Simulation and Justification in the Age of Deep Learning: AlphaFold as a Paradigmatic Case of Opaque Models

Abstract

This paper examines the epistemic status of contemporary computational models based on deep learning through the case of AlphaFold, with the aim of assessing to what extent such systems can be understood within the traditional framework of computational simulations or whether they require distinct epistemic categories. Drawing on Paul Humphreys’ analytical framework—particularly the notions of computational templates and computational models—the paper explores the possibility of a formal and epistemic continuity between classical simulations and deep learning systems, while critically addressing issues of opacity, pragmatic validation, and the constitutive role of visual representation. The analysis shows that, although classical representationalist accounts are insufficient to capture the internal functioning of models such as AlphaFold, these systems cannot be straightforwardly reduced to mere algorithmic devices. In this sense, it is argued that AlphaFold confronts philosophy of science with the need to rethink both the criteria of epistemic justification and the scope of the notion of simulation in the era of deep learning.