A combined first-principles and data-driven approach to model building

• We develop a regression formulation that enforces first-principles relationships.
• Our methodology enforces response bounds, physical limits, and boundary conditions.
• A semi-infinite programming approach infers relationships among regression parameters.
• Constrained regression models are shown to be more robust and physically realizable.

We address a central theme of empirical model building: the incorporation of first-principles information in a data-driven model-building process. By enabling modelers to leverage all available information, regression models can be constructed using measured data along with theory-driven knowledge of response variable bounds, thermodynamic limitations, boundary conditions, and other aspects of system knowledge.We expand the inclusion of regression constraints beyond intra-parameter relationships to relationships between combinations of predictors and response variables. Since the functional form of these constraints is more intuitive, they can be used to reveal hidden relationships between regression parameters that are not directly available to the modeler. First, we describe classes of a priori modeling constraints. Next, we propose a semi-infinite programming approach for the incorporation of these novel constraints. Finally, we detail several application areas and provide extensive computational results.