Lab-scale Machine Learning: Tales of the good, the bad and the average

Abstract

Machine Learning and Artificial Intelligence are presented as the fix-all for current day problems. Also in research it is experiencing a golden age. However, before a Machine Learning model can be created, an enormous quantity of training data needs to be generated. This stands in stark contrast to general academic and industrial lab-scale data sets resulting from research projects. The latter give rise to small or even extremely small data sets (< 50 samples). This makes many of us wonder: "Is it possible to train an ML model with 30 samples instead of 30.000.000?" Using some simple regression models, I'll show that these can be successful in creating a suitable model in the (very) small data regime. Real life use-cases considering adhesive coatings, solvable inks, and spray-coating are discussed. I'll present a strategy to always obtain the best model and highlight caveats and ways to deal with them. [1] "A machine learning approach for the design of hyperbranched polymeric dispersing agents based on aliphatic polyesters for radiation curable inks",