Fig. 5: Simulated Dual-GP, human-in-the-loop, uncertainty-based exploration in 4D space using experimental data from an autonomous PLD experiment.
From: Active oversight and quality control in standard Bayesian optimization for autonomous experiments
a, b The ground truth response surface projected into the P vs. T and E1 vs. E2 planes. c, d The Dual-GP reconstruction closely matches the ground truth. e, f Random sampling performs better than (g, h) pure uncertainty exploration with traditional GPBO. The points in each map represent sampled points, where the red color indicates a high score. The axes of each map are normalized. i The root mean squared error of ground truth reconstruction vs sample number for all three cases shows that human-in-the-loop Dual-GP quickly outperforms random sampling and traditional GPBO when using uncertainty-based exploration, which is attractive for synthesis science applications where pure optimization is not the goal of the experiment. Examples of the human score (ranking 1–10) are shown in j–l and compared with the original Raman score which was based on a more complicated peak-fitting model.
