Table 1 Summary of representative machine learning works on the direct prediction of lattice thermal conductivities in recent 5 years.

the κ_L of 120 compounds at 300 K, obtained from first-principles calculations¹⁵

maximum phonon frequency, integrated Grüneisen parameter up to 3 THz, average atomic mass, and volume of the unit cell

GPR, SISSO

the 2146 κ_L of 119 compounds at various temperatures from 100 to 1000 K, obtained from first-principles calculations¹⁶

before filtering: 126 crystal features mainly generated based on Wigner-Seitz cells, 145 elemental features, and temperature

RF

the κ_L of 110 compounds at 300 K, obtained from first-principles calculations²⁸

a set of descriptors from simple elemental and structural representations

BO

the κ_L of 5486 compounds at 300 K, obtained from the AGL method⁴⁴

7 crystal properties, 25 CW elemental properties, 25 crystal structure fingerprints, and 18 statistical properties

XGBoost

the κ_L of 1700 RMLs and their corresponding 1700 GMLs at 30 K, obtained from MD simulations⁴⁵

the thickness sequence, the thickness- and period-based index of randomization, the standard deviation of thickness, the maximum and minimum deviation of layer thickness with respect to the mean layer thickness of RML

NN

the κ_L of 86 HH compounds at 300 K, obtained from first-principles calculations or experimental measurements⁴⁶

lattice constants and 24 elemental features

SISSO

initialization: the κ_L of 100 porous graphene structures at 300 K, obtained from MD simulations⁴⁷

a gray image representing the spatial distribution of the holes

CNN + active learning

initialization: the κ_L of 100 hybrid carbon-boron nitride honeycombs (C-BNHCs) at 300 K, obtained from MD simulations⁵⁰

a grayscale image representing the top-view schematic of C-BNHC

CNN + active learning

(1) the κ_L of 2668 compounds at 300 K, obtained from high-throughput first-principles calculations. (2) the κ_L of 132 compounds at 300 K, obtained from experimental measurements⁵¹

a graph representing the connection of atoms in the crystal

CGCNN + transfer learning

(1) the harmonic three-phonon scattering phase space of 320 crystals, obtained from first-principles calculations. (2) the κ_L of 45 crystals at 300 K, obtained from first-principles calculations⁵³

290 elemental features

NN, RF + transfer learning

initialization: the κ_L of 300 randomly generated Si/Ge RMLs at 300 K, obtained from MD simulations⁵⁷

N-bit array: N is the number of unit cells in the RML, and one can input a value of 1 or 2 depending on whether the corresponding unit cell consists of Si or Ge atoms, respectively

CNN + active learning