The quantity of data available to scientists in all disciplines is increasing at an exponential rate, yet the insight necessary to distill data into scientific knowledge must still be supplied by human experts. This widening gap between data and insight can be bridged with data-driven modeling, in which computational methods shift much of the work in creating models from humans to computers. Traditional approaches to data-driven modeling require that the form of the model be fixed in advance, which requires substantial human effort and limits the complexity of problems that can be addressed. In contrast, a newer approach to automated modeling based on evolutionary computation (EC) removes such restrictions on the form of models. This free-form modeling has the potential both to reduce human effort for routine modeling and to make complex problems more tractable. Although major advances in EC-based modeling have been made in recent years, many challenges remain. These challenges include three features often seen in biological systems: complex nonlinear behavior, multiple time scales, and hidden variables. This work addresses these challenges by developing new approaches to ECbased modeling, with applications to neuroscience, systems biology, ecology, and other fields. The contributions of this work consist of three primary lines of research. In the first line of research, EC-based methods for the automated design of analog electrical circuits are adapted for the modeling of electrical systems studied in neurophysiology that display complex, nonlinear behavior, such as ion channels. In the second line of research, EC-based methods for symbolic modeling are extended to facilitate the modeling of dynamical systems with multiple time scales, such as those found throughout ecology and other fields. Finally, in the third line of research, established EC-based algorithms are extended with the capability to model dynamical systems as systems of differential equations with hidden variables, which can contribute in an essential way to the observed dynamics of a physical system yet historically have presented a particularly difficult challenge to automated modeling.