Developing a modern Fortran statistics and machine learning library (FSML)

 

Advances in computing, statistics, and machine learning (ML) techniques have significantly changed research practices across disciplines. Despite Fortran’s continued importance in scientific computing and long history in data-driven prediction, its statistics and ML ecosystem remains thin. FSML (Fortran Statistics and Machine Learning) is developed to address this gap and make data-driven research with Fortran more accessible. 

The following points are considered carefully in its development and each come with their own challenges, solutions, and successes: 

• Good sustainable software development practices: FSML is developed openly, conforms to language standards and paradigms, uses a consistent coding and comment style, and includes examples, tests, and documentation. A contributor’s guide ensures consistency for future contributions. 

• Accessibility: FSML keeps the code clean and simple, avoids overengineering, and has minimal requirements. Additionally, an example-rich html documentation and tutorials are automatically generated with the FORtran Documenter (FORD) from code, comments, and simple markdown documents. Furthermore, it is developed to support compilation with LFortran (in addition to GFortran), so it can be used interactively like popular packages for interpreted languages. 

• Community: FSML integrates community efforts and feedback. It uses the linear algebra interfaces of Fortran’s new de-facto standard library (stdlib) and the fortran package manager (fpm) for easy building and distribution. Its permissive licence (MIT) allows developers to integrate FSML into their projects without the restrictions often imposed by other licenses. Its simplicity, documentation, contributor’s guide, and GitHub templates remove barriers for new contributors and users. 

• Communication: FSML updates are shared through a variety of methods with different communities. This includes a journal article (https://doi.org/10.21105/joss.09058) for visibility among academic colleagues, frequently updated online documentation (https://fsml.mutz.science/), social media updates, as well as a blog and Fortran Discourse posts to keep Fortran’s new and thriving online community updated. 

Early successes of FMSL’s approach and design include: 1) Students with little coding experience were able to learn the language and use library with only Fortran-lang’s tutorials and FSML’s documentation; 2) early career researchers with no prior experience in Fortran used FSML’s functions to conduct research for predicting future climate extremes; 3) FSML gained a new contributor and received a pull request only days after its first publicised release. 

The development of FSML demonstrates the merits of using good and open software development practices for academic software, as well as the potential of using the new Fortran development ecosystem and building bridges to the wider (non-academic) developer community.