Research
Papers and publications in computational chemistry, spectroscopy, and ML.
Information-Theoretic Limits of Spectroscopic Molecular Identification
Tubhyam Karthikeyan
We present a group-theoretic and information-theoretic framework for bounding what vibrational spectra can reveal about molecular structure. Using representation theory, we define the Information Completeness Ratio R(G,N) that quantifies the fraction of vibrational degrees of freedom observable via IR and Raman spectroscopy for molecules with point group G and N atoms. We prove that for centrosymmetric molecules, IR and Raman provide complementary (non-overlapping) vibrational information, and conjecture — with strong numerical evidence from Jacobian rank analysis on 999 QM9 molecules — that the combined forward map is generically injective up to symmetry equivalence.
Hybrid State-Space Attention for Multi-Task Vibrational Spectroscopy
Tubhyam Karthikeyan
We introduce Spekron, a hybrid state-space and attention architecture for multi-task vibrational spectroscopy. The model combines wavelet-domain embeddings with selective state-space sequence modeling (Mamba), Mixture of Experts routing, and a Variational Information Bottleneck for disentangling chemical and instrumental information. Pretrained on 60K+ spectra from RRUFF and OpenSpecy, Spekron achieves competitive calibration transfer on the corn and tablet benchmark datasets with as few as 10 labeled transfer samples.