Uncertainty Analysis of Molecular Quantum Properties Prediction Using Chemical Foundation Models

Large pre-trained foundation models are becoming prevalent and have a high risk impact in domains of the physical sciences. Despite significant advancements in predicting molecular and reaction properties, many foundation models struggle when applied to real-world scenarios. In such cases, uncertainty analysis of prediction results can help engender trust in the model outcomes and indicate reliability to decision makers. In this study, we introduce a method for uncertainty quantification and characterization tailored to chemical foundation models, with a focus on predicting quantum molecular properties. Our approach is tested through an investigation into the uncertainties inherent in the predictions of a chemical foundation model, specifically when predicting quantum molecular properties such as the highest occupied molecular orbital (HOMO) energy, the lowest unoccupied molecular orbital (LUMO) energy, and the dipole moment (R2). We apply our method to a SMILES-based foundation model and examine how these uncertainties vary between a fine-tuned version of the model and a version where the model parameters remain frozen during these property predictions. The results demonstrate the effectiveness of our approach in identifying and quantifying uncertainties, offering insights into model reliability and the impact of model fine-tuning on prediction results.