One Interaction, Three Fields: The Disciplinary Phase Transition Driven by AI Foundation Models

This paper begins with a single observation. Biology, chemistry, and physics have independently—without coordinating across fields—arrived at the same conclusion: the causal structure of nature is learnable. AlphaFold demonstrated this for protein folding, MACE-MP-0 for interatomic interactions, and the General Physics Transformer and Fourier Neural Operator for continuum physics. These models span the range from pioneering architecture (FNO) to full foundation model (AlphaFold, MACE, GPhyT), yet all instantiate the same computational structure. This paper argues that this convergence is not coincidental. To substantiate this claim, we introduce the Learned Interaction Pipeline—a four-stage computational structure (tokenize, interact, aggregate, decode)—and demonstrate that all four models instantiate it, differing only in domain-specific implementation choices. We offer one explanation for this convergence: at the atomic scale, the three fields study the same object, and AI foundation models, by learning directly at that scale, have rendered the methodological rationale for disciplinary boundaries structurally redundant. But the paper's central claim does not depend on this explanation. Whether the convergence reflects the unity of nature or the versatility of AI architectures, the observable consequence is the same: domain-specific methods are being replaced by domain-agnostic ones across all three fields simultaneously. Because AI methodology drove this replacement, the pace of progress across these fields is now increasingly coupled to the pace of AI advancement.