Skip to main content
. 2020 Jul 30;11(32):8517–8532. doi: 10.1039/d0sc01101k

Fig. 2. Simulated performance of various agents in discovering stable Fe–X binary compounds. Ndiscovered and NDFT correspond to the total number of materials that are within 0.1 eV per atom of the instantaneous convex-hull, and the DFT budget (in terms of total number of DFT calculations agent is allowed to request), respectively. Agents are labeled as follows: M, AB, GPheur, SVGP and QBC which respectively correspond to “greedy” agent with a surrogate model, AdaBoost based agent, heuristic GP agent, stochastic variational GP agent, and query-by-committee based agent. The parameters ε (fraction of random exploration) and α (mixing parameter for uncertainty) are followed by their respective values in labels. The “one-shot” points refer to direct acquisition of N = 200 or 500 candidates, with the noted agent (M refers to the NN regressor, and AB refers to the same regressor boosted with AdaBoost), with no sequential (batch-mode) acquisitions. Presence of “div” in the label means the agent uses diversification. Ideal-limit shows the maximum achievable performance, i.e. if every agent-designated material was deemed stable and therefore “discovered”, outside of which is not accessible by any agent and hence grayed out.

Fig. 2