Fig. 3.

The iterative active learning framework for extracting knowledge with metrics. (A) The iteration is initialized with 486 data points, which are uniformly dispersed in the parameter space. Through training 20,000 NNs of the same structure (“4/32/32/1”), a committee of more than 12,000 models is established. The committee’s accuracy is estimated to be less than 12% relative error by acquiring selective data from simulations. The newly collected data are utilized in the next iteration. After five iterations, 362 new data points are collected, and a group of NNs with less than 5.0% relative error is obtained. (B) The premise that $\epsilon \left(\underset{¯}{x}\right)$ and $\stackrel{\sim }{\epsilon }\left(\underset{¯}{x}\right)$ are colocalized (as described Eq. 4) is verified with the newly selected points in each iteration. As expected, $\stackrel{\sim }{\epsilon }\left(\underset{¯}{x}\right)$ is a good approximation to $\epsilon \left(\underset{¯}{x}\right)$.