Deep Scientific Reasoning under Physical Constraints: Structure-Aware Spectrum Prediction

Structured scientific spectra encode rich physical information while obeying hard constraints, such as conservation laws and characteristic spectral geometry. Accurate prediction of these spectra is central to materials discovery, yet existing methods often treat them as unconstrained sequences and therefore fail to enforce the underlying physical structure. Taking electronic density of states (eDOS) as the prototypical example, we introduce \textbf{DeepSciReasoner}, a general paradigm for predicting scientific spectra under physical constraints. The framework combines structure-aware spectrum decoding with constraint-preserving physical reasoning, allowing predictions to capture rich spectral structure while respecting the underlying physics. We evaluate DeepSciReasoner on eDOS, phonon density of states (phDOS), X-ray absorption near-edge structure (XANES), and Raman spectra, where it substantially improves prediction accuracy while maintaining physical consistency. These results establish DeepSciReasoner as a reusable blueprint for structured scientific spectrum prediction under hard physical constraints.

Many important materials properties are revealed through spectra: curves that show how a material responds across energy, frequency, or wavelength. These spectra help scientists understand whether a material may be useful for electronics, energy devices, or other technologies, but computing them accurately with first-principles simulations is expensive. Machine learning could make this faster, but many existing models treat spectra as ordinary sequences of numbers and can produce predictions that look plausible while violating basic physical rules, such as preserving the total amount of spectral signal or keeping sharp peaks and gaps in the right places. We introduce DeepSciReasoner, a machine learning framework that predicts scientific spectra while respecting these physical constraints. The model first learns how atoms in a crystal contribute to different parts of a spectrum, then refines the predicted spectrum by moving spectral signal around in a way that preserves the total amount. We test the method on several types of materials spectra, including electronic density of states, phonon density of states, X-ray absorption spectra, and Raman spectra. Across these tasks, DeepSciReasoner improves prediction accuracy, better identifies important spectral gaps, and uses data more efficiently than prior methods. This work shows that combining neural networks with physical reasoning can make AI predictions for scientific discovery more reliable and useful.