TRIZ-Aware Loss v2.0 with Casimir-Regularized Fields: Geometric Invariance for Explainable Neural Machine Translation in Metrological and Scientific Domains

This record contains the full text and supplementary materials for the article introducing TRIZ-Aware Loss v2.0, a physics-inspired and geometrically invariant loss function designed for neural machine translation of metrological, physical, and highly technical scientific texts. This release represents the next part of the UMT-NMT framework, extending its principles toward semantic-field regularization and geometric stability. Version 2.0 expands the original TRIZ-Loss formulation by incorporating Casimir-like short-range semantic forces, cross-domain bridge alignment between classical and quantum constants, manifold-smoothness priors, and dynamic stabilization of the embedding geometry. Together, these mechanisms form a unified field-regularized objective that preserves physical dimensions, constants, formulas, and structural relations during translation, enabling more explainable and faithful outputs in domains where even small distortions are unacceptable. Supplementary Video 1 — Casimir Field Evolution (00:05).This animation presents a time-lapse visualization of cosine-similarity dynamics among the four principal semantic anchors (<anchor_SI>, <anchor_Planck>, <anchor_Dirac>, <anchor_Josephson>) during fine-tuning, demonstrating the progressive convergence and stabilization induced by the Casimir regularization term. © 2025 Yurii Kuzmenko, Olena Trofimova.For academic use only. Redistribution and commercial use prohibited.