Scientific Machine Learning Benchmark (smlb)

Introduction

smlb is a toolbox focused on enabling rigorous empirical assessments of data-driven modeling approaches for applications in the natural sciences. It is particularly useful when developing or fine-tuning data-driven algorithms to ensure statistically sound decisions. Its focus is on models for experimental and computed properties of molecules and materials. It emphasizes correctness, flexibility, and domain support.

smlb was designed to help answer questions that arise during the development of domain-specific machine-learning models. Examples of such questions include

• Which of these uncertainty estimate approaches most closely matches the true error distribution?
• How does removing slow or failing features affect the predictive accuracy of my model?

smlb provides

• ready-to-use synthetic, computational and experimental datasets
• bindings to other software, including domain-specific features and general machine-learning packages
• standard loss functions and error metrics, also for predictive distributions (uncertainties)
• reproducibility by systematic control of pseudo-random number generation

Other uses include integration tests to ensure that local changes to a modeling pipeline do not have overall adverse effects.

See the Overview for a more detailed description.

Getting started

To get started, follow installation instructions and run the tutorial.

Other

To contribute, see the Contributing instructions.

Related work
Acknowledgments