Architectural Invariance in Spectral Autoencoders: The Robustness of 1D Convolutional Kernels Against kappa-Deformed Loss Regimes in Native Raman Taxonomy

In the unsupervised machine learning analysis of raw, un-preprocessed Raman spectra, the massive and non-linear background variance introduced by fluorescence baselines typically creates a "luminescence trap". Standard fully connected Dense Autoencoders (DAEs) optimized with classical Shannon-based Mean Squared Error (MSE) loss routinely succumb to this artifact, grouping samples by instrumental background noise rather than authentic chemistry. While our previous work demonstrated that deforming the metric space using relativistic Kaniadakis kappa-statistics acts as an automatic mathematical brake that unlocks objective taxonomy within DAEs, the universal applicability of this statistical braking mechanism across different deep architectures remains an open question. In this study, we implement a rigorous comparative analysis matching classical Shannon-MSE against Kaniadakis kappa-loss function topologies within a 1D Convolutional Autoencoder (Conv-1D AE) framework, processing 96 native, un-preprocessed carbonaceous Raman spectra. Quantitative cross-classification contingency matrices and generative centroid profiles (pseudospectra) reveal a striking morphological and taxonomic invariance across both the sub-unitary (kappa = 0.5) and relativistic (kappa = 2.0) regimes. The natural data partitions, peak geometries, and reconstructed baseline curves remain stable, yielding functionally equivalent classifications regardless of the severe gradient penalties introduced by the deformed loss metric. Rather than representing a trivial lack of sensitivity, this crucial "negative" result reveals a profound architectural divergence in how deep networks process sequential data. While point-by-point Dense layers are highly vulnerable to global metric adjustments, the local receptive fields and sliding spatial kernels of Conv-1D layers prioritize topological shape correlations, local derivatives, and peak-to-valley geometries over absolute numerical distances. This structural feature-extraction capability acts as an intrinsic spatial shield, anchoring the unsupervised taxonomy directly to the authentic molecular signatures (D and G carbon bands). We conclude that Conv-1D autoencoders possess an inherent architectural robustness that ensures stability against cost function deformations, establishing spatial convolution as a vital design choice for baseline-independent spectral library generation.