Reproducible, Replicable, Reusable Science in Computational Geodynamics
Abstract
Reproducibility (the ability to re-run an experiment with the same outcome), and replicability (that the experiment can be independently verified) are core concepts in science and they ought to be straightforward to implement in computational disciplines.
It can be surprisingly hard to implement workflows that make results reproducible, particularly after a number of years have elapsed. Replicability contains a component of subjectivity: what does it take to have an independent result ? Do the algorithms need to be independent or only the implementation ? What does it mean to get the same answer in a numerical environment where errors are not derived from experimental "noise" ? Going beyond each of these concepts is the idea that numerical models should be re-usable too. This means that it is possible to build new models and results from existing ones. This extends the expectations of publishing results to include the fact that people should be able to validate the numerical analysis and extend it with their own idea and data. I will discuss how to approach reproducibility for numerical tools, how to frame questions so that they can be reproduced with different numerical methods, and how to publish computational results and the associated codes in a way that allows others to extend them - even when massively parallel computing is required. On a personal note, these ideas were refined as a result of many conversations with Louise Kellogg at CIG and CIDER workshops.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMDI33C0047M
- Keywords:
-
- 0545 Modeling;
- COMPUTATIONAL GEOPHYSICS;
- 8120 Dynamics of lithosphere and mantle: general;
- TECTONOPHYSICS;
- 8124 Earth's interior: composition and state;
- TECTONOPHYSICS;
- 8180 Tomography;
- TECTONOPHYSICS