TY - JOUR
T1 - Conceptual Limits of Analogy-Based Reasoning in Generative Materials Models
AU - Carlos Ramirez
AU - Elena Torres
AU - Pablo Ortega
AU - Sofia Mendes
JF - Journal of Artificial Intelligence for Materials Science
JO - J. Artif. Intell. Mater. Sci.
SN - 3149-8957
Y1 - 2023
VL - 2
IS - 2
SP - 115
N2 - Generative materials models, including variational autoencoders, generative adversarial networks, and diffusion models, have become central to modern artificial intelligence for materials science. Yet, their pervasive reliance on analogy-based reasoning remains largely unexamined and conceptually undertheorized. These models routinely treat latent-space interpolation, transfer learning, and structural substitution as forms of analogical mapping—assuming that what holds between known materials will hold for novel ones—without acknowledging the fundamental epistemological limits of such reasoning. This critical critique identifies four interlocking problems that undermine the reliability of analogy-driven generation: analogy functioning as a substitute for genuine physical understanding, the propagation of false analogies, boundary blindness to domains where analogies break, and the reification of statistical correlations into ontological claims. The consequences of these unacknowledged limits extend beyond technical inaccuracy to wasted experimental resources, overconfident predictions, and a subtle distortion of scientific understanding in materials discovery. Rather than abandoning analogy entirely, this paper argues for hybrid frameworks that explicitly bind analogical transfer with physical invariants, causal verification, and uncertainty quantification. By confronting these conceptual limits head-on, the field can move toward more robust, epistemologically grounded generative models that augment rather than replace mechanistic insight.
UR - https://iamrp.net/u551520197
ER -