SHENZHEN, China, October 12, 2026 — XtalPi (HKEX: 2228) has open-sourced LCMS2Yield, an AI model that works with experimental measurements to determine product concentrations and chemical reaction yields without requiring a pure reference standard for each target compound. The GitHub release includes model weights, inference code, and datasets of molecular structures and corresponding UV correction factors used for training and evaluation.
Preparing pure reference standards adds time and effort to conventional yield measurements, particularly in high-throughput experiments involving newly synthesized compounds. LCMS2Yield predicts full-spectrum UV correction factors from molecular structures. Researchers combine these predictions with liquid chromatography–UV (LC-UV) measurements and an internal standard of known concentration to calculate product concentrations and reaction yields.
The model combines a Graph Transformer with a Kolmogorov–Arnold Network (KAN). It was pretrained on quantum chemistry–derived correction factors for 22,609 compounds, then fine-tuned on experimentally measured correction factors for 1,845 compounds. This approach uses computational data to learn relationships between molecular structures and correction factors, with experimental data adapting the model to laboratory measurements.
Across 178 reactions spanning four common reaction types, the method achieved a mean absolute yield error of 4.74% compared with reference yields obtained using conventional calibration curves. The team also tested concentration estimates using 144 FDA-approved drug compounds. The research was previously published in JACS Au. Applicability depends on the compounds’ UV absorption characteristics and the analytical conditions.
The repository provides setup instructions and usage examples. Researchers and developers can run predictions, evaluate the model with their own experimental data, reproduce predictions using the released datasets, and develop model improvements.
The research team views LCMS2Yield as an example of connecting scientific AI with automated experimentation: models support analysis and decision-making, while experiments provide data and validation. The open-source release makes the model and supporting resources available for further research and integration into laboratory workflows.
GitHub repository: https://github.com/xtalpi-inc/lcms2yield
Research paper: AI-Predicted Full-Spectrum UV Correction Factors: An Analyte-Standard-Free Framework for Reaction Yield Quantification
