A unified understanding of minimum lattice thermal conductivity

Yi Xia,1, ∗ Dale Gaines II,2 Jiangang He,3 Koushik Pal,2 Mercouri

G. Kanatzidis,4 Vidvuds Ozoliņš,5, † and Chris Wolverton2, ‡
1
Department of Mechanical and Materials Engineering, Portland State University, Portland, OR 97201, USA
2
Department of Materials Science and Engineering, Northwestern University, Evanston, IL 60208, USA
3
School of Mathematics and Physics, University of Science and Technology Beijing, Beijing 100083, China
4
Department of Chemistry, Northwestern University, Evanston, IL 60208, United States,
Materials Science Division, Argonne National Laboratory, Argonne, IL 60439, United States
5
Department of Applied Physics, Yale University, New Haven, CT 06511, USA,
Energy Sciences Institute, Yale University, West Haven, CT 06516, USA
(Dated: June 22, 2023)
arXiv:2306.11917v1 [cond-mat.mtrl-sci] 20 Jun 2023

We propose a first-principles model of minimum lattice thermal conductivity (κminL ) based on a

unified theoretical treatment of thermal transport in crystals and glasses. We apply this model
to thousands of inorganic compounds and discover a universal behavior of κmin L in crystals in the
high-temperature limit: the isotropically averaged κminL is independent of structural complexity
and bounded within a range from ∼0.1 to ∼2.6 W/[m·K], in striking contrast to the conventional
phonon gas model which predicts no lower bound. We unveil the underlying physics by showing
that for a given parent compound κmin L is bounded from below by a value that is approximately
insensitive to disorder, but the relative importance of different heat transport channels (phonon
gas versus diffuson) depends strongly on the degree of disorder. Moreover, we propose that the
diffuson-dominated κmin
L in complex and disordered compounds might be effectively approximated
by the phonon gas model for an ordered compound by averaging out disorder and applying phonon
unfolding. With these insights, we further bridge the knowledge gap between our model and the well-
known Cahill-Watson-Pohl (CWP) model, rationalizing the successes and limitations of the CWP
model in the absence of heat transfer mediated by diffusons. Finally, we construct graph network
and random forest machine learning models to extend our predictions to all compounds within the
Inorganic Crystal Structure Database (ICSD), which were validated against thermoelectric materials
possessing experimentally measured ultralow κL . Our work offers a unified understanding of κminL ,
which can guide the rational engineering of materials to achieve κmin
L .

I. INTRODUCTION Pohl [9] (referred to as the CWP model) to incorporate

the Debye model of vibrations by adopting a wavelength-
Knowing the lower limit to the lattice thermal conduc- dependent mean free path (MFP). An earlier approach
tivity of crystals is of fundamental interest and technolog- by Slack [2] as well as the more recent study by Clarke
ical importance, particularly relevant to thermal energy [7] also assumed that phonons are the dominant heat car-
conversion and management applications [1]. The mini- riers, but used different assumptions about the phonon
mum lattice thermal conductivity (κmin MFP. The CWP, Slack, and Clarke theories of κmin are all
L ), a concept first L
proposed by Slack [2], sets the upper limit of thermoelec- based on the conventional phonon gas model (PGM) and
tric conversion efficiency for a given electronic transport the Boltzmann gas kinetic equation [14], where the pri-
profile [3–6] and provides the maximum thermal insula- mary heat carriers are propagating phonons [10]. On the
tion for a substrate used in thermal barrier coatings [7, 8]. other hand, recognizing the potential failure of the PGM
The fundamental physics of heat transport in crystals at in the disordered regime, Allen and Feldman [15, 16]
the lower limit of thermal conductivity continues to be developed the diffuson theory of heat transport in dis-
an active research topic [2, 7, 9, 10]. ordered solids. The Allen-Feldman theory is based on
Motivated by the observation that the experimentally Kubo-Greenwood formula in the harmonic approxima-
accessible κmin is often found in amorphous solids, the tion and the primary heat carriers are diffusons, which
L
initial attempts to understand κmin [11, 12] relied on are described by the off-diagonal terms of the heat flux
L
a model proposed by Einstein [13], who assumed that operator. Recently, a phenomenological κmin L model was
the mechanism of heat transport in crystals was a ran- developed by Agne, Hanus, and Snyder [10] based on the
dom walk of the thermal energy between neighboring diffuson theory, aiming to resolve the overestimation of
atoms vibrating with random phases. Einstein’s random κmin
L from the CWP model in comparison to experiments.
walk model was later generalized by Cahill, Watson, and Despite significant advances, a comprehensive under-
standing of κmin
L is still elusive. A crucial yet missing
piece of the puzzle is a first-principles-based theory that
accurately describes κmin
L over the whole spectrum of ma-
∗ [email protected] terials between the prototypical classes of simple crystals
† [email protected] and disordered amorphous solids. The development of
‡ [email protected] such an advanced model would shed light on crucial open
 2

questions. For instance, the heat transfer pathways me- Here, e, D, and R denote the polarization vector, the dy-
diated by localized diffusons and propagating phonons namical matrix, and the lattice vector, respectively. α/β,
are often classified into independent channels [17–21]. Is m, and p/q are indices labeling the Cartesian coordinate,
there a unified physical picture of the two different kinds the unit cell, and the atoms within the unit cell.
of heat carriers when approaching κmin L ? Why does the Physically, we can further decompose κmin L into two
CWP model work remarkably well for solids with strong parts: the diagonal part with s = s′ , which corre-
disorder, despite retaining the PGM nature and neglect- sponds to the PGM, denoted as PGM-κmin L , and the off-
ing all optical phonons [9, 11, 12]? There are recent ex- diagonal part, which accounts for the diffuson channel,
perimental discoveries of crystalline compounds with ul- denoted as OD-κmin L . The total κL
min
will be referred to
tralow and glass-like lattice thermal conductivity κL that as PGM+OD-κmin L . Immediately, we see that the eval-
resembles κmin
L [22–26]. What is the general behavior of uation of Eq.[1] only requires the knowledge of the har-
κmin
L in crystals, i.e., how does it vary with structural monic phonon dispersion, making it amenable to large-
complexity and atomic composition? Ultimately, is there scale DFT calculations. We applied Eq.[1] to a selected
a physical lower or upper bound to κmin L ? set of 2576 inorganic compounds, taking advantage of the
tabulated harmonic force constants within the phonon
database (phononDB) generated by Togo et al. [31, 32]
II. RESULTS AND DISCUSSIONS using crystalline structures from the Materials Project
(MP) [33–35].
Recent theoretical advances towards a unified theory Temperature enters Eq. (1) both explicitly in the pref-
of thermal transport in crystals and glasses have enabled actor and implicitly via the phonon occupation numbers
a consistent treatment of different heat carriers such as nsq . Because κmin
L increases monotonically with increas-
propagating phonons and localized diffusons [28, 29]. ing temperature up to the Debye temperature as a re-
Herein, we construct a κminL model based on the uni- sult of enhanced heat capacity [9], and κmin
L in crystals is
fied theory of Simoncelli, Marzari, and Mauri [28, 30] usually approached at high temperatures, we focus on ex-
by adopting the approximation that each phonon mode’s ploring the high-temperature regime. We performed cal-
lifetime is equal to one half of its vibrational period, culations at T = 300, 600, and 900 K. We find that κmin L
τqs = π/ωqs . This assumption follows the original pro- of most compounds calculated at 600 K exhibit consid-
posal by Einstein [13] which was later adopted by the erable increment over those calculated at 300 K, whereas
CWP model [9]. We choose not to use the minimum κmin
L varies by less than 5% between T = 600 and 900 K
MFP criteria (e.g., by assuming that MFP is equal to for approximately 95% of the studied compounds, with
the smallest atomic spacing) because for optical phonon the largest decrease being only 14%. Considering the
modes with vanishing group velocities it will lead to in- larger availability of experimental measurements of κL at
finite phonon lifetimes, which is unphysical and tends to 600 K than 900 K, we will focus on the T = 600 K results
diminish their off-diagonal contribution to thermal con- in the following discussion. For anisotropic crystals, the
ductivity (see below). Using the formula for thermal components of κmin L are further averaged along the three
conductivity from Ref. [28], we arrive at the following principle crystallographic axes. We refer the readers to
expression: the Methods section for a more detailed discussion of the
structure types, chemical compositions, and criteria for
πℏ2 X X (ωqs + ωqs′ )2 numerically converging κmin L .
′ ′
κmin
L = vqs,s ⊗ vqs ,s
kB T 2 V Nq q ′
2 The calculated values of PGM-κmin min
L , OD-κL , and
s,s
′ ′ ′
(1) min
PGM+OD-κL as functions of the number of atomic
ωqs nsq (nsq + 1) + ωqs nsq (nsq + 1) sites in the primitive cell, Na , are plotted in Fig.1. We
·
4π 2 (ωqs′ − ωqs )2 + (ωqs + ωqs′ )2 associate the value of Na with the degree of structural
complexity of a compound. We see from Fig.1A that
where ℏ, kB , T , V , Nq are, respectively, the reduced PGM-κmin L covers a large range of values, spanning from
Planck constant, the Boltzmann constant, the absolute approximately 0.01 to 1 W/[m·K] in the high temper-
temperature, the volume of the unit cell, and the to- ature limit (600 K and above). There is a clear trend
tal number of sampled phonon wave vectors. Phonon that PGM-κmin L decreases monotonically with increasing
modes are denoted by the wave vector q and mode index Na , which can be clearly seen from the log-log plot in
s. The key quantities entering Eq.[1] are phonon mode- the inset of Fig.1A). In contrast, the total PGM+OD-
resolved frequencies ωqs , population numbers nsq , and the κmin
L in Fig.1B exhibits a much narrower spread from
′
generalized group velocity tensors vqs,s , with the latter approximately 0.1 to 2.6 W/[m·K], independent of Na .
calculated as [16, 28] The lower (0.1 W/[m·K]) and the upper (2.6 W/[m·K])
bounds were determined by analyzing the frequency dis-
′ i X X pq tribution histogram of all calculated κmin
L , as detailed in
vqs,s = esq (α, p)Dβα (0, m)
ωqs s
+ ωq ′
the Supplementary Materials section. There are only
m,p,q
α,β (2) four compounds within our calculations having values
′
× (Rm + Rpq )eiq·Rm esq (β, q). of PGM+OD-κmin L outside these bounds. The largest
 3

1 1
A 10 B 10 C

Ratio: OD-κLMin /PGM-κLmin [W/(m·K)]

0
10
2.6

PGM / PGM+OD-κLMin [W/(m·K)]

10-1
100 100
PGM-κLMin [W/(m·K)]

101
10-2

5 10 40 100
0.1
10-1 10-1
101

100
100
10-2 10-2
5 10 40 100

40 80 PGM
120 160 40 80 120 160 40 80 120 160
# of sites in primitive cell (Na ) # of sites in primitive cell (Na ) # of sites in primitive cell (Na )

FIG. 1. Contributions to minimum lattice thermal conductivity from the phonon gas model and the off-diagonal terms based
on Eq.[1]. (A) Calculated minimum lattice thermal conductivity (κmin L ) as a function of the number of atomic sites (Na ) in a
primitive cell using the phonon gas model (denoted as PGM-κmin L ), i.e., the diagonal part of Eq.(1). The inset is plotted with
−2/3
logarithmic scales and the dashed line is plotted following the scaling of Na obtained from Slack’s model [27], showing the
min
monotonic decay of PGM-κL with increasing Na . (B) The same as (A) but with additional red disks showing the calculated
κmin
L accounting for both the PGM and off-diagonal (OD) contributions, denoted as PGM+OD-κmin L . The gray dashed lines
indicate values of 0.1 and 2.6 W/[m·K] for the lower and the upper bounds of PGM+OD-κmin L , respectively. (C) The ratio of
OD-κmin
L and PGM-κmin L as a function of Na . The inset is displayed with logarithmic scales, showing the increasing trend of
OD-κmin min
L /PGM-κL . All results were obtained at 600 K.

deviation from the lower bound is found in crystalline pressed spectral weight and reduced group velocity. It
P 6(3)/mmc-Ar, which has a value of 0.043 W/[m·K], is worth noting that strategies based on the above ar-
and all the other compounds have values larger than gument have been successfully used to search for com-
0.081 W/[m·K]. Meanwhile, for the upper bound, the plex materials with intrinsically low lattice thermal con-
largest PGM+OD-κmin L is associated with P 4(2)/mnm- ductivity [26, 37]. However, our results show that these
TiO2 , with a value of 2.64 W/[m·K]. The comparison strategies may be no longer effective when phonons are
of PGM-κmin L and PGM+OD-κmin L reveals the increas- strongly scattered by anharmonicity or disorder when the
ing importance of the off-diagonal diffuson term OD-κmin L heat transport is dominated by the OD terms beyond
with increasing structural complexity Na . This is fur- the PGM. This is especially true for systems with κL ap-
ther illustrated by the ratio of OD-κmin L to PGM-κminL in proaching κmin
L . In this scenario, the breakdown of the
min
Fig.1C. We see that OD-κL quickly surpasses PGM- PGM is due to the crossover in the heat transport from
κmin
L above Na ∼ 10 and OD-κmin L becomes an order propagating phonons to localized diffusons, as initially
of magnitude higher than PGM-κmin L above Na ∼ 60. proposed by Allen and Feldman [15, 16]. As a result,
We also observe a group of compounds (e.g., P 3m1-ZnS) the total PGM+OD-κmin L provides a much more reason-
with ratios of OD-κmin
L to PGM-κmin L significantly smaller able estimation of κmin
L compared to the phonon term
than the majority trendline. Our analysis of their struc- PGM-κminL alone. Importantly, despite such a crossover
tures reveals that these compounds display extreme crys- from PGM-κmin L to OD-κmin
L , the bounds of the total
talline anisotropy, i.e., the lattice constant of the primi- PGM+OD-κmin L remain independent of Na . This sug-
tive cell along one axis is exceedingly large and thus Na is gests an effective interconversion between the two fun-
large. However, for these compounds, the ratios of OD- damentally different heat transfer channels, which might
κmin
L to PGM-κmin L are reduced along the other two axes be induced by variations in chemical composition, atomic
with much smaller lattice constants, thus giving rise to disorder, lattice distortion, structural complexity (Na ),
overall reduced ratios after averaging over the three spa- and other factors. This invites the following two ques-
tial directions. tions: (i) what are the key factors that determine the
The monotonic decay of PGM-κmin as a function of interconversion between PGM-κmin L and OD-κminL across
L
Na is not unexpected. As already reflected in the Slack various compounds? and (ii) is it possible to revert
model [27, 36], κL based on the PGM is proportional to such a interconversion, for example, via approximating
−2/3 PGM+OD-κmin L of compounds with large Na using only
Na provided that acoustic phonon modes with high
PGM-κminL ?
group velocities dominate κL . The monotonic decay of
PGM-κmin L can be attributed to the fact that the increase Answering the above two questions will help establish
in Na leads to a reduced Brillouin zone that effectively a unified understanding of κmin
L . The challenge arises
folds back the high-energy acoustic modes into the first from the fact that the crossover in the heat transport
Brillouin zone as optical modes, resulting in both sup- from phonons to diffusons is observed in a large set of
 4

0.7 0.7 11 11 8
A B C D
0.7 ● - PGM+OD ●●● ●●●

Group velocity (km/s)

●
● ● ● ● ● 10 10
● ● ● 6
0.6 ● 0.6 ●
● ●
w/o random perturbation 9 9
0.6 w/ random mass perturbation ●
4
w/ random IFCs perturbation ●
■ ■ ■ 8 8
w/ random POS perturbation 0.5 ▲ PGM 0.5 ■
▲ ▲ ■ ■ ●
Minimum κL (W/m·K)

Minimum κL (W/m·K)
OD

Minimum κL (W/m·K)
0.5 ▲ - Diagonal (PGM) ■
■ 7 7 2
▲

Frequency (THz)

Frequency (THz)
▲ ▲ ● PGM+OD
■
■ 0.4 ▲ 0.4 ■

■ ▲ 6 6 0
0 2 4 6 8 10
0.4 ▲
▲ ▲ ■
■ ■ ■ 8
■ 5 5 Frequency (THz)
0.3 ▲ 0.3
■

Group velocity (km/s)

■ ▲
■ ▲
▲ 4 4 6
0.3
■ 0.2 8-atom ▲
0.2 ▲ 64-atom
■ ■ primitive cell ■ primitive cell 3 3
▲ 4
▲
0.2 ▲ ▲ 2 2
■ ■ ■ 0.1 0.1 ▲
▲ ▲ ▲ 2
■ - Off-diagonal (OD) ▲ 1 1
0.1 0.0 0.0 0 0 0
1 8 64 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 Γ LΓ L 0 2 4 6 8 10
Number of sites in primitive cell Strength of random Strength of random Wave vector Wave vector Frequency (THz)
mass perturbation mass perturbation

FIG. 2. Impacts of disorder and structural complexity on minimum lattice thermal conductivity. (A) Calculated total κmin L
(PGM+OD-κmin L ) and the corresponding decomposed contributions (PGM-κL
min
and OD-κmin
L ) of a model system (fcc-Al) as
functions of the number of atomic sites (Na ) in the designed primitive cell and the presence of various kinds of disorder. We
artificially create three kinds of disorder in the perfect crystal of fcc-Al by applying small random perturbations (within 5% of
their original values) to atomic mass, interatomic force constants (IFCs), and atomic position (POS), denoted in blue, orange,
and yellow, respectively (with dashed line as a guide to the eye). (B) Calculated PGM+OD-κmin min
L (green disks), PGM-κL (blue
triangles), and OD-κminL (orange squares) of fcc-Al as a function of the strength of mass disorder, realized by applying random
percentagewise mass perturbation. For example, 20% random mass perturbation means a random mass ranging from -20%
and 20% of the original atomic mass of the element is added to each atom. The left and right panel show the results obtained
for an 8-atom and 64-atom primitive cell, respectively. The shaded areas indicate the uncertainty obtained by averaging many
independent random mass perturbations. (C) Phonon dispersions of the perfect fcc-Al (left panel) and an 8-atom primitive cell
with 20% random mass perturbations (right panel). (D) Phonon group velocities of the perfect fcc-Al (upper panel) and an
8-atom primitive cell with 20% random mass perturbations (lower panel). The κmin L shown in panel A and B were all calculated
at 300 K.

compounds with diverse characteristics. To reveal and cell is used as the primitive cell for the perfect fcc-Al
disentangle the impacts of chemical composition, inter- crystal. In contrast, when small perturbations are ap-
atomic interaction, and structural complexity on κmin
L , we plied, we notice different trends in PGM-κmin L and OD-
use a simple elemental structure, i.e., face-centered cubic κmin
L across the unit cells with increased N a (i) when
:
aluminium (fcc-Al), as a base model and then manually Na =1, perturbations lead to only small changes in both
generate perturbations on this model to mimic the varia- PGM-κmin L and OD-κminL , and the PGM-κL
min
dominates
min
tions in structures and chemistries among different com- over the OD-κL , as expected for a simple crystal; (ii)
pounds. Specifically, random perturbations are applied when Na =8, perturbations start to convert PGM-κmin L to
on the atomic mass, interatomic force constants (IFCs), OD-κminL , with the former still larger than the latter; (iii)
and atomic position. We also vary the number of atoms when Na =64, the values of PGM-κmin L and OD-κmin L are
(Na ) in the unit cell by repeating the primitive cell of nearly exchanged, and OD-κmin L becomes the major con-
fcc-Al. As a consequence, we can use these models to ap- tribution to the total κmin
L . Surprisingly, the total κL
min

proximate a variety of situations ranging from ordered to largely remains constant with both varying perturbations
disordered by varying the strength of perturbations and and Na .
Na . Note that we strictly enforce physical constraints The above results reveal that the total κmin L is not
when applying these perturbations, such as the transla- sensitive to small perturbations of atomic masses, po-
tional invariance of the crystal. sitions, and IFCs of a given base crystal structure. In
Fig.2A summarizes the calculated total κmin L and its the other words, the capability of lattice heat transfer of
decompositions (PGM and OD) when weak perturba- a given system that approaches maximum phonon scat-
tions are applied respectively to atomic masses, IFCs, tering (τqs ≈ π/ωqs ) can neither be reduced nor enhanced
and atomic positions in unit cells with increasing Na . significantly with small adjustments in atomic compo-
First, we observe that neither the total κminL nor its de- sitions, crystal structure, and interatomic interactions.
composition change by simply increasing Na without in- This is in contrast to the case of weak phonon scatter-
troducing disorder. This is as expected because nothing ing in some materials (e.g. diamond and BN), wherein
has been changed physically and merely a larger unit small mass perturbations such as isotope scattering can
 5

lead to a considerable reduction in κL [38]. Conversely, ment between the two, as can be inferred by comparing
the relative contributions of PGM and OD terms are OD-κminL without any disorder and the total κmin L with
very sensitive to small perturbations, especially for in- strong disorder in Fig.2B. These results help explain why
creasingly large unit cells. The significantly decreased we find that the total PGM+OD-κmin L in Fig.1B lies in
PGM contribution can be mostly attributed to the re- a relatively narrow range, which is bounded both from
duced (diagonal) phonon group velocities. As shown in above and below over a variety of compounds. It can be
Fig.2C and D, the introduction of mass disorder strongly interpreted as follows: (i) complex compounds are noth-
breaks the energy degeneracy of phonon bands and sup- ing but variants of simple compounds with additional
presses the (diagonal) group velocities, whereas phonon compositional and structural disorder; (ii) despite poten-
frequencies change by much less. This results in an in- tially dissimilar amounts of structural complexity, differ-
creased OD contribution due to the lifting of degeneracy, ent compounds share close values of κminL if their averaged
which makes the otherwise vanishing off-diagonal veloci- structural properties, such as mass, position, and inter-
ties appreciable [39]. We can also infer from Fig.2A that atomic interaction, are similar; (iii) it is likely that the
the PGM-κmin L tends toward zero when Na approaches averaged structural properties depend weakly on struc-
infinity while OD-κminL approaching the total κmin
L . In- tural complexity for compounds with similar chemistries.
terestingly, such a disorder-induced interconversion be- We apply these newly developed insights to bridge the
tween PGM-κmin L and OD-κmin L , when approached from knowledge gap between our κmin model and the CWP
L
an inverse perspective, might be leveraged to estimate model. This is motivated by the missing connection be-
the total κmin
L without explicitly computing the OD con- tween the CWP model and Allen and Feldman’s the-
tributions. That is, we may approximate the total κmin L ory of diffusons, a question initially raised by Cahill and
of some very complex materials (large Na ) wherein OD- Pohl [12]. Considering the PGM nature of the CWP
κmin
L dominates by computing only the PGM-κmin L of a model, we first compare the PGM-κmin calculated using
L
simplified model (e.g. Na = 1). The latter could be our model to those from the CWP model (denoted as
obtained by averaging the atomic masses, positions, and CWP-κmin L ). To eliminate the uncertainty caused by an
IFCs of the complex structure. κmin based on such an ap- anisotropic average of κmin
L L , we only selected compounds
proximation can be easily calculated using macroscopic with cubic symmetry, as shown in Fig.3A. In contrast to
properties such as materials’ density and elastic proper- the expected failure of the CWP model due to the lack of
ties. OD contribution, we find that CWP-κmin deviates signif-
L
We have so far qualitatively answered the two ques- icantly from PGM-κmin L and displays much larger values,
tions raised earlier by means of showing the impacts which seems to overcome the underestimation inherent
of weak disorder and structural complexity on κmin L in in PGM-κmin L . The latter is further confirmed by plot-
an idealized model. However, the assumption of weak ting CWP-κmin L against PGM+OD-κmin L in Fig.3B. Over-
disorder makes the above picture applicable to only a all, we find that CWP-κmin L is comparable to PGM+OD-
small group of compounds which share similar structures, κmin
L , although the latter tends to show smaller values
chemical compositions, and interatomic interactions. To statistically.
better describe complex materials with large structural The above comparison reveals that the CWP model
and compositional fluctuations, we move on to investi- seems to work remarkably well, which is quite surprising
gate models with increasingly strong disorder. We show because (i) the CWP model is relatively simple and only
in Fig.2B the calculated total and decomposed κmin L as needs sound velocity and number density as inputs [see
a function of the strength of random mass disorder (our Eq(3) in Ref.[9]], and (ii) the CWP model does not at all
tests show that perturbing IFCs has similar effects). We account for the OD contribution, which is dominated by
see from Fig.2B that PGM-κmin L decreases sharply with optical phonon modes [28]. We attribute the success of
the onset of the disorder and then continues to decrease the CWP model to the fact that the CWP model could
in a slower manner with increasing disorder. In contrast, be a good approximation to the sophisticated PGM+OD-
OD-κminL increases first and then decreases with enhanced κmin model based on Eq.(1) by averaging out disorder
L
disorder, while its percent contribution keeps increasing and phonon unfolding, in the same spirit of our model de-
(see Supplementary Materials). The resulting total κmin L tailed earlier. This is achieved by neglecting the details
displays an overall reduced value. By comparing the left of both chemical composition and atomic arrangement
and right panels in Fig.2B, we also observe that a larger using a single parameter, i.e., the number density n, to
unit cell with similar strength of disorder tends to shift describe the structure in the CWP model. In such a sim-
the maximum of OD-κmin L towards the weaker disorder plified picture with only one atom in an averaged unit
regime and make OD-κmin L more important. cell, only three acoustic branches arise, which induces
Overall, we see that, in the full range of disorder from an effective back conversion from the OD contribution
weak to strong, the total κmin
L of complex unit cells might to the PGM contribution, thus making the PGM domi-
be approximated by computing only PGM-κmin L for the nant again. This picture is supported by Fig.1C, which
perfectly ordered crystal, despite the latter displaying shows that PGM-κmin L tends to dominate over OD-κmin L
certain underestimation. Surprisingly, the presence of in simple crystals. Importantly, the above analysis im-
strong disorder might accidentally lead to a better agree- plies that it is inappropriate to use the CWP model as a
 6

A 2. B 2.
C
2.
1.

PGM+OD-kLMin @WêHm◊KLD

PGM+OD-kLMin @WêHm◊KLD
2x
PGM-kLMin @WêHm◊KLD

0.5 y= x 1. 1.
y= x

2x

2x
0.5

y=
y=

y=
0.2
0.5 0.5

x
y=

y=
0.1

5x

5x
0.

0.
y=

y=
0.2 0.2

0.1 0.1
0.1 0.2 0.5 1. 2. 0.1 0.2 0.5 1. 2. 0.1 0.2 0.5 1. 2.
CWP-kLMin @WêHm◊KLD CWP-kLMin @WêHm◊KLD 2ê3
0.5kB n HYêrL 1ê2

FIG. 3. Comparisons of the minimum lattice thermal conductivity based on Eq.[1] with the CWP model for cubic crystals at
600 K. (A) Comparisons of the phonon gas model of κminL (PGM-κmin min
L ) with the CWP model (CWP-κL ). (B)Comparisons of
the unified minimum thermal transport model that combines both phonon gas model and the off-diagonal terms (PGM+OD-
κmin
L ) with the CWP model. (C) The linear correlation of the PGM+OD-κL
min
with an approximated formula of the CWP
2/3 1/2
model in the high temperature limit, i.e., 0.5kB n (Y /ρ) , wherein n, Y , and ρ are the number density, Young’s modulus,
and mass density, respectively. The dashed lines in (A), (B), and (C) denote the deviations from the diagonal within a factor
of two.

second heat transport channel on top of the PGM model et al. [10], who propose to mitigate such an overestima-
as adopted in recent studies [18, 40], which will result in tion by resorting to an alternative heat transfer through
a double-counting of the PGM contribution. diffusons. We note that our conclusion only applies to the
The interpretation of the CWP model based on the general behavior of the CWP model but may not work
structure averaging and phonon unfolding picture offers for a specific compound, as indicated by the large spread
additional insights into its potential limitations or uncer- of data points along the diagonal in Fig.3C. Furthermore,
tainties. On one hand, neglecting structural details might we see from the relation of κmin
L ∝ kB n2/3 (Y /ρ)1/2 that
min
lead to a lack of additional flattening of phonon disper- the bound to κL of crystals can be attributed to the
sions caused by disorder or symmetry breaking, thus giv- competing parameters of n, Y , and ρ, all of which might
ing rise to an overestimated κmin have physical bounds for crystals.
L . On the other hand,
OD contributions from the three acoustic branches are
still missing, which leads to a general underestimation of With our κmin
L model established, we wanted to in-
κmin vestigate the distribution of κmin across all experimen-
L . Considering these two competing factors, there is
L
no clear answer on the net effect. However, through our tally known compounds in the Inorganic Crystal Struc-
numerical experiment presented in Fig.3B, we see that ture Database (ICSD) [45]. Due to the high computa-
CWP-κmin is likely to overestimate κmin tional cost of calculating phonon properties from first-
L L , thus indicating
that the dominant uncertainty probably comes from the principles, it would be unfeasible to calculate them all,
first factor. This might explain why experiments tend to making it necessary to use a more efficient model. We
find lower lattice thermal conductivities than the values decided on two complementary approaches to machine-
predicted by the CWP model [10]. learn κmin
L : MatErials Graph Network (MEGNet) [46] –
a state of the art crystal graph convolutional neural net-
To establish a quantitative measure of the extent work – and a random forest – a more interpretable ap-
that the CWP model overestimates κmin L compared to proach (as seen in Fig.4A). MEGNet directly uses crystal
our model, in Fig.3C, we numerically fit our calcu- structure information to construct crystal graphs as fea-
lated PGM+OD-κmin L based on the relation of κmin L ∝ tures. In contrast, the random forest requires featuriza-
2/3 1/2
kB n (Y /ρ) suggested by Clarke [7], wherein n is tion. For the random forest, we used a variety of features
the number density, Y is Young’s modulus, and ρ is generated by Matminer [47] using the Magpie [48] preset
the mass density. We find that PGM+OD-κmin L ≈ including the mean, standard deviation, minimum, max-
0.5kB n2/3 (Y /ρ)1/2 , which is considerably smaller than imum, mode, and range of elemental properties such as
that from the CWP model in the high temperature atomic number (N ), mass (m), electronegativity (EN ),
limit [9], i.e., κmin
L ≈ 1.1kB n2/3 (Y /ρ)1/2 if sound velocity Mendeleev number (Nm ), melting temperature (Tm ), col-
is approximated as vs = (2vT + vL )/3 ≈ 0.94(Y /ρ)1/2 [7, umn, row, covalent radius (Rcovalent ), number of elec-
10]. Our analysis unambiguously demonstrates that the trons and unfilled slots in the s, p, d, and f valence or-
CWP model tends to overestimate κmin L intrinsically, thus bitals, bandgap (EgGS ), magnetic moment (MGS ), and
offering a different perspective from the proposal by Agne volume per atom in the ground state (VGS ). We also in-
 7

A B D E

F H

FIG. 4. Machine learning models of minimum lattice thermal conductivity. (A) Depiction of the two machine-learning models,
and the error on the test set for (B) MEGNet and (C) the random forest. (D) Feature importance from the random forest. (E)
MEGNet latent graph features reduced to two dimension through t-SNE. Black points show the training set and other points
show the ICSD dataset, colored by their estimated κmin
L . (F) Kernel density estimate of the MAE distribution for κL
min
for both
min min
models on the test set. (G) Kernel density estimate of κL for both models on the ICSD dataset. (H) κL of seven selected
compounds from experiments (TlInTe2 at 600 K [41], Tl3 VSe4 at 300 K [18], CsAg5 Te3 at 600 K [25], Cu2 Se at 600 K [23],
AgTlTe at 600 K [22], BiCuSeO at 600 K [42], Cu3 SbSe3 at 600 K [43], and Sr8 Ga16 Ge30 at 300 K [44]), direct calculation using
our PGM+OD-κmin L model, MEGNet, and the random forest at the corresponding temperatures. All κmin L were isotropically
averaged.

cluded some simple structural features such as the space having strong bonding and relatively low densities. Part
group, and the mean and standard deviations of both of the difference in predicted values between the mod-
bond length (LB ) and angle (θB ), totaling 141 features els arises from a limitation of the random forest, as it
for each compound. We refer the readers to the Meth- is restricted to interpolating between closest data points
ods section for more details on training machine learning in our training set, preventing it from predicting smaller
models. or larger values than those it was trained on. On the
The two models achieve similar accuracies on our test other hand, MEGNet is capable of extrapolation and
set, with MEGNet slightly outperforming the random finds more compounds with higher values of κmin L when
forest (Fig.4B,C, and F), and both also predict similar compared to the random forest and finds a smoother dis-
distributions for κmin on the ICSD data set (Fig.4G). The tribution for intermediate values, as shown in Fig.4G.
L
distribution is bounded on the low end near 0.1 W/[m·K] The random forest model provides advantages in the
and a maximum around 4 W/[m·K]. Only ∼100 of the form of faster computation, interpretable feature impor-
∼36,000 structures show MEGNet-predicted κmin L values tance, and easily visualized features. From the random
above 2.64 W/[m·K] (the maximum observed in the cal- forest, we extracted the 15 most important features and
culated training set) and are almost entirely various al- have plotted them in Fig.4D. Our most information-
lotropes of Be, C, and Co or carbides and nitrides, all dense features are related to bonding strength (LB , θB ,
 8

EN , Nvalence ) and factors concerning the atomic proper- glass material. This finding implies that although κL
ties (m, N , ρ, Rcovalent , VGS ). In some cases, the mean of a material might exhibit a temperature dependence
properties are more useful, such as for bond length (LB ), similar to that of glasses, κL may still not achieve the
bond angle (θB ), and valence electron counts (Nvalence ), minimum.
while in other cases, the differences between atomic prop- Overall, the machine-learned κminL from both MEGNet
erties carry more information (σm , σN , σRow , σEN ). This and the random forest agree very well with our direct
is not unexpected, as phonon transport is sensitive to calculations. This is rather encouraging because most of
both similarities and differences within compounds. For the compounds are not contained in the training dataset
example, we tend to expect shorter bonds to be stronger for the machine learning models except for TlInTe2 . The
which results in higher characteristic frequencies, while largest deviation between experiment, machine learning
mass difference among atoms could significantly alter the models, and direct calculation is found in Cu2 Se. The
phonon spectrum. Notably, the most information-dense discrepancy might be attributed to the following factors:
features being related to bonding strength is consistent (i) we used an ordered structure (β-Cu2 Se) to compute
with the positive correlation with Y in the relation of κmin
L , while in the experimental structure the copper ions
κmin
L ∝ kB n2/3 (Y /ρ)1/2 . are highly disordered around the Se sublattice [23]; (ii)
Finally, we performed dimensionality reduction using β-Cu2 Se often exhibits off-stoichiometry with copper va-
t-SNE (t-Distributed Stochastic Neighbor Embedding) cancies, resulting in relatively large variations in the ex-
with PCA (Principal Component Analysis) initialization perimentally measured values of κL in the range of 0.4-
in order to project MEGNet’s latent graph features into 0.6 W/[m·K] among different samples [23]. Therefore,
two dimensions for visualization, as seen in Fig.4E. The despite the disordered nature of β-Cu2 Se, our models,
distribution of the training set and the ICSD dataset including both machine learning and direct calculation,
are considered together, and the overlap generally tells seem to provide reasonable estimations using ordered
us how well the training set covers the ICSD dataset, structures.
indicating potential transferability of the model. The The above results on Cu2 Se indicate that κL of amor-
points containing ICSD materials with low and high pre- phous compounds might be approximated with a rea-
dicted κmin
L are segregated in the latent space, allowing sonable accuracy using ordered crystals. To confirm this
the model to clearly separate them from each other. Ad- hypothesis, we further apply our model to amorphous sil-
ditionally, the training set points span the extent of the ica (a-SiO2 ). We calculated PGM+OD-κmin L for two or-
axes and their distribution varies smoothly, with the ex- dered phases of SiO2 (α-quartz and α-cristobalite) that
ception of less density near some outer edges and a small are free of lattice instabilities, which are compared to
cluster towards the top center of the figure. Neverthe- the CWP model and the experimental measurement [9]
less, the coverage is especially dense in the area of low in Fig.5A. We see that PGM+OD-κmin L of both α-quartz
predicted κmin
L which is most relevant for thermoelectric and α-cristobalite agree well with experiment, displaying
or thermal barrier coating materials, suggesting that our overall improvement over the CWP model, particularly
training set sufficiently covers the space, but coverage at high temperatures. This implies that specific atomic
could be improved by increasing the size of our training arrangement might not be very relevant when approach-
set in the future. ing κmin
L , although it is responsible for subtle differences,
We further validate these predictions against experi- which are probably more profound at low temperatures.
mental measurements and direct calculations using Eq.[1] To provide more evidence for this observation, we next
Since κmin
L is relatively rare in ordered crystals, we de- examine other phases of SiO2 using our machine learn-
liberately choose thermoelectric materials exhibiting ul- ing model. Particularly, we used MEGNet instead of the
tralow κL , which presumably best represent κmin L , as random forest because the former better encodes struc-
shown in Fig.4H. We find that the experimental val- tural information. Since our MEGNet model is trained
ues are either higher or comparable to PGM+OD-κmin L , on structures from the Materials Project (MP) [33–35],
consistent with the definition of κminL . It is noteworthy we collected all structures with the chemical formula
that compounds including CsAg5 Te3 [25], Cu3 SbSe3 [43], SiO2 available within MP, totaling 317 structures. The
and BiCuSeO [42] have κL already close to κmin L from calculated κmin
L using MEGNet is shown in Fig.5B. We
our model, suggesting that further improvement in ther- find a relatively large spread of κminL when Na is small
moelectric efficiency should be focused on optimizing (Na < 25), while the spread becomes much smaller when
their electronic transport properties. We note that Na is large (Na > 75). The mean value of κmin L of the
CsAg5 Te3 and Cu3 SbSe3 indeed resemble amorphous 317 structures is 1.45 W/[m·K], close to the experimental
solids and display nearly temperature-independent and value of 1.65 W/[m·K] for the amorphous phase [9] and
glasslike κL [25, 43]. Whereas for compounds such as better than the value of 1.18 W/[m·K] from the CWP
Tl3 VSe4 [18] and TlInTe2 [41], there is still room for fur- model at 600 K [9]. These results confirm our earlier
ther reducing κL , although their κL are already very low. hypothesis that κL of amorphous compounds can be ap-
Also, we find it interesting to see a large gap between our proximated by ordered crystals. Moreover, a reliable pre-
calculated κmin
L and the experimentally measured κL in diction may be achieved by averaging over different struc-
Sr8 Ga16 Ge30 [44, 49], a typical electron-crystal phonon- tures, effectively capturing the distinct local structures in
 9

A 2. B 2. C 1.5
1.5
a-cristobalite
Exp. Ha-SiO 2 L
Exp.

MEGNet-kLMin @WêHm◊KLD
1.8 MEGNet
Exp.
CWP
kLMin @WêHm◊KLD

1.5 1.

kL @WêHm◊KLD
1.6
a-quartz

kL HWêm◊KL
1.4
1. CWP Mean 0.5

1.2
CWP
0.5 1. 0.
200 400 600 800 50 100 150 200
Temperature HKL
S3 s 2 Ge Se Si
SiO
2
 of atomic sites As 2 G eA
Cd

FIG. 5. Prediction of lattice thermal conductivity in amorphous solids using machine learning models of minimimum lattice
thermal conductivity. (A) Comparison of temperature-dependent κL of amorphous silica (a-SiO2 ) between the experiment
(gray disks) [9] and theoretical κmin
L (solid/dashed lines). Results from the CWP model are indicated by solid magenta lines.
The dashed lines display the calculated PGM+OD-κmin L for two ordered phases of SiO2 , namely α-quartz and α-cristobalite.
(B) Predicted κmin
L (orange disks) at 600 K using MEGNet for 317 structures of ordered SiO2 from Materials Project [33] as a
function of the number of atomic sites in a primitive cell. The upper, middle, and lower dashed lines denote the values from the
experiment, mean of MEGNet predictions, and the CWP model, respectively. (C) Comparison of κL at 300 K of six amorphous
compounds between experiments [9], our MEGNet model, and the CWP model [9].

amorphous solids. W/[m·K] in at least one of the three Cartesian direc-

We apply our MEGNet model to other amorphous tions. However, there are only 5 compounds exhibiting
solids by averaging over ordered phases and compared κmin
L less than 0.07 W/[m·K] (see listed compounds and
to the experiments and the CWP model in Fig.5C. Our κmin
L in Supplementary Materials). This observation sug-
machine learning model achieves an accuracy compara- gests that, even though it may be difficult to find such
ble with the CWP model. The advantage of our model materials due to their rarity, anisotropic materials, in-
is the calculation requires very low computational cost cluding layered crystals, have the potential to achieve a
(within a second for each compound) and simple input lattice thermal conductivity of less than 0.1 W/[m·K],
(only structural information), and thus is readily applica- consistent with the prior work [51].
ble to many other compounds on a large scale. We note Secondly, we computed all κmin L using the harmonic
that compared to the amorphous SiO2 our predicted κL phonon dispersion in the absence of finite temperature
of amorphous Si is considerably smaller than both the induced phonon frequency shifts; the latter are important
CWP model and the experiment. We attribute such a in a range of low-thermal-conductivity materials [52, 53].
discrepancy mainly to the observations reported in the As a first attempt to investigate the effects of phonon
literature that (i) propagating modes contribute signifi- anharmonicity on κmin L , we computed κL
min
at 300 K for
cantly more in amorphous Si than amorphous SiO2 , and Tl3 VSe4 using harmonic and anharmonic phonon disper-
(ii) phonon lifetimes in amorphous Si are notably larger sions, respectively. The anharmonic phonon dispersion
than π/ω [50], both of which indicate that κL in amor- was computed using the self-consistent phonon theory at
phous Si still does not reach the minimum value. We 300 K in a previous study [52]. We find that account-
hypothesize that this might also be the case for amor- ing for anharmonic effects only leads to a small change
phous CdGeAs2 and Ge. (about 7% decrease) in κmin L , which is much smaller
Before closing, we briefly comment on the limita- than the change (about 162.5% increase) in κL when
tions of our theoretical and machine learning models. phonon lifetimes are limited by intrinsic three phonon
Firstly, both the identified lower bound to κminL and the scattering [52]. The relative insensitivity of κmin L to an-
constructed machine learning models are based on the harmonic effects can be explained by the relationship
isotropically averaged κmin
L . Consequently, it is reason-
κmin
L ∝ kB n2/3 (Y /ρ)1/2 ∝ kB n2/3 vs [7], which indicates
able to question how κminL behaves in anisotropic struc- that κmin
L is largely unchanged unless anharmonic effects
tures. This anisotropy is of particular interest because lead to a sizable change in sound velocity vs .
κL as small as 0.05 W/[m·K] below our isotropically aver- Thirdly, it is not entirely justified to use one half of the
aged lower bound at room temperature has been realized vibrational period as the minimum phonon lifetime, i.e.,
in layered WSe2 crystals [51]. Therefore, we analyzed τ = π/ω. The adoption of τ = π/ω was first proposed
the anisotropic κmin
L for all of the calculated compounds by Cahill, Watson, and Pohl [9], motivated by Einstein’s
(see Fig.S3 in Supplementary Materials). We find there thermal conductivity model [13]. Specifically, Einstein’s
are 67 compounds in total that have κmin L less than 0.1 model relies on a critical assumption that there is no co-
 10

herence between the motions of neighboring atoms, which III. CONCLUSIONS

seems to be a reasonable choice for glassy materials. Ein-
stein further derived the energy exchange between neigh- In summary, we have developed a first-principles mini-
boring atoms by integrating over one half of a period of mum lattice thermal conductivity model based on a uni-
oscillation, however, without an explicit explanation in fied theory of thermal transport in crystals and glasses.
his original paper [13]. We hypothesize this might be mo- By applying such a model to thousands of crystalline
tivated by the fact that it is the shortest time period that compounds, we have discovered a universal bound to
an atom returns to its original position after exchanging minimum lattice thermal conductivity independent of
energy with neighboring atoms. Cahill et al. interpreted structural complexity. In striking contrast to the con-
Einstein’s result to mean that each atom undergoes a ventional phonon gas model, such an unusual behavior is
thermal energy loss or gain in the span of half an oscilla- found to be deeply rooted in the conversion of the heat
tion period [9] and subsequently assumed that τ = π/ω. transfer mechanism from the phonon gas model to the
Mathematically, τ adopted in computing κmin L can have diffuson picture with the presence of disorder, while the
an arbitrary value, which presumably is small in order value of the total minimum lattice thermal conductiv-
to reduce the PGM contribution. However, physically, τ ity largely remains unchanged. With these insights, we
cannot be arbitrarily short because our derived equation bridge the knowledge gap between the Cahill-Watson-
for κmin
L relies on a critical assumption that phonons are Pohl model and our unified model by pointing out that
well-behaved quasiparticles (τ ω ≈ 2π). The very small τ the former could be viewed as an approximation of the
implies strong interactions and will invalidate the quasi- latter. We further construct machine learning models
particle picture, making the phonon properties (such as based on a graph network and random forest to enable
group velocities and frequencies) less well-defined and in- fast and accurate prediction of minimum lattice thermal
validating the usage of Eq.[1]. To explore the potential conductivity on a large scale, which is validated against
uncertainty due to the adopted τ , we performed sensitiv- thermoelectric materials with ultralow and glasslike lat-
ity analysis of κmin
L by varying τ for a randomly selected tice thermal conductivity. We demonstrate the appli-
set of compounds (see Supplementary Materials). We cability of our graph network model to simulate lattice
find that κmin
L decreases for most compounds when τ is thermal conductivity in amorphous solids. These find-
decreased. Specifically, a sharp decrease in κminL is found ings highlight a unified understanding of the lower limit
when τ is reduced from a larger value to π/ω, a value of lattice thermal transport in solids. The first-principles-
we adopted for computing κmin L . When τ is less than based theory and machine learning model built in this
π/ω, slower decrease or no change in κmin L is observed, work are universal and readily applicable in research rel-
especially for lower values of κmin
L . Since τ cannot be ar- evant to thermal energy conversion and management.
bitrarily small as argued previously, we hypothesize that
τ = π/ω might be a reasonable choice to achieve a rel-
atively low value of κmin
L while approaching the limit of ACKNOWLEDGMENTS
the quasiparticle picture.
Y.X., D.G., K.P., and C.W. acknowledge financial sup-
port received from (i) Toyota Research Institute (TRI)
through the Accelerated Materials Design and Discov-
ery program (thermal conductivity calculations), (ii) the
Department of Energy, Office of Science, Basic Energy
Finally, despite encouraging results from the com-
Sciences under grant DE-SC0014520 (theory of anhar-
parison between the calculations and the experiments,
monic phonons), and (iii) the U.S. Department of Com-
our theoretical model only accounts for harmonic heat
merce and National Institute of Standards and Tech-
flux [39, 54] and Eq.[1] fails to describe phonon satellite
nology as part of the Center for Hierarchical Materi-
structures deviating from the quasiparticle picture [55].
als Design (CHiMaD) under award no. 70NANB14H012
For the machine learning model, local structural mo-
(DFT calculations). Y. X. was also supported by Port-
tifs may not be well captured by ordered structures for
land State University Lab Setup Fund. M.G.K. were
amorphous solids, and our simple averaging scheme with
supported in part by the National Science Foundation
equal weight for each structure is probably not optimal.
Grant DMR-2003476. V.O. acknowledges financial sup-
We deem establishing a more reliable physical bound to
port from the National Science Foundation Grant DMR-
phonon lifetime, including higher-order anharmonic ef-
1611507. We acknowledge the computing resources
fects on the phonon frequency shifts and the heat flux
provided by (i) the National Energy Research Scien-
beyond the phonon quasiparticle picture, and machine
tific Computing Center (NERSC), a U.S. Department
learning structural motifs inherent in amorphous solids
of Energy Office of Science User Facility operated un-
are interesting avenues of further research to improve our
der Contract No. DEAC02-05CH11231, (ii) Quest high-
model.
performance computing facility at Northwestern Uni-
versity which is jointly supported by the Office of the
Provost, the Office for Research, and Northwestern Uni-
 11

versity Information Technology, and (iii) Bridges2 at and efficiency.

Pittsburgh Supercomputing Center (PSC) through al-
locations dmr160027p and mat220007p from the Ad-
vanced Cyber-infrastructure Coordination Ecosystem: Appendix B: Machine learning models
Services & Support (ACCESS) program, which is sup-
ported by National Science Foundation grants #2138259,
MEGNet: For the MEGNet model, we used frozen ini-
#2138286, #2138307, #2137603, and #2138296. The
tial atom embeddings transferred from another MEGNet
authors (Y.X., C.W, and M.K, initial DFT and concep-
model [46] trained on 133,420 formation energies from
tion of research problem) also acknowledge support for
Materials Project [33]. Crystal graphs were constructed
the initial stages of this research from the U.S. Depart-
using 50 features per bond with a gaussian smearing
ment of Energy under Contract No. DE-SC0014520.
width of 0.5 Å up to a cutoff radius of 5 Å for the bond
attributes, transferred atom embedding weights of size 16
for the atom attributes, and one single global attribute
Appendix A: First-principles calculation of κmin
L embedding the temperature in Kelvin. For the model
hyperparameters, we used only a single MEGNet block
The key ingredients for modeling κmin
L based on Eq.[1] and set the number of dense layers in the MEGNet block
from first principles are materials harmonic vibrational to n1 = 16, n2 = 16, and n3 = 8, followed by 2 Set2Set
spectra, which can be explicitly calculated if materi- passes, and the learning rate was set to 10−3 . In order
als structure and harmonic interatomic force constants to prevent overfitting, we set weight dropout to 25% dur-
are known. To construct a large database for κmin L , ing both training and prediction and added a small L2
we used the phonon database (phononDB) generated by weight regularization parameter of 10−5 . The κmin L data
Togo [31], who used crystalline structures from the Ma- was split 80%:10%:10% into a training set, validation set,
terials Project (MP) [33–35] and calculated harmonic in- and testing set. Predictions were then run over 10 tri-
teractions using Phonopy [32]. It is worth noting that als and the average was taken as the final output. The
these phonon calculations were performed by means of model was run for 500 epochs with a patience of 250,
the Vienna Ab Initio Simulation Package (VASP) [56– and the model with the best validation performance was
59], which employed the projector-augmented wave selected for further predictions. The MEGNet training
(PAW) [60] method in conjunction with the revised mean absolute error (MAE) was 0.033 W/[m·K], the val-
Perdew-Burke-Ernzerhof version (PBEsol [61]) of the idation MAE was 0.0406 W/[m·K], and the testing MAE
generalized gradient approximation (GGA) [62] for the was 0.0479 W/[m·K]. Finally, we performed dimension-
exchange-correlation functional [63]. We performed post ality reduction on the combined phononDB and ICSD
processing of the phononDB (version of 2018-04-17) to dataset using t-SNE to project the latent graph features
generate harmonic force constants and downselected the into 2 dimensions. We intialized the t-SNE with PCA
compounds that are free of lattice instabilities (imagi- and chose the settings of 1,000 iterations, a perplexity of
nary phonon frequencies) with supercell structures con- 5, and a learning rate of 39,514 (n).
structed from diagonal matrices (required by our κmin
L im- Random Forest: Features for the random forest were
plementation within ShengBTE [64]), totaling 2576 com- created through Matminer [47] using the Magpie pre-
pounds. These compounds cover wide ranges of chemical set [48], along with global symmetry features and site
compositions and space group symmetries, with 189, 922, statistic fingerprints, totaling 141 features for each com-
310, 660, and 495 compounds from cubic, orthorhom- pound as detailed in the main text. The random for-
bic, tetragonal, trigonal/hexagonal, triclinic/monoclinic est was constructed using 1,000 estimators, and we veri-
crystal systems, respectively. Other compounds and fied that including more estimators did not improve the
their harmonic phonon properties, including TlInTe2 , accuracy significantly. For this model, we used 5-fold
Tl3 VSe4 , CsAg5 Te3 , Cu2 Se, TlAgTe, CuBiSeO, SnSe, cross validation on the 90% training set for parameter
α-cristobalite, and α-quartz, were calculated following selection and left a holdout test set of 10%. In order
the similar DFT settings. For β-Cu2 Se which has imagi- to prevent overfitting, we changed the allowable maxi-
nary phonon frequencies, we have performed anharmonic mum depth of each tree, where larger maximum depths
phonon renormalization at finite temperature (600 K) us- showed better prediction accuracies at the cost of a larger
ing self-consistent phonon theory [52, 53, 65, 66]. gap between training and testing accuracies. A maxi-
We implemented the κmin
L model based on Eq.[1] within mum depth 8 was then chosen to balance training, vali-
ShengBTE [64]. The off-diagonal group velocities were dation, and testing accuracy. For the random forest, the
calculated following the derivations by Allen and Feld- mean training MAE was 0.0375 W/[m·K], mean valida-
man [16] and Simoncelli et al. [28]. We calculated tion MAE was 0.0458 W/[m·K], and the testing MAE
κmin
L for these 2576 compounds at three temperatures of was 0.0570 W/[m·K]. Feature importance was extracted
300 K, 600 K, and 900 K using a constant mesh density from a random forest trained on only 600 K data.
(mesh density = 50) following Phonopy conventions [32]. The training dataset used in both MEGNet and ran-
The convergence is carefully monitored across different dom forest models consists of our calculated κmin L at
compounds to achieve good balance between accuracy three temperatures of 300K, 600K, and 900K, totaling
 12

7728 datapoints. For the ICSD dataset, 36,199 crystal Appendix C: Data availability
structures were pulled from the Open Quantum Materi-
als Database (OQMD) [67, 68], and all machine learning The codes, data sets, and machine learning models are
predictions of κmin
L were then performed at 600 K unless available via public repository (https://github.com/
otherwise specified. yimavxia/Minikappa)

[1] Lon E. Bell, Cooling, Heating, Generating Power, and [18] Saikat Mukhopadhyay, David S. Parker, Brian C. Sales,
Recovering Waste Heat with Thermoelectric Systems, Alexander A. Puretzky, Michael A. McGuire, and Lucas
Science 321, 1457–1461 (2008). Lindsay, Two-channel model for ultralow thermal con-
[2] Glen A Slack, The thermal conductivity of nonmetallic ductivity of crystalline Tl3VSe4, Science 360, 1455–1458
crystals, Solid state physics 34, 1–71 (1979). (2018).
[3] G. A. Slack, CRC Handbook of Thermoelectrics (CRC [19] Yixiu Luo, Xiaolong Yang, Tianli Feng, Jingyang Wang,
Press, Boca Raton, FL, 1995) pp. 407–440. and Xiulin Ruan, Vibrational hierarchy leads to dual-
[4] G. Jeffrey Snyder and Eric S. Toberer, Complex thermo- phonon transport in low thermal conductivity crystals,
electric materials, Nat. Mater. 7, 105–114 (2008). Nature Communications 11, 2554 (2020).
[5] Joseph R Sootsman, Duck Young Chung, and Mer- [20] Yi Xia, Vidvuds Ozoliņš, and Chris Wolverton, Micro-
couri G Kanatzidis, New and old concepts in thermoelec- scopic Mechanisms of Glasslike Lattice Thermal Trans-
tric materials, Angewandte Chemie International Edition port in Cubic Cu12 Sb4 S13 Tetrahedrites, Phys. Rev. Lett.
48, 8616–8639 (2009). 125, 085901 (2020).
[6] Jian He and Terry M. Tritt, Advances in thermoelectric [21] Riley Hanus, Janine George, Max Wood, Alexander
materials research: Looking back and moving forward, Bonkowski, Yongqiang Cheng, Douglas L. Abernathy,
Science 357 (2017), 10.1126/science.aak9997. Michael E. Manley, Geoffroy Hautier, G. Jeffrey Snyder,
[7] David R. Clarke, Materials selection guidelines for low and Raphaël P. Hermann, Uncovering design principles
thermal conductivity thermal barrier coatings, Surface for amorphous-like heat conduction using two-channel
and Coatings Technology 163-164, 67–74 (2003), pro- lattice dynamics, Materials Today Physics 18, 100344
ceedings of the 29th International conference on Metal- (2021).
lurgical Coatings and Thin Films. [22] Ken Kurosaki, Atsuko Kosuga, Keita Goto, Hiroaki
[8] Jayant Gopal Thakare, Chandan Pandey, MM Mahapa- Muta, and Shinsuke Yamanaka, Ag9 TlTe5 and AgTlTe:
tra, and Rahul S Mulik, Thermal barrier coatings—A high ZT materials with extremely low thermal conductiv-
state of the art review, Metals and Materials Interna- ity, in ICT 2005. 24th International Conference on Ther-
tional 27, 1947–1968 (2021). moelectrics, 2005. (IEEE, 2005) pp. 323–326.
[9] David G. Cahill, S. K. Watson, and R. O. Pohl, Lower [23] Huili Liu, Xun Shi, Fangfang Xu, Linlin Zhang, Wen-
limit to the thermal conductivity of disordered crystals, qing Zhang, Lidong Chen, Qiang Li, Ctirad Uher, Tris-
Phys. Rev. B 46, 6131–6140 (1992). tan Day, and G. Jeffrey Snyder, Copper ion liquid-like
[10] Matthias T. Agne, Riley Hanus, and G. Jeffrey Snyder, thermoelectrics, Nature Materials 11, 422–425 (2012).
Minimum thermal conductivity in the context of diffuson- [24] Xu Lu, Donald T. Morelli, Yi Xia, Fei Zhou, Vidvuds
mediated thermal transport, Energy Environ. Sci. 11, Ozolins, Hang Chi, Xiaoyuan Zhou, and Ctirad Uher,
609–616 (2018). High Performance Thermoelectricity in Earth-Abundant
[11] D G Cahill and R O Pohl, Lattice Vibrations and Heat Compounds Based on Natural Mineral Tetrahedrites,
Transport in Crystals and Glasses, Annual Review of Advanced Energy Materials 3, 342–348 (2013).
Physical Chemistry 39, 93–121 (1988). [25] Hua Lin, Gangjian Tan, Jin-Ni Shen, Shiqiang Hao, Li-
[12] David G. Cahill and R.O. Pohl, Heat flow and lattice vi- Ming Wu, Nicholas Calta, Christos Malliakas, Si Wang,
brations in glasses, Solid State Communications 70, 927 Ctirad Uher, Christopher Wolverton, and Mercouri G.
– 930 (1989). Kanatzidis, Concerted Rattling in CsAg5Te3 Leading to
[13] A. Einstein, Elementare Betrachtungen über die thermis- Ultralow Thermal Conductivity and High Thermoelectric
che Molekularbewegung in festen Körpern, Annalen der Performance, Angewandte Chemie International Edition
Physik 340, 679–694 (1911). 55, 11431–11436 (2016).
[14] C. Kittel, Introduction to Solid State Physics (Wiley, [26] Siqi Lin, Wen Li, Shasha Li, Xinyue Zhang, Zhiwei Chen,
2004). Yidong Xu, Yue Chen, and Yanzhong Pei, High Ther-
[15] Philip B. Allen and Joseph L. Feldman, Thermal Conduc- moelectric Performance of Ag9GaSe6 Enabled by Low
tivity of Glasses: Theory and Application to Amorphous Cutoff Frequency of Acoustic Phonons, Joule 1, 816–830
Si, Phys. Rev. Lett. 62, 645–648 (1989). (2017).
[16] Philip B. Allen and Joseph L. Feldman, Thermal con- [27] Glen A Slack, Nonmetallic crystals with high thermal
ductivity of disordered harmonic solids, Phys. Rev. B 48, conductivity, Journal of Physics and Chemistry of Solids
12581–12588 (1993). 34, 321–335 (1973).
[17] Philip B Allen, Joseph L Feldman, Jaroslav Fabian, [28] Michele Simoncelli, Nicola Marzari, and Francesco
and Frederick Wooten, Diffusons, locons and propagons: Mauri, Unified theory of thermal transport in crystals
Character of atomic vibrations in amorphous Si, Philo- and glasses, Nature Physics 15, 809–813 (2019).
sophical Magazine B 79, 1715–1731 (1999). [29] Leyla Isaeva, Giuseppe Barbalinardo, Davide Donadio,
and Stefano Baroni, Modeling heat transport in crystals
 13

and glasses from a unified lattice-dynamical approach, [43] Kriti Tyagi, Bhasker Gahtori, Sivaiah Bathula, A. K. Sri-
Nature Communications 10, 3853 (2019). vastava, A. K. Shukla, Sushil Auluck, and Ajay Dhar,
[30] Michele Simoncelli, Nicola Marzari, and Francesco Thermoelectric properties of Cu3SbSe3 with intrinsically
Mauri, Wigner Formulation of Thermal Transport in ultralow lattice thermal conductivity, J. Mater. Chem. A
Solids, Phys. Rev. X 12, 041011 (2022). 2, 15829–15835 (2014).
[31] Atsushi Togo, Phonon database at Kyoto university, [44] G. S. Nolas, J. L. Cohn, G. A. Slack, and S. B. Schujman,
http://phonondb.mtl.kyoto-u.ac.jp/ (2018). Semiconducting Ge clathrates: Promising candidates for
[32] Atsushi Togo and Isao Tanaka, First principles phonon thermoelectric applications, Applied Physics Letters 73,
calculations in materials science, Scripta Materialia 108, 178–180 (1998), https://doi.org/10.1063/1.121747.
1 – 5 (2015). [45] Alec Belsky, Mariette Hellenbrandt, Vicky Lynn Karen,
[33] Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, and Peter Luksch, New developments in the Inorganic
Wei Chen, William Davidson Richards, Stephen Dacek, Crystal Structure Database (ICSD): accessibility in sup-
Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand port of materials research and design, Acta Crystal-
Ceder, and Kristin A. Persson, Commentary: The Ma- lographica Section B: Structural Science 58, 364–369
terials Project: A materials genome approach to accel- (2002).
erating materials innovation, APL Materials 1, 011002 [46] Chi Chen, Weike Ye, Yunxing Zuo, Chen Zheng, and
(2013), https://doi.org/10.1063/1.4812323. Shyue Ping Ong, Graph Networks as a Universal Ma-
[34] Shyue Ping Ong, William Davidson Richards, Anubhav chine Learning Framework for Molecules and Crys-
Jain, Geoffroy Hautier, Michael Kocher, Shreyas Cho- tals, Chemistry of Materials 31, 3564–3572 (2019),
lia, Dan Gunter, Vincent L. Chevrier, Kristin A. Pers- https://doi.org/10.1021/acs.chemmater.9b01294.
son, and Gerbrand Ceder, Python Materials Genomics [47] Logan Ward, Alexander Dunn, Alireza Faghaninia,
(pymatgen): A robust, open-source python library for Nils E.R. Zimmermann, Saurabh Bajaj, Qi Wang, Joseph
materials analysis, Computational Materials Science 68, Montoya, Jiming Chen, Kyle Bystrom, Maxwell Dylla,
314–319 (2013). Kyle Chard, Mark Asta, Kristin A. Persson, G. Jeffrey
[35] Shyue Ping Ong, Shreyas Cholia, Anubhav Jain, Miriam Snyder, Ian Foster, and Anubhav Jain, Matminer: An
Brafman, Dan Gunter, Gerbrand Ceder, and Kristin A. open source toolkit for materials data mining, Computa-
Persson, The Materials Application Programming Inter- tional Materials Science 152, 60–69 (2018).
face (API): A simple, flexible and efficient API for ma- [48] Logan Ward, Ankit Agrawal, Alok Choudhary, and
terials data based on REpresentational State Transfer Christopher Wolverton, A general-purpose machine
(REST) principles, Computational Materials Science 97, learning framework for predicting properties of inorganic
209–215 (2015). materials, npj Computational Materials 2, 16028 (2016).
[36] Glen A Slack, The thermal conductivity of nonmetallic [49] B.C. Chakoumakos, B.C. Sales, D.G. Mandrus, and G.S.
crystals, Solid state physics 34, 1–71 (1979). Nolas, Structural disorder and thermal conductivity of
[37] Eric S. Toberer, Alex Zevalkink, and G. Jeffrey Snyder, the semiconducting clathrate Sr8Ga16Ge30, Journal of
Phonon engineering through crystal chemistry, J. Mater. Alloys and Compounds 296, 80–86 (2000).
Chem. 21, 15843–15852 (2011). [50] Jason M. Larkin and Alan J. H. McGaughey, Thermal
[38] Ke Chen, Bai Song, Navaneetha K. Ravichandran, Qiye conductivity accumulation in amorphous silica and amor-
Zheng, Xi Chen, Hwijong Lee, Haoran Sun, Sheng phous silicon, Phys. Rev. B 89, 144303 (2014).
Li, Geethal Amila Gamage Udalamatta Gamage, Fei [51] Catalin Chiritescu, David G Cahill, Ngoc Nguyen, David
Tian, Zhiwei Ding, Qichen Song, Akash Rai, Hanlin Johnson, Arun Bodapati, Pawel Keblinski, and Paul
Wu, Pawan Koirala, Aaron J. Schmidt, Kenji Watan- Zschack, Ultralow thermal conductivity in disordered,
abe, Bing Lv, Zhifeng Ren, Li Shi, David G. Cahill, layered WSe2 crystals, Science 315, 351–353 (2007).
Takashi Taniguchi, David Broido, and Gang Chen, [52] Yi Xia, Koushik Pal, Jiangang He, Vidvuds Ozoliņš, and
Ultrahigh thermal conductivity in isotope-enriched Chris Wolverton, Particlelike Phonon Propagation Dom-
cubic boron nitride, Science 367, 555–559 (2020), inates Ultralow Lattice Thermal Conductivity in Crys-
https://www.science.org/doi/pdf/10.1126/science.aaz6149. talline Tl3 VSe4 , Phys. Rev. Lett. 124, 065901 (2020).
[39] Robert J. Hardy, Energy-Flux Operator for a Lattice, [53] Yi Xia, Vinay I. Hegde, Koushik Pal, Xia Hua, Dale
Phys. Rev. 132, 168–177 (1963). Gaines, Shane Patel, Jiangang He, Muratahan Aykol,
[40] Minghui Wu, Enamullah, and Li Huang, Unusual lattice and Chris Wolverton, High-Throughput Study of Lat-
thermal conductivity in the simple crystalline compounds tice Thermal Conductivity in Binary Rocksalt and
TlXTe2 (X=Ga,In), Phys. Rev. B 100, 075207 (2019). Zinc Blende Compounds Including Higher-Order Anhar-
[41] Manoj K. Jana, Koushik Pal, Avinash Warankar, Pankaj monicity, Phys. Rev. X 10, 041029 (2020).
Mandal, Umesh V. Waghmare, and Kanishka Biswas, [54] Tao Sun and Philip B. Allen, Lattice thermal conductiv-
Intrinsic Rattler-Induced Low Thermal Conductivity in ity: Computations and theory of the high-temperature
Zintl Type TlInTe2, Journal of the American Chemical breakdown of the phonon-gas model, Phys. Rev. B 82,
Society, Journal of the American Chemical Society 139, 224305 (2010).
4350–4353 (2017). [55] Olivier Delaire, Jie Ma, Karol Marty, Andrew F May,
[42] Yan-Ling Pei, Haijun Wu, Di Wu, Fengshan Zheng, and Michael A McGuire, Mao-Hua Du, David J Singh,
Jiaqing He, High Thermoelectric Performance Realized A Podlesnyak, G Ehlers, MD Lumsden, et al., Giant an-
in a BiCuSeO System by Improving Carrier Mobility harmonic phonon scattering in PbTe, Nature materials
through 3D Modulation Doping, Journal of the Ameri- 10, 614–619 (2011).
can Chemical Society, Journal of the American Chemical [56] G. Kresse and J. Hafner, Ab initio molecular dynamics
Society 136, 13902–13908 (2014). for liquid metals, Phys. Rev. B 47, 558–561 (1993).
 14

[57] G. Kresse and J. Hafner, Ab initio molecular-dynamics [64] Wu Li, Jesús Carrete, Nebil A. Katcho, and Natalio
simulation of the liquid-metal˘amorphous-semiconductor Mingo, ShengBTE: A solver of the Boltzmann transport
transition in germanium, Phys. Rev. B 49, 14251–14269 equation for phonons, Computer Physics Communica-
(1994). tions 185, 1747 – 1758 (2014).
[58] G. Kresse and J. Furthmüller, Efficiency of ab-initio to- [65] Ion Errea, Matteo Calandra, and Francesco Mauri, An-
tal energy calculations for metals and semiconductors us- harmonic free energies and phonon dispersions from the
ing a plane-wave basis set, Comput. Mater. Sci. 6, 15–50 stochastic self-consistent harmonic approximation: Ap-
(1996). plication to platinum and palladium hydrides, Phys. Rev.
[59] G. Kresse and J. Furthmüller, Efficient iterative schemes B 89, 064302 (2014).
for ab initio total-energy calculations using a plane-wave [66] Terumasa Tadano and Shinji Tsuneyuki, Self-consistent
basis set, Phys. Rev. B 54, 11169–11186 (1996). phonon calculations of lattice dynamical properties in cu-
[60] P. E. Blöchl, Projector augmented-wave method, Phys. bic SrTiO3 with first-principles anharmonic force con-
Rev. B 50, 17953–17979 (1994). stants, Phys. Rev. B 92, 054301 (2015).
[61] John P. Perdew, Adrienn Ruzsinszky, Gábor I. Csonka, [67] James E. Saal, Scott Kirklin, Muratahan Aykol, Bryce
Oleg A. Vydrov, Gustavo E. Scuseria, Lucian A. Con- Meredig, and C. Wolverton, Materials Design and Dis-
stantin, Xiaolan Zhou, and Kieron Burke, Restoring the covery with High-Throughput Density Functional The-
Density-Gradient Expansion for Exchange in Solids and ory: The Open Quantum Materials Database (OQMD),
Surfaces, Phys. Rev. Lett. 100, 136406 (2008). JOM 65, 1501–1509 (2013).
[62] John P. Perdew, Kieron Burke, and Yue Wang, General- [68] Scott Kirklin, James E Saal, Bryce Meredig, Alex
ized gradient approximation for the exchange-correlation Thompson, Jeff W Doak, Muratahan Aykol, Stephan
hole of a many-electron system, Phys. Rev. B 54, 16533– Rühl, and Chris Wolverton, The Open Quantum Ma-
16539 (1996). terials Database (OQMD): assessing the accuracy of
[63] P. Hohenberg and W. Kohn, Inhomogeneous Electron DFT formation energies, npj Computational Materials
Gas, Phys. Rev. 136, B864–B871 (1964). 1, 15010 (2015).