Adrian Mirza and Dr. Kevin Jablonka were sipping coffee in a quiet corner of Friedrich Schiller University Jena when their AI system, SECS, proposed a molecular structure in just 47 seconds—a task that once took seasoned chemists hours, if not days. That moment marked a quiet revolution in chemical analysis, one now unfolding across labs in Jena and beyond. Determining the exact structure of a newly synthesized molecule has long been one of chemistry’s most painstaking puzzles, relying on a patchwork of techniques like NMR spectroscopy, infrared scans, and mass spectrometry. Each method offers a clue, but the full picture often remains elusive—especially with impure or novel compounds. SECS, an open-access AI developed by researchers from the University of Jena, Helmholtz-Zentrum Berlin, the Helmholtz Institute for Polymers in Energy Applications, and Swiss software firm Zakodium Sárl, is changing that. By translating raw NMR spectra into plausible molecular structures in minutes, it’s not just accelerating discovery—it’s making it more accessible.
SECS combines contrastive learning with an evolutionary algorithm, creating a shared mathematical language between spectra and molecular structures. It then iteratively tweaks candidate molecules—adding or removing atoms and bonds—testing each against the spectral data until it finds the best fit. The result is a ranked list of possible structures, each with a similarity score grounded in chemical context. In benchmark tests, SECS identified the correct molecule as its top prediction in over 80% of cases, matching the accuracy of human experts. In a pilot study where chemists tackled 20 complex structure-identification problems, the AI performed on par with seasoned specialists—without fatigue, bias, or the need for a lab coat.
Yet the team is clear: SECS isn’t here to replace chemists. "We do not see SECS as a replacement for human expertise," Mirza emphasizes. Instead, it serves as a powerful second opinion. When the AI’s top suggestions align with a chemist’s hypothesis, confidence grows. When they diverge, it’s a signal to dig deeper—potentially uncovering errors or unexpected chemistry. The system’s ability to handle real-world data, including impurities that muddy spectral signals, makes it especially valuable for practical research. And because SECS is open-source—its code, models, and a test version freely available—it invites collaboration and adaptation across the global scientific community.
Already, researchers are using the web-based tool to analyze one-dimensional proton NMR data, with plans to expand to more complex spectra. In an era where discovery often hinges on speed and precision, SECS offers both. It’s not just a tool for solving molecular puzzles—it’s a step toward democratizing scientific insight, one spectrum at a time.