TY - JOUR
T1 - A Conceptual Framework for Detecting Scientific Illusions in Generative Materials Outputs
AU - Alejandro Torres
AU - Miguel Fernandez
JF - Journal of Artificial Intelligence for Materials Science
JO - J. Artif. Intell. Mater. Sci.
SN - 3149-8957
Y1 - 2024
VL - 3
IS - 2
SP - 130
N2 - Scientific illusions represent an undetected yet pervasive failure mode in generative materials AI, where model outputs appear scientifically credible and satisfy superficial visual or computational checks but are fundamentally invalid, thereby consuming experimental resources, misleading research trajectories, and eroding community trust in AI-driven discovery pipelines. A scientific illusion is formally defined as any generative model output that meets surface-level plausibility constraints—such as reasonable bond lengths or predicted low formation energy—while violating deeper physical, chemical, or thermodynamic principles that render the material unrealizable or non-existent in nature. The four primary types of scientific illusions specific to materials generation—structural, property, stability, and novelty—arise through well-characterized mechanisms including spurious correlation learning, mode averaging across disparate training distributions, boundary artifacts in latent representations, and systematic training bias toward plausible but unphysical regions, as synthesized from recent surveys and critical reviews in the field. This paper proposes a five-component conceptual framework for pre-validation detection that operates entirely at the conceptual and computational screening level, enabling researchers to flag illusions before any laboratory commitment. Adoption of this framework promises to transform generative materials practice by shifting evaluation paradigms from post-hoc experimental triage to proactive illusion-aware design, ultimately accelerating credible discovery while safeguarding scientific integrity.
UR - https://iamrp.net/f818549890
ER -