Figure 4
From: Efficient Closed-loop Maximization of Carbon Nanotube Growth Rate using Bayesian Optimization
BO increased CNT growth rate (\(\nu \)) up to a factor of 8 and improved its prediction over time – thus effectively demonstrating learning. (a) The raw growth rate of seed and planned experiments for the two BO campaigns, BO-1 (bottom panel) and BO-2 (top panel), increased as BO optimized the objective function \(\sqrt{\nu }\). The inset in (a) shows example growth curves obtained by ARES from a seed and planned experiment. (b) Central moving average of \(\nu \) (\({\nu }_{c.m.a.}\), calculated using the experimental data in panel (a) with a sample window size of 13 datum points) and predicted growth rates (\({\nu }_{pred.}\), provided by BO). (c) Normalized difference (Δ) between the central moving average and predicted growth rate for the two campaigns. BO improved the growth rate after only ~105 experiments regardless of how the seed was generated or the number of experiments within the seed.
