Towards practical molecular structure elucidation: A lightweight multi-spectral CNN-Transformer framework

Chemists have long relied on a set of tools to figure out what a molecule looks like just by listening to how it vibrates or how its atoms respond to magnetic fields. These tools, known as spectroscopy, produce complex patterns of lines and peaks that act as a unique fingerprint for every chemical substance. For decades, scientists have tried to teach computers to read these fingerprints and instantly draw the corresponding molecular structure, a process called structure elucidation. While early attempts required massive amounts of human help and extra data to work, the goal has always been to create a system that can look at a raw spectrum and simply tell you the shape of the molecule, much like a human expert would, but with perfect speed and consistency.

A team of researchers has now built a new kind of computer program that comes closer to this goal than ever before, using a method that requires almost no outside help. Instead of feeding the computer a list of possible answers or adding extra chemical facts to guide it, the researchers let the machine look directly at the raw data from three different types of spectroscopy: infrared light, Raman light, and nuclear magnetic resonance. They designed a lightweight system that combines two powerful types of artificial intelligence. The first part acts like a filter, scanning the raw lines of the spectrum to find important patterns. The second part acts like a translator, taking those patterns and writing them out as a string of text that describes the molecule's shape. This text format, known as a SMILES string, is a standard way chemists write down structures so computers can read them.

The results of this new system are surprisingly strong. When tested on simulated data, the model could correctly identify the structure of a molecule from just its infrared spectrum about 69 percent of the time. When it used only the Raman spectrum, the accuracy jumped to nearly 78 percent. The nuclear magnetic resonance data alone was less helpful, getting the answer right about 38 percent of the time. However, when the researchers combined all three types of data, the system became incredibly precise, getting the correct structure right 84 percent of the time on the very first guess. If the researchers allowed the computer to make five guesses and pick the best one, the success rate soared to over 96 percent. These numbers represent a significant leap forward because the system achieved them without using any pre-processed data, chemical formulas, or external databases to derive the answer.

To make the system even better, the researchers added a special training step that taught the computer what not to do. After the model made its initial guesses, the researchers looked at the mistakes it made and showed it those errors again, instructing it to avoid those specific wrong paths in the future. This process, which they call unlikelihood fine-tuning, pushed the accuracy even higher. For example, when using all three spectra together, the model's top guess became correct 84.23 percent of the time, and if allowed five guesses, it was right 96.27 percent of the time. This improvement happened without adding any new chemical knowledge, proving that the model simply learned to be more careful by studying its own past failures.

The researchers also took the time to understand how the computer was thinking, peering inside the "black box" to see which parts of the data mattered most. They discovered that the Raman spectrum was the most important clue, carrying the most weight in helping the computer decide on the molecule's overall shape. The infrared spectrum was the second most important, while the nuclear magnetic resonance data played a smaller role. Surprisingly, the computer did not treat all three data sources equally at every step. Instead, it seemed to use the Raman data to build the main skeleton of the molecule, then used the infrared data to identify specific groups of atoms, and finally used the nuclear magnetic resonance data to double-check the connections.

A particularly interesting discovery was how the computer handled the very first step of its thinking. The system starts by generating a special starting symbol, which acts like a flag to tell the rest of the program what kind of molecule it is looking at. The researchers found that this starting symbol effectively separates molecules that form a closed ring from those that are just open chains. Before the computer even begins to write out the full structure, this single starting point has already decided whether the molecule is circular or linear, acting as a global anchor that guides the entire process. This insight suggests that the model has learned a fundamental rule of chemistry: the overall shape of a molecule dictates how its parts fit together.

The team did not stop at simulated data; they tested their system on real-world experimental spectra from standard chemical databases. Even with the messy, imperfect data found in real laboratories, the model performed well. On a dataset of thousands of real infrared spectra, the system correctly identified the structure on the first try about 22 percent of the time, and improved to over 38 percent when allowed five guesses. While this is lower than the performance on perfect simulated data, it proves the system can handle the noise and complexity of the real world. The researchers noted that the system still struggles with certain difficult tasks, such as counting exactly how many carbon atoms are in a long chain, which remains a challenge for any method relying solely on these spectra.

This work demonstrates that it is possible to build a powerful tool for identifying chemicals that relies almost entirely on the raw data itself, free from the heavy constraints of previous methods. By combining a simple, efficient architecture with a smart way of learning from mistakes, the researchers have created a system that sets a new standard for accuracy. The findings suggest that while we may not yet have a perfect machine that can solve every molecular puzzle instantly, we are moving toward a future where computers can read the language of light and magnetism with a clarity that rivals human experts, offering a reliable and interpretable path forward for chemical analysis.