, inquiry-based learning environment. Log trace data of 143 middle school students' interactions with an interactive map in Research Quest (an inquiry-based, online learning environment) were analyzed. Students used the interactive map to make scientific observations for an evidence-based hypothesis. The examination of classification error using an artificial neural network, compared against the majority class for prediction, suggests that student performance on several metrics of critical thinking can be classified based on different patterns in interactions with visual representations. Two alternative methods are compared in this study for training and evaluating data-mined models of student performance. In accordance with the general consensus in the literature, the error estimates for models' predictions were less variable using a student-level cross-validation. Implications of these findings for open-ended inquiry-based learning environments are discussed.