Fig. 2: Demonstration of how machine learning helps in achieving a knowledge upgrade.

a Flow chart highlighting the pathways leading to a knowledge upgrade with (② supervised training) or without (① unsupervised training) existing domain knowledge to extract meaningful features for ML. Upgraded knowledge is relative to existing knowledge in each domain, determined by the scientific problems we aim to solve. Two examples showcase the possible existing domain knowledge (b, d) and upgraded knowledge (c, e, f) for practical problems during ML. b, c Towards fast materials screening, the ML approach reveals key conditions for efficient Ce³⁺-activated scintillators and predicts good candidates¹⁵. b Left panel: 4 f vacuum-referred binding energies (VRBE) E_4f (m, Q, A) of the divalent (Q = 2 + ; red squares) and trivalent (Q = 3 + ; blue squares) lanthanide ions; m represents the number of electrons in the 4 f shell: m = n for Q = 3 + , and m = n + 1 for Q = 2 + ; “A” represents the chemical environment of the lanthanide ions. Right panel: Scheme depicting the changes in 4f-shell and 5d-shell electron binding energies in Ce³⁺ from a vacuum to a chemical environment A. c, Stacked-band scheme showing the DFT-computed relative valence and conduction band edges and ML-predicted VRBEs of an electron in the 4f and 5d₁ levels of the Ce³⁺ activator for elpasolite compounds; known scintillators are highlighted with blue bars. d–f Towards the discovery of new aggregation-induced emission (AIE) materials, the ML approach predicts and understands the AIE effect based on quantum mechanics. d Triphenylamine (TPA) core: green circle represents central nitrogen atom, and gray ellipses represent adjacent carbons with charges E_i¹⁶. e Classifiers trained by N-3C and 3 C show similar performance, while the single N yields the worst result. f Classification of the qualities of the parameter D at different threshold levels; the dot-dash line refers to the best threshold