In recent years, Satisfiability Modulo Theories (SMT) solvers have emerged as powerful tools in many formal methods applications, including verification, automated theorem proving, planning and software synthesis. The expressive power of SMT allows problems from many disciplines to be handled in a single unified approach. While SMT solvers are highly effective at handling certain classes of problems due to highly tuned implementations of efficient ground decision procedures, their ability is often limited when reasoning about universally quantified first-order formulas. Since generally this class of problems is undecidable, most SMT solvers use heuristic techniques for answering unsatisfiable when quantified formulas are present. On the other hand, when the problem is satisfiable, solvers using these techniques will either run indefinitely, or give up after some predetermined amount of effort. In a majority of formal methods applications, it is critical that the SMT solver be able to determine when such a formula is satisfiable, especially when it can return some representation of a model for the formula. This dissertation introduces new techniques for finding models for SMT formulas containing quantified first-order formulas. We will focus our attention on finding finite models, that is, models whose domain elements can be represented as a finite set. We give a procedure that is both finite model complete and refutationally complete for a fragment of first-order logic that occurs commonly in practice.

This thesis presents a framework for understanding first-order theories by investigating their models. A common application is to help users, who are not necessarily experts in formal methods, analyze software artifacts, such as access-control policies, system configurations, protocol specifications, and software designs. The framework suggests a strategy for exploring the space of finite models of a theory via augmentation. Also, it introduces a notion of provenance information for understanding the elements and facts in models with respect to the statements of the theory. The primary mathematical tool is an information-preserving preorder, induced by the homomorphism on models, defining paths along which models are explored. The central algorithmic ideas consists of a controlled construction of the Herbrand base of the input theory followed by utilizing SMT-solving for generating models that are minimal under the homomorphism preorder. Our framework for model-exploration is realized in Razor, a model-finding assistant that provides the user with a read-eval-print loop for investigating models.