Machine learning workflows beyond linear models in low-data regimes.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Data-driven methods and machine learning accelerate chemical discovery and sustainability.
• Non-linear models are powerful for large datasets but face skepticism in low-data scenarios due to interpretability and overfitting concerns.
• Linear regression is traditionally favored in data-limited situations for its simplicity and robustness.

Purpose of the Study:

• To introduce automated workflows for non-linear machine learning models in data-limited chemical research.
• To address challenges of overfitting and interpretability in non-linear model application.
• To demonstrate the performance of non-linear models against traditional linear regression in low-data regimes.

Main Methods:

• Development of automated workflows featuring Bayesian hyperparameter optimization.
• Incorporation of an objective function to mitigate overfitting in both interpolation and extrapolation.
• Benchmarking on eight diverse chemical datasets with sizes ranging from 18 to 44 data points.

Main Results:

• Properly tuned and regularized non-linear models perform comparably to or better than linear regression on small chemical datasets.
• Interpretability assessments confirm that non-linear models capture chemical relationships effectively.
• De novo predictions demonstrate the predictive power of the developed non-linear workflows.

Conclusions:

• Automated non-linear workflows can overcome traditional limitations in data-limited chemical research.
• Non-linear models, when properly optimized, offer a viable and powerful alternative to linear regression.
• These tools enhance chemists' capabilities for tackling problems with sparse data, complementing existing linear approaches.