KAIST

< (From Left) KAIST Dr. Jongchan Kim, Professor Sanghoo Park >

Just as every person has a unique fingerprint, every material has its own unique ‘optical fingerprint.’ Spectroscopy, which has identified materials without contact in fields ranging from semiconductor processes to environmental monitoring, disease diagnosis, and space research, has been called the ‘eyes of science.’ A KAIST research team has implemented spectroscopic analysis, which previously relied on the experience of experts, into AI-based automatic and real-time interpretation technology, greatly expanding its applicability in various industrial fields such as semiconductors, environment, and medicine.

The research team led by Professor Sanghoo Park of our university's Department of Nuclear and Quantum Engineering announced on the 3rd that they have developed ‘AI-based deep spectral interpretation technology’ that allows artificial intelligence to automatically interpret various spectral data in real-time, overcoming limitations such as noise, contamination, defects, and complex overlapping signals.

A spectrum is a graph that spreads out light emitted or absorbed by a material like a rainbow. Existing spectroscopic analysis had to manually analyze signals appearing as numbers in this spectrum by comparing them one by one with well-known reference data. Instead of this method, the research team enabled the artificial intelligence to recognize the entire spectrum as a single ‘image’ and learn its patterns.

< Deep learning-based spectrum technology >

As a result, even in situations where noise was mixed in the data or some parts were lost, the AI accurately identified material information as if it were recognizing an object in a photo. Furthermore, it equipped a function to self-check whether the prediction results are scientifically valid, significantly increasing the reliability of the analysis.

The research team verified this technology by applying it to absorption spectroscopy data widely used in atmospheric and plasma chemistry. As a result, they succeeded in predicting the concentrations of eight chemical substances, including ozone and nitrogen oxides, with very high accuracy even among complexly mixed signals. It was not only more accurate than existing manual analysis but also showed stable performance even in environments with poor data quality.

This research is expected to be a turning point in converting vast amounts of spectroscopic data, which were previously discarded due to the difficulty of analysis, into ‘immediately usable information.’ In particular, it has high potential for use in various high-tech industrial fields, such as improving yield in semiconductor plasma processes, stable control of nuclear fusion plasma, environmental monitoring in smart cities, and non-contact disease diagnosis.

< Research Image >

Professor Sanghoo Park said, “This technology is an achievement that significantly lowers the entry barrier for spectroscopic data analysis, which used to rely on the experience of experts,” and added, “It can be immediately applied to overall industries requiring spectral analysis, such as environmental monitoring, healthcare, and plasma diagnosis.”

In this study, doctoral students Jongchan Kim and Seong-Cheol Huh participated as co-first authors, and Jin Hee Bae and Su-Jin Shin also contributed to the research. The results were published online on January 12th in the prestigious international academic journal in the field of measurement and analytical chemistry, ‘Sensors and Actuators B: Chemical.’

※ Paper title: Deep spectral deconvolution for image-based broadband spectral data analysis DOI: https://doi.org/10.1016/j.snb.2025.139369

Meanwhile, this research was conducted with support from the Ministry of Science and ICT’s Global TOP Strategic Research Group Support Program, the KAIST Leap Research Project, and the Korea Institute of Materials Science (KIMS).