Spectroscopy as AI Language: New Framework Treats Spectra as 'Physical Tokens'

Artificial intelligence has mastered language by breaking continuous streams of words into discrete units—tokens—and learning how they relate to one another. A researcher at the University of Science and Technology of China now proposes that spectroscopy could work the same way for the physical sciences. When light hits matter, it produces a spectrum: peaks and valleys, shifts and intensities, patterns that reveal what a substance is and how it behaves. What if those spectral features could be treated as "physical tokens," discrete units of information that machines could learn to read and use to design new materials?

Yi Luo, working at the Hefei National Research Center for Physical Science at the Microscale, published this framework in Precision Chemistry on August 24, 2026. The core insight is simple but ambitious: spectroscopy has always been a tool for identifying substances after the fact, a fingerprint left behind by an experiment. But what if it became something more active—a language that connects observation, physical law, and design in a single system?

The framework defines a physical token at three levels. At the quantum level, it corresponds to the energy transitions that quantum mechanics allows. At the experimental level, it appears as the actual peaks, line widths, intensities, and shapes that a spectrometer records. At the machine-learning level, it can be discretized into spectral intervals, patches, or abstract representations that algorithms can process. The relationships among these tokens—how a peak shifts when the environment changes, how intensity varies with composition—encode what Luo calls the "syntax" of material behavior. In other words, the grammar of how matter speaks through its spectrum.

This reframing opens a path toward what Luo calls spectrum-guided inverse design. Instead of starting with a desired material property and working backward through chemistry to guess what to make, a researcher could translate that goal into a theoretical spectral blueprint. Physics-informed generative models would then propose candidate structures that should produce that spectrum. Robots would synthesize those candidates, measure their spectra, and feed the results back into the model. The loop closes: desired function becomes spectral target becomes proposed structure becomes experiment becomes refined model. No human intuition required—only physics, data, and iteration.

For this to work at scale, Luo argues, science needs new infrastructure. High-quality databases that combine spectral data with structural and property information across multiple techniques—infrared, Raman, nuclear magnetic resonance, X-ray diffraction. Models that understand how all these techniques relate to one another and to the underlying physics. Complete pipelines that connect spectrum to structure to function without gaps. Tighter integration between computational prediction, robotic synthesis, and real-time spectroscopic measurement.

The vision is self-driving laboratories. A researcher defines a goal—a material with specific strength, conductivity, or optical properties. The AI system plans a series of experiments, proposes candidate structures, and identifies which spectra would confirm or refute each hypothesis. Robots execute the synthesis and measurement. New spectral data arrives, and the model learns, refining its next proposals. Humans step back from the bench work and focus on asking better questions.

What makes this framework distinct is that it keeps spectroscopy anchored in physics. The tokens are not arbitrary; they emerge from quantum mechanics and the interaction of light with matter. The relationships among them follow physical law. An AI system trained on this framework would learn patterns that are not just statistically correlated but physically meaningful—constrained by the same rules that govern the universe. That constraint is what allows the system to generalize beyond the data it has seen, to propose materials that have never been made before but should work because the physics says so.

The proposal is still theoretical. No self-driving laboratory yet operates at this level of autonomy. But the framework offers a roadmap: treat spectra as a machine-readable language for matter, build the databases and models to support it, and watch what becomes possible when discovery moves from human intuition to closed-loop, physics-informed design.