What is an engineering model? philosophical arguments and counterarguments on why machine learning models belong in engineering

Abstract

This paper examines the ontological and epistemological status of engineering models to make explicit the often implicit criteria that govern model acceptance and evaluate whether machine learning (ML) models satisfy these criteria. We argue that the distinction engineers draw between "physics-based" and "data-driven" models lacks ontological robustness, and we propose a definition that identifies decision support and evidence of performance as constitutive requirements, while remaining neutral with respect to derivational history and mathematical form. From this foundation, three claims arise. First, ontologically, engineering models are structured compressions of regularities that differ in their degree of theoretical content but not in kind. Second, epistemologically, the warrant for trusting a model in engineering practice derives from calibration, validation, and domain-appropriate uncertainty quantification, and these criteria apply equally to ML models and traditional models. Third, pragmatically, model equivalence should be defined operationally in terms of decision support under reliability constraints rather than physical interpretability. The resistance to ML models in traditional engineering disciplines, thus, may reflect not only philosophical concerns but also an attachment to a conception of proper engineering that ML appears to challenge. This paper concludes that ML models represent not a departure from the engineering modeling tradition but its logical completion, insofar as they make explicit the empiricism that was always implicit in code equations, constitutive laws, and phenomenological models.