Solving dependencies for conda packages used in computational chemistry is spectacularly annoying. But while I’m at it – why not write something here.

It’s been three months of work with machine learning, and over four full years of theoretical modelling overall. I learned an important lesson: the values your model predicts are not important. It is relatively easy to make a prediction using numpy, scipy and a bunch of polynomials.

It is far more important to know the applicability domain of the model, its error and trust intervals of obtained parameters. Clean the dataset, reduce the number of parameters, write out the necessary equations and try to reduce the error using first-principle approach. Modern science studies deviations and reproducibility, and rarely values themselves. Which is a good thing.