Limitations of theory-informed machine learning algorithms for the prediction and exploration of molecular properties: Solvation free energy as a case study

Significant advancements in machine learning have accelerated and improved structure–property predictions for materials discovery. However, data are often scarce due to large parameter spaces consisting of chemistry, structure, synthesis, and processing variables. At small data limits, theory can be leveraged to inform machine learning (ML) models with domain knowledge and improve generalization. Here, we determine how the accuracy of first-principles calculations affects theory-informed ML predictions of the experimental solvation free energy ΔGsolvexp in both “small” (101–102) and “large” (103–104) data size limits. We compare several existing theory-informed techniques to a baseline (no theory) model: feature-informed, difference, and ratio. At small data limits, all theory-informed models exhibit lower RMSE, reducing training data size by more than 65% compared to the baseline model. With larger training set sizes and as theory prediction accuracy declines, the difference and ratio models exhibit larger errors than the baseline model, indicating negative transfer occurs. No negative transfer is observed in the feature-informed model; however, the model is unable to extrapolate outside of the known ΔGsolvexp distribution, whereas the difference and ratio models exhibit lower extrapolation error. Finally, we employ each model in an active learning algorithm and compare two exploration acquisition functions: maximum difference and maximum variance. Sampling with the maximum difference policy reduces RMSE and variance of the feature-informed model faster than maximum variance early in the exploration campaign, as it identifies label bounds in fewer iterations. Our findings highlight the balance between leveraging theory and relying on data-driven models in high-throughput materials discovery.