4.0.0 Open Citrination update

Improved optimizer

We've improved the quality of our optimizer and our models are now able to more quickly identify optima. We tested this new approach against 5 synthetic benchmark functions, and this new method was proven effective in high- and low-dimensional space.

User Impact: This behavior is now default on all new design runs on Citrination.

Improved data type representation

We've improved internal representations of complex data types such as inorganic and organic compounds and their properties. This includes representing materials as distributions over plausible realizations such as mixtures for which the exact composition is unknown, or multiple inorganic compounds that reasonably match the characterization data.

User Impact: Because of this change, we no longer support Infinity bounds for real inputs or outputs. New data views will require users to provide real upper bounds on real descriptors. If not specified, the upper bound of each real descriptor will be 2X the maximum value in the training set.

Default uncertainty recalibration

Recalibration is now applied to our uncertainty estimates to keep them well calibrated with significantly less computational effort.

User Impact: This change is on by default for all new views on Citrination with this release. Users should not see a performance impact with new data views or models

Randomized pivots now default in Lolo models

This change creates robust interpolation between training points placed on a grid.

User Impact: This feature is on by default for new views on Citrination with this release.


--Article published 3/29/19--


Feedback and Knowledge Base