vIR-OLO: an open-source YOLO-based interface for functional- group detection and localization in infrared spectra
The paper introduces vIR-OLO, an open-source, YOLO-based framework that reformulates functional group identification in infrared spectra as an object detection task to simultaneously predict group presence and localize characteristic spectral regions, thereby enhancing interpretability and supporting human-in-the-loop workflows.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Chemists have long relied on a technique called infrared spectroscopy to understand what molecules are made of. When a beam of invisible infrared light passes through a sample, the molecules absorb specific colors of that light, creating a unique pattern of peaks and valleys on a graph. This pattern acts like a fingerprint, revealing the presence of specific clusters of atoms known as functional groups. Identifying these groups is essential for everything from discovering new medicines to analyzing ancient artifacts. However, reading these fingerprints has traditionally been a slow, manual job. An expert must look at the graph, recognize the shape and position of each peak, and decide which chemical group it represents. This process requires years of training and constant reference to databases, and even experts can make mistakes when the patterns are unclear or the molecules are unusual.
Recently, scientists have tried to use computer programs to automate this reading. Most of these programs act like a simple classifier: they look at the entire graph and output a list of the groups they think are present. While this is faster than doing it by hand, it has a significant flaw. The computer gives an answer but offers no visual proof of where it found the evidence. It is like a teacher grading a test and circling the final score without showing which questions were answered correctly. Without seeing the specific part of the graph the computer used to make its decision, a human scientist cannot easily verify the result or correct it if the computer is wrong. This lack of transparency makes it difficult to build trust in the machine's judgment or to use the computer's work to improve future models.
A team of researchers at McMaster University has addressed this gap by creating a new tool called vIR-OLO. Instead of treating the problem as a simple list-making exercise, they reformulated it as a task of finding and marking specific objects within an image. They adapted a type of artificial intelligence originally designed for spotting cars or people in photographs and taught it to spot chemical peaks in infrared graphs. Rather than just saying "alcohol is present," the software draws a box around the exact section of the graph where the alcohol peak appears. This allows a human user to see exactly what the computer is looking at, verify the location, and even adjust the box if the computer missed the mark. The system is designed to be a partner in the lab, not a replacement for human expertise, enabling a workflow where a scientist can quickly review, correct, and refine the computer's suggestions.
To build this system, the researchers gathered thousands of infrared graphs from a public database and manually drew boxes around the peaks corresponding to twelve different types of functional groups, such as alcohols, acids, and aromatics. They then trained three different versions of their software, each with a slightly different set of rules for how to group these chemicals. One version kept all twelve groups separate, another merged similar groups like ketones and aldehydes into a single "carbonyl" category to reduce confusion, and a third version split some groups into left and right sections to handle complex patterns. After training, they tested the software on hundreds of unseen graphs. The results showed that the system was highly effective at identifying which groups were present, correctly detecting them in more than 84 percent of cases across all versions.
More importantly, the system proved capable of pinpointing the location of these groups with high precision. When the computer drew a box around a peak, that box overlapped with the correct area of the graph more than 80 percent of the time. This level of accuracy means the software is not just guessing; it is reliably locating the specific features that define a chemical group. The researchers found that merging similar groups into a single category sometimes made the system slightly more consistent at finding the right answer, while keeping the groups separate provided more detailed chemical information but required more data to avoid confusion. The software also learned to recognize how certain groups often appear together, such as phenols and aromatics, mirroring the chemical relationships found in nature.
The researchers released their software as an open-source tool, meaning anyone can download and use it without paying a fee. The program includes a user-friendly interface that lets scientists load their own data, run the detection, and manually edit the results. If a user corrects a mistake, the system can learn from that correction and improve its future predictions. This creates a cycle of continuous improvement where the tool becomes more accurate as more people use it and refine their own datasets. By making the detection process visible and editable, the researchers have provided a practical way for chemists to work alongside artificial intelligence, turning a black-box prediction into a transparent, collaborative process that can accelerate the discovery and analysis of new materials.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.