# What are some examples of active learning methods used in atomistic machine learning?

Many machine learning attempts in atomistic applications (see this answer) seem to parameterize models on calculated data (i.e., CCSD(T), DFT, etc.). This approach suggests some automatic procedure for active machine learning in this context.

So the question is: what is the current status of active machine learning methods for atomistic applications? Specifically, I want to know which algorithms are more appropriate for chemical problems.

• My edit should not only bump up this unanswered question, but should also help make it more answerable. Jun 12 '20 at 22:38
• It has also been posted on Twitter 3 times (once, and then 2 re-tweets with new comments, and new tags). You can re-tweet all three of the Twitter posts to get more attention. Jun 12 '20 at 23:02

## 1 Answer

Inverse designing of materials with known target properties is of great importance (to reduce time, labour, financial etc. costs) than the traditional way of materials design which is guided by human intuition, followed by trial and error loops (see figure below from Active learning in materials science with emphasis on adaptive sampling using uncertainties for targeted design ).

Active learning optimization methods are generally used to iteratively guide experiments such that only a small number of promising experiments are carried out to discover new stable materials with target properties. During the iterative loop, the algorithm can learn from both successful and failed previous experiments to guide the next experiment.

To formulate a simple example, imagine you want to discover new stable quaternary compound in the quaternary system $$\ce{Ba - Sr - Mn - O}$$. One would start the experiment from $$\ce{BaSrMn2O6}$$, and the active learning algorithm learns from the existing data and the success/failure of this experiment to guide the next point to experiment in the quaternary system, such that smaller number of steps will be needed to reach convergence, which is to find a new stable compound. The same example can be reiterated to find new quaternary compound with target dielectric constant etc. It is specifically very useful when the compound search space is huge, but only very few experimental data is available on such compounds.

In theory, we want to find the design x, that maximises a desired function f(x). Bayesian global optimisation (BGO) is a powerful algorithm used to search for extrema of objective functions which are costly and complex. BGO is applied in active learning to reach convergence. Describing BGO is beyond the scope of this answer and I suggest you look into 1 to understand the details. I also recommend the paper Experimental search for high-temperature ferroelectric perovskites guided by two-step machine learning that deploys ML classification, regression and active learning techniques to discover new ferroelectric materials.

Finally, I have written an answer here describing the current trend of ML in materials science. Check that out too.

• +10. This might have been what OP was asking for! Jun 13 '20 at 1:35