In this paper
Abstractmain.pdfIntroductionTheoryAllen–Feldman conductivity and its site projectionGauge freedomGauge-free sensitivities and an Euler sum ruleExact gradient and decompositionLocal softnessModels and calculationsResultsThe heat-flow map is robustThe sensitivity map is anti-correlated with the heat-flow mapMechanism: local softness and mode mixingTemperature dependenceParacrystalline siliconMaterial design: additivity and substituent placementAn amorphous/crystalline interfaceDiscussionConclusionsAcknowledgmentsRemarks on the original formulationThe Euler ruleDerivation of the gradientReferencesSupplemental Material for “Where heat flows and where to change it: gauge-free sensitivity maps of thermal conductivity in disordered solids”SM.pdfBond compensationBroadening, widths and heat capacitiesValidationPotentials and structuresVibrational densities of states evaluationLocal softness as a diagonal of the pseudo-inverseSite currents in an arbitrary gaugeGauge dependence of the site projectionAnti-correlation and mediation at KSign and magnitude of the mass mapDecomposition of the covariance and pair-resolved rotation termThe 32768-atom replicaParacrystalline modelsDesign testsInterfaceFully anharmonic check: per-atom Green–KuboBondsReferencesWorksPhysics and Astronomy
Where heat flows and where to change it: gauge-free sensitivity maps of thermal conductivity in disordered solids
Abstract
Designing materials for thermal management requires knowing where heat flows, not only how much. Within harmonic (AllenFeldman) theory we develop two local descriptions and test them on amorphous silicon. The first, a projection of the conductivity onto atoms, depends on how the potential energy is divided among atoms (a "gauge"); we find the dependence moderate, 23% of the map's spatial standard deviation at single sites and 9.5% over regions of radius 0.4 nm, apart from a uniform offset of 0.8% of . The second avoids the division altogether. It is the exact response of to the mass of each atom and the stiffness of each bond, it obeys an Euler sum rule, and it contains the heat-flow map as one of three exact terms. In amorphous and paracrystalline silicon, with ACE, GAP-18 and Keating force constants, the two maps are anti-correlated: where little heat flows a heavier atom lowers , and where much heat flows it tends to raise it. The explanation is local softness. The harmonic vibrational amplitude of each atom predicts both maps, and at room temperature the anti-correlation largely disappears at fixed amplitude, in a 32,768-atom replica as well; below about 100 K a component remains that softness does not capture. Thus soft regions, which carry little heat, act through the mixing of vibrational modes about 12 meV apart. In paracrystalline silicon the grains conduct more heat per atom, while the disordered matrix controls the response. Exact recalculations show that the response map guides the placement of dispersed substituents and locates the region that limits transport across an amorphous/crystalline interface.
main.pdf
Introduction
Heat conduction in solids, especially disordered ones, is a fundamental problem for basic science and for applications. Heat must be removed from nanoscale devices through amorphous dielectrics and across interfaces [1, 2]; thermoelectrics need a low lattice conductivity [3]; phase-change memories and thermal barrier coatings depend on the conductivity of disordered phases [4, 5]. The theory is well developed. Crystals are treated by Boltzmann transport with first-principles phonons [6]; glasses by the harmonic theory of Allen and Feldman (AF) and its unified extensions that include anharmonic damping [7, 8]; and any of them by Green–Kubo molecular dynamics. Each delivers a single number, the conductivity . Machine-learned interatomic potentials now give force constants of near first-principles quality for – atoms [9], so local descriptions of heat transport are within reach, and they are what design needs: where in a material does heat flow, and where would a change matter? We use amorphous silicon as a proving ground. It is the simplest tetrahedral network, its structure and vibrations are well characterized, and accurate potentials exist. The methods apply to any harmonic network.
Amorphous silicon (a-Si) conducts heat poorly, and it is worth knowing both why and where. The AF theory answers the first question [10–13]: heat is carried mostly by diffusons, delocalized but non-propagating vibrations that pass energy between modes of nearly equal frequency. Long-wavelength propagons add a contribution that depends on sample size, and localized vibrations (locons) carry almost nothing; the measured conductivity of a-Si films is about 1–3 W/(m K), depending on thickness [14, 15]. The second question, where in the network heat is carried and where a local change would alter , is harder. It is the subject of this paper.
In earlier work we projected onto atoms, the site-projected thermal conductivity (SPTC) [16–19], and applied it to the phase-change material GeSbTe [20]. The projection has a feature familiar to anyone who has computed a Mulliken charge: it is not unique. The heat current depends on how the potential energy is divided among atoms, and this division (a “gauge”) is arbitrary [21, 22]; for machine-learned potentials the atomic energies are themselves a by-product of the fit [23]. The total is indifferent to the gauge; the site values are not. Thus a reader is entitled to ask whether a map of SPTC describes the network or the bookkeeping.
We answer in two ways. First, we measure the gauge dependence: it is moderate at single sites and small over regions a few Å across. Second, and more interesting, we ask a question with no gauge ambiguity at all: if one changes the mass of one atom, or the stiffness of one bond, how does change? Since the AF expression for involves no division of the energy among atoms, neither does its derivative (Sec. II). We derive the exact gradient of the AF conductivity with respect to the dynamical matrix, contract it into mass and bond sensitivity maps, and decompose the mass map exactly into three terms, one of which is the SPTC itself. The answer is at first surprising. In the regions that carry the least heat a heavier atom lowers , and in those that carry the most it tends to raise it: the sign of the response is anti-correlated with the heat flow, and averaged over a few Å its magnitude is larger where little heat flows. A single local quantity, the harmonic vibrational amplitude of each atom, mediates the relation, and at fixed the anti-correlation between the two maps largely vanishes. Soft, large-amplitude sites (“rattlers”) contribute little to transport, but a change there mixes the vibrational modes that do.
Soft regions are no strangers to the physics of glasses. Quasi-localized low-frequency modes identify “soft spots” that predict where plastic rearrangements occur under shear [24–26], and elastic heterogeneity has been linked to vibrational anomalies and to the thermal conductivity of glasses [27, 28]. Our results give the thermal counterpart: in a-Si the soft regions are both where heat flows least and where its flow can be changed most.
The paper is organized as follows. Section II defines the SPTC and the sensitivity maps, derives the exact gradient and its decomposition, and shows that is a local diagonal of the pseudo-inverse of the dynamical matrix. Section III describes the models and calculations. Section IV establishes the robustness of , the anti-correlation of the two maps and its mechanism, its temperature dependence, paracrystalline silicon, consequences for design, and an application to an amorphous/crystalline interface. § 3.0.3 discusses limitations and connections to other work.
Theory
Allen–Feldman conductivity and its site projection
Let be the force-constant matrix, the mass of atom (atoms are labelled throughout, as in Hardy’s notation [29]), and the dynamical matrix, with eigenpairs , . Cartesian indices run over the degrees of freedom and is the atom of . With the bond-weighted operators , where denotes the elementwise (Hadamard) product, , (minimum image), the AF conductivity is
where holds the eigenvectors and [11]
with the volume, the heat capacity of mode (quantum unless stated) and the half-width of the Lorentzian; with per-mode widths , . In SI units ( in , in m, in J/K) is in W/(m K). Modes below 0.5 meV, which include the three translations, are excluded. We quote in kelvin, as the temperature ; K corresponds to meV. Writing as a trace over sites defines the site projection
Anchoring each bond current at the row atom is one choice of local energy, the “AF gauge”. Any other choice, in practice the site energies of the potential, gives a different with the same sum [21]. The weight is symmetric in through the factor ; this leaves unchanged and makes independent of whether a bond current is anchored at its row or its column atom (Appendix A).
Gauge freedom
A local energy assigns to each atom a quadratic form with . The total energy fixes but not the : this is the gauge freedom. The AF choice gives each atom its diagonal block and half of each mixed block; a many-body potential with site energies gives , which couples all atoms within the range of . Energy conservation gives the site currents with [29], and in the mode representation the current matrix element is proportional to
where and are the antisymmetric and symmetric parts of , with mass weighted (Supplemental Material, Sec. S8). In the AF gauge is antisymmetric and . Three results follow.
(i) is gauge invariant up to . Under with , the antisymmetric part of the total current changes by . Only depends on the gauge, and in Eq. (4) it is multiplied by , which vanishes where the Lorentzian concentrates the weight. For Stillinger–Weber, where every gauge can be computed exactly, the gauge differences of for a 196-atom model (Sec. III) are 0.29, 0.50, 0.95 and 1.9% at , 2, 4 and 8 K, linear in .
(ii) Site values are not invariant. For an individual site the antisymmetric part changes by . This piece sums to zero over sites, but it is not the time derivative of a bounded quantity and carries no factor , so the argument that protects does not protect : the single-site uncertainty remains as . For Stillinger–Weber it is 16, 13, 11 and 12% of the spread of at , 2, 4 and 8 K.
(iii) Regional values are protected up to a surface term. Summed over a region , the change involves , which vanishes by whenever atoms and lie deeper than inside . Only a shell of thickness contributes, so the local gauge ambiguity of a regional conductivity should fall roughly as for a region of size . It does, roughly, with a small floor: for si4096 (ACE, Å) it is 23% of the spatial spread of at single sites, 9.5% at Å and 4.6% at 10 Å, falling roughly as beyond about 4 Å, as expected for a surface term (Supplemental Material, Sec. S9). A uniform offset of 0.8% of , the global difference of result (i), remains; we quote it separately. The ranking of sites is nearly gauge-invariant beyond a few Å (Spearman 0.984 at 4 Å, 0.995 at 10 Å), and for a map of where heat flows that is the claim that matters. Thus the site projection is a well-defined description of where heat flows at the scale of a few Å, but not of individual atoms (Figure 1 and Figure 2). The gauge-invariant generalization of a local conductivity is the wave-vector-dependent , whose gauge dependence vanishes as [21]. The energy density of density-functional theory has the same ambiguity, and unique atomic energies follow from integrating over volumes on whose boundaries the gauge terms vanish [30, 31]; result (iii) is an approximate analogue for the heat current.
Figure 1. The 4096-atom a-Si network (ACE, K). (a) The whole cell: translucent surfaces enclose the 12% of the volume where the heat-flow map (blue) and the mass-sensitivity map (orange), smoothed with a Gaussian of 2.2 Å, are largest. The two regions share 0.27% of the volume, against 1.44% for independent maps. (b)–(d) A 9 Å slab, with bonds coloured and weighted by fields averaged over 4 Å spheres: (b) the heat-flow map , (c) the gauge-free mass-sensitivity map , and (d) the harmonic vibrational amplitude (300 K). Darker colours and thicker bonds indicate larger field values. In the soft, large-amplitude regions, which carry little heat, is positive and large: a heavier atom there lowers .
Figure 2. The gauge question at a glance: the heat-flow map of the slab of Figure 1 at single sites (top row) and averaged over 4 Å spheres (bottom row). (a), (d) AF gauge; (b), (e) site-energy gauge of the ACE potential, on one colour scale per row; (c), (f) their difference, in units of the spatial standard deviation of in that row, on one diverging scale for both rows. The two gauges look alike at a glance because their difference is small compared with the range of ; the difference maps show where they part, after removal of the uniform offset between the gauges (0.8% of ). Between the gauges, Spearman 0.92 and gauge uncertainty 23% of the spatial spread at single sites, 0.984 and 9.5% at 4 Å. At 45% of single sites the gauges differ by more than 0.3 of the spread, against 11% of 4 Å regions.
Gauge-free sensitivities and an Euler sum rule
The site energies do not enter Eq. (1), which is a function of the dynamical matrix and the positions alone. Thus its derivatives with respect to masses and force constants are unique, and they require no partition of among atoms; this is the sense in which we call them gauge free. At finite , Eq. (1) is the conductivity of the AF gauge, and a conductivity built from another gauge differs from it by [result (i) above]; but the derivatives of such a conductivity would differ from ours by the derivative of that difference, and all of them coincide as . How large is the difference in practice? For Stillinger–Weber, where every gauge can be computed exactly, we obtained the mass maps of the AF gauge and of two site-energy gauges by finite differences in the 196-atom model of result (i): they differ by 0.1, 0.1, 0.2 and 0.4% of the spread of at , 2, 4 and 8 K, growing with and smaller than the gauge differences of in the same cell (0.3–1.9%), and they rank the sites identically (Spearman rank correlation 1.000; the statistics are described in Sec. III). Thus, at least for this potential, the maps are gauge free for practical purposes at finite as well.
We define the mass sensitivity
where, since frequencies scale as , is the response of to the local frequency scale at atom ; and the bond sensitivity , where is a small dimensionless parameter, the fractional change of the force constants of the pair: the blocks , are scaled by , with the self blocks compensated by , where , so that the perturbed matrix remains symmetric. For pair potentials is symmetric and this compensation preserves the acoustic sum rule exactly. It is well to observe that for many-body potentials is not symmetric, and the row sum of atom changes by , with (and that of by the opposite amount). No change of a single non-symmetric block keeps both the symmetry of and the sum rule exactly. Fortunately the residual is small: for nearest-neighbour bonds is 1–2% of the block (median; 3–4% at the 90th percentile) for ACE and 4% for Keating. Moreover, it cancels exactly when summed over all bonds of an atom, because is symmetric, so the bond sensitivities, summed over all pairs within the range of , obey the Euler rule below. The alternative that scales only , which preserves both properties for each bond, gives bond maps with Spearman correlation 0.986–0.998 with ours for nearest-neighbour bonds (sign changes for 0.3–2.9% of them), but violates the Euler rule by 4–22% (Supplemental Material, Table S1).
| model | asym (%) | Spearman | change (%) | sign (%) | top 100 | Euler, ours | Euler, alt. |
| si4096 ACE 2K | 1.6 / 3.9 | 0.986 | 6.1 | 2.9 | 0.91 | 22% | |
| si4096 ACE 10K | 1.6 / 3.9 | 0.996 | 3.0 | 1.5 | 0.97 | 4% | |
| si4096 Keating 2K | 3.6 / 6.8 | 0.998 | 4.1 | 1.5 | 0.95 | 15% | |
| R1 ACE 2K | 0.1 / 0.5 | 0.998 | 0.1 | 0.3 | 0.93 | 10% | |
| R2 ACE 2K | 1.1 / 2.8 | 0.994 | 2.9 | 1.5 | 0.98 | 13% | |
| R3 ACE 2K | 1.0 / 2.5 | 0.991 | 2.8 | 1.2 | 0.94 | 13% | |
| si512 GAP 10K | 2.4 / 5.1 | 0.995 | 4.9 | 2.6 | 0.99 | 7% |
Table S1. Bond compensation. Our definition compensates the self blocks by ; the alternative scales in all four blocks and preserves both the symmetry and the acoustic sum rule bond by bond. Since is symmetric, the two differ exactly by . Columns: relative size of (median / 90th percentile, %), Spearman correlation of the two maps, median relative change (%), sign changes (%), all for nearest-neighbour bonds ( Å); overlap of the 100 largest bonds; relative Euler residual of our definition and of the alternative.
Scaling all masses by , together with and , maps every frequency and gives . Euler’s theorem for this homogeneous function yields (Appendix B)
Here the temperature derivative acts only through the heat capacities. Equation (6) checks one number, the sum of the gradient over all sites. Thus it tests the current and eigenvalue parts and the and derivatives together, but it is blind to the spatial distribution: the rotation term of the decomposition derived in Sec. II D sums to zero identically. The site-by-site test is the comparison with finite differences (Supplemental Material, Sec. S3). The rule is also a consistency identity, not a partition: individual are of both signs and typically larger than (for si4096 at K the median of is 1.8 times the mean of , and the standard deviation of is 3.3 times it), and their sum reproduces through large cancellations. We read as a response. The same rule holds for the bond sensitivities summed over all pairs within the range of , not only nearest-neighbour bonds.
Exact gradient and decomposition
The gradient has an explicit part from and a spectral part from the eigenpairs. The spectral part is first-order perturbation theory in which divided differences handle degeneracies, the Daleckii–Krein formula for functions of Hermitian matrices [32–34]; the bilinear structure of Eq. (1) is a double operator integral in the sense of Birman and Solomyak [35]. We find (Appendix C)
for . On the diagonal is the partial derivative at fixed eigenvectors, which acts on only. For near-degenerate pairs ( below a tolerance relative to the spectral width) ; results are insensitive to the tolerance. Both terms of Eq. (7) are symmetric matrices. In the first, is symmetric because and both change sign when and are exchanged. In the second, and are both antisymmetric, and the elementwise product of two antisymmetric matrices is symmetric. This is as it must be: is symmetric, so any change is symmetric too, and the first-order change involves only the symmetric part of .
Since , the mass sensitivity is . Inserting Eq. (7) gives an exact three-term decomposition,
where is the weight of mode on atom and the local overlap of two modes. The first term arises because the explicit part of the gradient contracted with is exactly twice the AF-gauge . The eigenvalue-shift term is a sensitivity-weighted local density of states. The rotation term sums to zero over sites, because : it only redistributes sensitivity in space.
Local softness
The classical mean-square displacement of atom in the harmonic approximation is
where the sum excludes the three translations and is the pseudo-inverse of (Supplemental Material, Sec. S7). It is the quantity used by Gautam et al. to identify “rattling” sites [18]. Two features matter here. First, is a local spectral moment of the same mode weights that enter , weighted toward low frequencies. Second, the right-hand form needs no eigenvectors: it can be evaluated with three sparse linear solves per atom, at any system size; a complete map costs solves. Since is the same for every site of a monatomic network, enters our rank statistics only through the local compliance , a structural property of the harmonic network; its ranking, and every correlation built from it, is the same at all temperatures, and we use the classical form throughout, quoting magnitudes at 300 K. The quantum amplitude, which tends to a weighting at low temperature, is examined in Sec. IV D. We also use the local low-frequency weight .
Models and calculations
Structures. The reference model is the 4096-atom continuous random network of Djordjević, Thorpe and Wooten [36] scaled to Si (si4096; cubic cell of 43.42 Å, density 2.334 g/cm), relaxed in its fixed cell with each potential. After relaxation with the ACE potential it has 5 three-fold and 23 five-fold atoms (0.68% coordination defects at a 2.8 Å cutoff). For validation we use a 512-atom model of the same construction [36] (si512, fully four-fold). Four independent networks test generality: the 10 000-atom Barkema–Mousseau model [37] (BM10k), relaxed with ACE at zero pressure (0.56% defects) and annealed at 800 and 1000 K and relaxed again (1.24% and 1.51%); and three 1000-atom paracrystalline models R1–R3 of Rosset et al. [38]. For the interface application we use the 2022-atom a-Si/c-Si model of Bernstein, Aziz and Karplus [39], relaxed with GAP in earlier work [17] and relaxed again here with ACE. The Stillinger–Weber gauge tests of Sec. IIB use a small network with original Stillinger–Weber parameters [40]: a 216-site diamond cell with 20 sites removed at random, displaced randomly by 0.3 Å and relaxed (196 atoms, 17% three-fold). It is not meant as a model of a-Si. Stillinger–Weber is the one potential here whose site energies, and so every gauge, are known exactly, and a small defective cell suffices for that test; no result about a-Si rests on it. Coordination statistics of all models are in the Supplemental Material, Table S3.
Potentials. The atomic cluster expansion (ACE) [41, 42] potential for Si–H of Rosset and Deringer (SiH-ACE-25) [43] is used throughout; GAP-18 [44] (through LAMMPS/QUIP) is used for si512 and si4096 (for si4096 from sparse finite differences; Supplemental Material, Sec. S4). We also use the Keating valence force field [45] in the form of Wooten, Winer and Weaire [46], , with eV/Å (47.5 N/m), (13.5 N/m) and Å, on the bond topology of the input network. Each potential is used on its own relaxed structure. The vibrational densities of states of all models are compared with inelastic neutron scattering [47] in the Supplemental Material (Figure S1): the ACE and GAP-18 spectra reproduce the measured one closely. The Keating spectrum does not: its transverse-optic-like (TO-like) peak lies near 48 meV, about 10 meV below the measured band, and its transverse-acoustic-like (TA-like) band extends too high and fills the minimum near 30 meV (Kamitakahara et al. also found the TA-like band of their Keating model too high [47]).
Figure S1. (a) Vibrational density of states of all models, Gaussian broadened (FWHM 4 meV), against the inelastic neutron data for a-Si of Kamitakahara et al. [3] (crosses). BM10k from a kernel-polynomial expansion (1500 moments, 32 random vectors) of its sparse ACE dynamical matrix. (b) Participation ratio against energy, si4096 (ACE and Keating); dotted lines at 10 meV and at 56.5 meV, the boundaries used for propagons and locons. (c) Local VDOS of si4096 (ACE) averaged over the 10% of sites with the largest and smallest and (4 Å averages), and over all sites. Every curve in all panels is normalized to unit area over 0–80 meV, the neutron data over their measured range; panels (c) and (d) therefore compare shapes, not the numbers of atoms in each group. (d) Interface model: VDOS projected on the crystal more than 4 Å from the interfaces, on atoms within 4 Å of the two interfaces, and on the amorphous region more than 4 Å from them.
Dynamical matrices and conductivity. ACE dynamical matrices were computed by central finite differences (0.01 Å) in LAMMPS [48], on the GPU with the KOKKOS package. For BM10k we built a sparse dynamical matrix by finite differences with graph colouring (atoms displaced together if more than Å apart), truncated at 11 Å with the acoustic sum rule restored; it reproduces the dense matrices of si512 and si4096 to and . Unless stated otherwise we use K, quantum heat capacities and K (0.17 meV); K (0.86 meV) is used as a second, smoother scale. For si4096 and smaller models Eqs. (1)–(10) are evaluated by dense diagonalization (about 15 min for and , 17 min for the full gradient at 4096 atoms on one workstation). For BM10k, and come from a double Chebyshev expansion of Eq. (3) on the GPU [49, 50] (Supplemental Material, Sec. S6), and from block-preconditioned conjugate gradients on Eq. (11).
Validation. For si512 the analytic gradient agrees with finite differences from an independent code to a median relative error of over all sites (masses) and – over 48 bonds. The Euler rule, Eq. (6), holds to in every -broadened run, the three terms of Eq. (9) reconstruct to , and to . The conjugate-gradient agrees with the exact pseudo-inverse of the same matrix to (maximum relative difference over all 4096 sites of si4096; mean ), and truncation of the matrix at 11 Å changes by at most 0.4% (Pearson 0.99999 with the full dense result).
Statistics. Most comparisons ask whether two maps order the sites alike, so we use the Spearman (rank) correlation , which is insensitive to units, to monotonic rescaling and to the few extreme sites that are common in . The partial correlation correlates what remains of the ranks of and after linear regression on the ranks of ; it is near zero if accounts for the relation. Neighbouring sites are correlated, so we give 95% intervals from a spatial block bootstrap (300 resamples of periodic cubic blocks: 64 for si4096 and BM10k, 27 for the 1000-atom models and 8 for si512, so that blocks exceed the 8 Å diameter of an averaging sphere), at single sites and for fields averaged over spheres of 4 Å radius. As a yardstick for chance agreement we shift one map rigidly through the periodic cell, which keeps its spatial autocorrelation, and recompute the correlation (500 random translations, excluding those shorter than 8 Å, twice the averaging radius; single-site values are read at the nearest site, and 4 Å fields are evaluated at the shifted points). Shuffling sites, the usual alternative, destroys the autocorrelation and so underestimates chance correlations between smooth maps. The 95% level of the shifted correlation depends on the maps compared; for it is 0.12 and 0.04 at single sites and 0.28 and 0.09 at 4 Å (si512 and si4096; last row of Table I), for against in the 1000-atom models 0.15 at single sites and 0.29 at 4 Å, and we quote it alongside other correlations where it matters; smaller correlations cannot be told from chance, and this is the sense in which we call one “near zero”.
Results
The heat-flow map is robust
For si4096 with ACE, W/(m K) at K, of which diffusons contribute 0.956, propagons 0.241 and locons 0.021 W/(m K) (propagons below 10 meV, diffusons 10–56.5 meV, locons above 56.5 meV; a pair of modes in different bands contributes half to each). The AF-gauge map of and the map in the site-energy gauge of the potential (obtained by probing the per-atom energies with finite displacements; Supplemental Material, Sec. S9) have Spearman correlation 0.92 at single sites and 0.984 over 4 Å spheres. The gauge uncertainty (the standard deviation over the two gauges divided by the spatial standard deviation of their mean, after removing the uniform offset of result (i): is 0.78% lower in the site-energy gauge) is 23% of the spatial spread of at single sites and 9.5% at 4 Å (Figure 2).
Table I compares maps across potentials and models. Two machine-learned potentials, ACE and GAP-18, agree nearly as well as two gauges do (si512: 0.86 at single sites and 0.92 at 4 Å for K; si4096: 0.89 and 0.93, and 0.90 and 0.94 at K). GAP-18 gives a lower than ACE at matched broadening: for si4096 1.009 against 1.218 W/(m K) at K and 0.867 against 0.948 W/(m K) at 10 K (−17% and −9%; −14% and −8% for si512). ACE and Keating, each on its own relaxed structure, agree at 0.84–0.89 for si4096. The independent BM10k network reproduces the si4096 statistics: the relative spread of (standard deviation over mean) is 0.12–0.13 in both, and increases with the mean bond length of the atom (Spearman +0.61 to +0.63 in BM10k, +0.57 in si4096). Thus, at the 4 Å scale is a robust structural descriptor of where heat flows.
The sensitivity map is anti-correlated with the heat-flow map
Figure 3(a) compares the mass sensitivity with for si4096. The two are anti-correlated: Spearman −0.31 [−0.36, −0.26] at single sites and −0.41 [−0.49, −0.34] at 4 Å (95% intervals). Table II collects the models at K; K is in the Supplemental Material (Table S4). The anti-correlation appears with all three potentials and in every network of 1000 atoms or more: it is stronger with Keating (−0.66 at 4 Å) and in the paracrystalline models (−0.43 to −0.55), holds for classical as well as quantum heat capacities, and for broadening up to K. It is not driven by outliers: removing the top 1% of sites leaves it unchanged. With GAP-18 it is weaker than with ACE but well clear of chance at 4096 atoms (−0.31 [−0.37, −0.24] at 4 Å and K against a null of 0.09; −0.28 at single sites). At 512 atoms it is weak with both potentials: for GAP-18 −0.13 at K and −0.00 at 10 K at 4 Å, below the shifted nulls of 0.30 and 0.23, and for ACE −0.25 (null 0.21) against −0.41 at 4096 atoms. The weakness at 512 atoms is thus common to both potentials, as a finite-size effect would be, and is not a property of GAP-18. In the 32 768-atom replica of si4096 it is stronger still (−0.56 at 300 K; Sec. IV D). Thus the anti-correlation strengthens as the cell supports more low-frequency modes: from si512 to si4096, and in the 32 768-atom replica of si4096, which adds longer-wavelength modes to the same network. We return to this in § 3.0.3.

Figure 3. Mass sensitivity against heat flow for si4096 (ACE, K, fields averaged over 4 Å spheres and multiplied by ). (a) Density of the regions (grey, logarithmic) and the mean in each tenth of the regions ordered by , coloured by their mean vibrational amplitude . (b) Local softness mediates the relation: Spearman correlations between the three maps, and the partial correlation of and at fixed .
TABLE S4. As Table II of the main text, at K. : Spearman rank correlation of maps and (4 Å averages, 95% block-bootstrap intervals); : partial correlation of and at fixed .
| model, potential, | (W/(m K)) | ||||
| si4096, ACE, 10 K | 0.948 | −0.41 [−0.46, −0.36] | −0.70 [−0.73, −0.66] | +0.68 [+0.64, +0.72] | +0.12 [+0.06, +0.18] |
| si4096, Keating, 10 K | 1.166 | −0.85 [−0.87, −0.83] | −0.90 [−0.91, −0.88] | +0.90 [+0.89, +0.91] | −0.23 [−0.32, −0.16] |
| si4096, GAP-18, 10 K | 0.867 | −0.22 [−0.27, −0.16] | −0.61 [−0.66, −0.56] | +0.55 [+0.50, +0.59] | +0.18 [+0.12, +0.25] |
Table S4.
| comparison | site | 4 Å |
| gauge: AF vs site energies (ACE, si4096) | 0.92 | 0.984 |
| ACE vs GAP-18 (si512) | 0.86 | 0.92 |
| ACE vs GAP-18 (si512, K) | 0.92 | 0.96 |
| ACE vs GAP-18 (si4096) | 0.89 | 0.93 |
| ACE vs GAP-18 (si4096, K) | 0.90 | 0.94 |
| ACE vs Keating (si4096) | 0.84 | 0.89 |
| null (si512 / si4096) | 0.12 / 0.04 | 0.28 / 0.09 |
Table 1. Robustness of the heat-flow map (Spearman correlations, single site and 4 Å averages; K unless noted). Null: 95th percentile of the correlation with a cyclically shifted copy of the other map, the yardstick of Sec. III.
Since has both signs (38% of the sites of si4096 are negative at K), the anti-correlation is first of all a statement about sign. In every model the 10% of regions with the least heat flow have positive mean : a heavier atom there lowers . The 10% with the most heat flow have negative or small mean . The magnitude follows less closely. Over 4 Å the regions with the least heat flow have the larger mean (2.3 times that of the most heat-carrying regions for si4096 with ACE, 2.4 with Keating, 1.3–6 for R1–R3), and Spearman is −0.18 (ACE) and −0.41 (Keating); but at single sites the magnitude is nearly uncorrelated with (−0.02 for ACE at K), and for R2 and for GAP-18 at 512 atoms it is uncorrelated at 4 Å as well (−0.14 for GAP-18 at 4096 atoms) (Supplemental Material, Table S5). Thus “where the response is large” and “where heat flows little” coincide only on average; what the heat flow sets reliably is the sign of the response. The picture is plain in three dimensions, where the two maps occupy largely different parts of the network: the 12% of the volume where the smoothed is largest and the 12% where is largest share 0.27% of the volume, five times less than independent maps would [Figure 1(a)].
TABLE S5. Signed and absolute mass sensitivity against the heat-flow map: fraction of sites with ; Spearman correlation of with and with at single sites and over 4 Å spheres, and with the absolute value of the 4 Å average of ; ratio of the mean of the 10% of 4 Å regions with the smallest to that of the 10% with the largest. 95% block-bootstrap intervals are ±0.02–0.20 (Sec. III of the main text gives the block counts).
| model | site: | site: | 4 Å: | 4 Å: | ratio | ||
| si4096 ACE 2K | 0.38 | −0.31 | −0.02 | −0.41 | −0.18 | −0.30 | 2.3 |
| si4096 ACE 10K | 0.20 | −0.31 | −0.20 | −0.41 | −0.32 | −0.41 | 1.8 |
| si4096 Keating 2K | 0.42 | −0.63 | −0.24 | −0.66 | −0.41 | −0.39 | 2.4 |
| R1 ACE 2K | 0.44 | −0.27 | −0.22 | −0.51 | −0.67 | −0.43 | 6.1 |
| R2 ACE 2K | 0.43 | −0.31 | +0.06 | −0.55 | −0.03 | −0.16 | 1.3 |
| R3 ACE 2K | 0.40 | −0.29 | −0.08 | −0.43 | −0.31 | −0.14 | 1.3 |
| si512 GAP 10K | 0.27 | −0.18 | −0.01 | −0.00 | +0.18 | −0.00 | 1.0 |
| si4096 GAP 2K | 0.36 | −0.28 | −0.02 | −0.31 | −0.14 | −0.25 | 2.0 |
| si4096 GAP 10K | 0.23 | −0.28 | −0.13 | −0.22 | −0.09 | −0.22 | 1.4 |
Table S5.
The extreme values of at K belong to under-coordinated atoms with quasi-localized resonances: 4 of the 11 most sensitive sites of si4096 are among its 5 three-fold atoms, and their sensitivity is dominated by a single mode (for example at 7.3 meV with participation ratio 0.06). These resonances decorate the map but do not cause the anti-correlation: the Keating network keeps the four-bonded topology of its input and has no coordination defects, yet its anti-correlation is stronger; a defect-free ACE model [51] would test this further.
| model, potential, | (W/(m K)) | ||||
| si4096, ACE, 2 K | 1.218 | [, ] | [, ] | [, ] | [, ] |
| si4096, Keating, 2 K | 1.734 | [, ] | [, ] | [, ] | [, ] |
| si4096, GAP-18, 2 K | 1.009 | [, ] | [, ] | [, ] | [, ] |
| R1, ACE, 2 K | 8.15 | [, ] | [, ] | [, ] | [, ] |
| R2, ACE, 2 K | 1.72 | [, ] | [, ] | [, ] | [, ] |
| R3, ACE, 2 K | 1.74 | [, ] | [, ] | [, ] | [, ] |
| si512, ACE, 2 K | 1.091 | [, ] | [, ] | [, ] | [, ] |
| si512, ACE, 10 K | 0.972 | [, ] | [, ] | [, ] | [, ] |
| si512, GAP-18, 2 K | 0.944 | [, ] | [, ] | [, ] | [, ] |
| si512, GAP-18, 10 K | 0.892 | [, ] | [, ] | [, ] | [, ] |
| BM10k, ACE, 2 Ka | 1.231 | — | [, ] | — | — |
Table 2. TABLE II. Anti-correlation of the mass map with the heat-flow map and its mediation by local softness. : Spearman rank correlation of maps and , for fields averaged over 4 Å spheres, with 95% block-bootstrap intervals. Last column: , the partial correlation of and at fixed , that is, the correlation of what remains of the two maps once the part of each that follows is removed (Sec. III). K, quantum heat capacities; classical. aSingle sites (455 sampled sites of the relaxed model; not computed).
Mechanism: local softness and mode mixing
Local softness mediates the anti-correlation. The heat-flow map is strongly anti-correlated with the local vibrational amplitude [Table II, Figure 3(b), and Supplemental Material, Table S4]: to at 4 Å in every model for which was computed, and in BM10k (single sites). The sensitivity map is positively correlated with : to , including for GAP-18 at 4096 atoms ( for GAP-18 at 512 atoms and K). Soft, large-amplitude regions carry little heat, and therefore is large and positive: a heavier atom lowers most. At K and fixed the partial correlation of and falls to , and for si4096 with ACE, Keating and GAP-18, and lies between and in every ACE and Keating case and for GAP-18 at 4096 atoms, against to without conditioning (Table II and Supplemental Material, Table S4). At K the partial correlation becomes positive for both machine-learned potentials ( [, ] for ACE, [, ] for GAP-18; Supplemental Material, Table S4), so at that broadening the amplitude more than accounts for the anti-correlation. The local low-frequency weight is an equally good mediator; conditioning on both adds nothing over either alone, so both measure the same local property. In spectral terms, the 10% of sites with the largest and the 10% with the smallest have about twice the weight below 12.4 meV of their opposite numbers (Supplemental Material, Sec. S5).
The independent BM10k network confirms the link without any eigenvectors. With from sparse linear solves and from the Chebyshev expansion at sampled sites (455, 522 and 547 in the three states), the single-site [, ] for the relaxed model and and after the 800 and 1000 K anneals. Restricted to four-fold sites it is to , so the relation is not carried by defects (Supplemental Material, Sec. S6).
Which term is mediated. Equation (9) shows how (Figure 4). Of the covariance between and at 4 Å, the current term positively correlated with by construction, opposes the total (share to in si4096 and R1–R3 at and 10 K), the eigenvalue-shift term carries 7–38%, and the rotation term 80–171%, overcoming the current term in each of these models; with GAP-18 at 4096 atoms the pattern is the same but the shares are larger (, , at K), because the total covariance is smaller; in the 512-atom cells, where the total covariance is small, all three shares are larger (Supplemental Material, Table S6). Conditioning on removes most but not all of the rotation term’s anti-correlation with : for si4096 with ACE at K it falls from to , and for R1–R3 from to . A residual remains with Keating ( at 2 K, at 10 K). Local softness thus sets the magnitude of the effect, and eigenvector rotation is the microscopic route; in the stiffest, most uniform network (Keating) part of the rotation term carries information that does not.
Figure 4. Which term is mediated: Spearman correlation (4 Å) of each term of Eq. (9) with , without conditioning (pale) and at fixed (solid), for si4096 (ACE and Keating, and 10 K), R1–R3 (ACE, 2 K) and si4096 (GAP-18, 2 K); bars are 95% block-bootstrap intervals. The other models of Table II are not shown. The correlations of the three maps themselves are in Table II and Figure 3(b).
Top: share of the covariance between and (4 Å averages) carried by each term of , for every model. Bottom: share of for si4096 by mode-pair separation (in units of ) and by the mean frequency of the pair ( K unless noted). The total covariance is close to zero (Spearman ), so the shares are not meaningful.
| model | |||
| si4096 ACE 2 K / 10 K | / | / | / |
| si4096 Keating 2 K / 10 K | / | / | / |
| R1 / R2 / R3, ACE 2 K | / / | / / | / / |
| si512 ACE 2 K / 10 K | / | / | / |
| si512 GAP-18 2 K / 10 K | / — | / — | / — |
| si4096 GAP-18 2 K / 10 K | / | / | / |
| si4096 ACE | si4096 Keating | ||
| separation 0–1 / 1–3 / 3–10 / | 0.01 / 0.28 / 0.41 / 0.30 | 0.08 / 0.18 / 0.34 / 0.39 | |
| same, K | 0.32 / 0.43 / 0.27 / −0.02 | 0.24 / 0.33 / 0.34 / 0.09 | |
| pair mean frequency 0–10 / 10–30 / 30–45 / meV | 0.39 / 0.54 / 0.09 / −0.03 | 0.23 / 0.69 / 0.06 / 0.01 | |
| same, K | 0.22 / 0.68 / 0.15 / −0.06 | 0.07 / 0.74 / 0.16 / 0.03 |
Table S6.
The mixing window. Resolving into mode pairs shows that the anti-correlation is carried by pairs of modes about 1–2 meV apart, mostly in the 10–30 meV band of low diffusons (54–74% of the covariance from pairs whose mean frequency lies in that band); modes above 45 meV contribute nothing. The window is set mainly by the network, not by the broadening: at K (0.17 meV) pairs within carry only 26–29% of the covariance. It is set by the local spectral correlation of the modes [Fig. 5]. For the 4286 (ACE) and 3719 (Keating) modes of the 10–30 meV band we computed, as functions of , the co-localization and the local overlap strength , which measure whether two modes live on the same atoms. Both decay to half their value over 1.5–2.6 meV (: 1.5 meV for ACE, 2.4 meV for Keating; : 2.3 and 2.6 meV), independent of . The rotation covariance of these pairs reaches half its total at 0.9 and 1.4 meV for ACE ( and 10 K) and at 1.3 and 1.9 meV for Keating: a fivefold change of widens the window by only half, and the potential with the broader local overlap has the broader window. By contrast, the correlations of the current matrix elements (, the -weighted overlap of rows, and ) show no common decay scale (Supplemental Material, Table S7). The mixing window is therefore the energy range over which two vibrational modes share the same atoms: a local mass change can only mix modes that overlap locally, and in a-Si that requires them to be within 1–2 meV. This is, in our view, the vibrational form of the island picture of localized electronic states in disordered networks, in which eigenstates close in energy are built from the same spatial islands [52, 53].
Decay widths (separation at half maximum, meV) of pair correlations among the 10–30 meV modes of si4096, and the half-point of the cumulative rotation covariance (30 logarithmic bins, 0.02–20 meV; the half-points are interpolated in within the bin where the cumulative share crosses one half; pairs of 10–30 meV modes).
| ACE | Keating | |
| co-localization | 2.3 | 2.6 |
| local overlap | 1.5 | 2.4 |
| non-monotonic | 2.7 | |
| -weighted row overlap , / 10 K | 1.3 / 1.7 | 3.7 / 3.5 |
| , / 10 K | 0.34 / no decay below 10 meV | 0.12 / 0.81 |
| half-point, 4 Å, / 10 K | 0.9 / 1.4 | 1.3 / 1.9 |
| half-point, single sites, / 10 K | 1.4 / 2.0 | 1.6 / 2.4 |
| share of from pairs with both modes in 10–30 meV, 4 Å, / 10 K | 0.43 / 0.51 | 0.59 / 0.64 |
Table S7.
The scale is suggestive of a Thouless energy [54, 55]. The AF diffusivity averaged over the 10–30 meV band is 0.94 mm/s for ACE and 1.60 mm/s for Keating, a ratio of 1.7, while the local-overlap widths differ by a factor of 1.6 (1.5 and 2.4 meV). Writing the width as gives the same length for both potentials: and 6.6 Å from the overlap width, and 8.3 and 9.1 Å from the half-point of the rotation covariance at K. These lengths are 2.5–4 bond lengths, close to the Ioffe–Regel scale, so the diffusive form is an order-of-magnitude statement. Modes closer in energy than remain correlated over a region about 0.6–0.9 nm across, and only such pairs are mixed by a local perturbation. It is reasonable to conjecture that the window is a Thouless energy of this kind; the two potentials are consistent with it, but they do not prove it.
Temperature dependence
All points from 50 K to the classical limit, at and 10 K, were computed from a single diagonalization of si4096 (ACE); the Euler rule closes to at each. The anti-correlation persists at all temperatures (Figure 6) and is strongest at low : Spearman at 4 Å is at 100 K and in the classical limit ( K), and and at K. The rotation term carries 93–122% of the covariance at every temperature, and half of it comes from pairs less than 1 meV apart at K (1.4–1.7 meV at K) at all . The mixing window is thus a property of the spectrum and the broadening, not of the thermal population. The map of is stable in (Spearman 0.97 between 100 and 600 K at K); the map of less so (0.79), because at low the soft modes weigh more.
Figure 6. Spearman correlation of and at 4 Å for si4096 (ACE) against temperature (“cl.”: classical heat capacities), without conditioning (filled) and at fixed classical (open), for and 10 K, with 95% block-bootstrap intervals. Stars: the 32 768-atom replica of si4096 ( K), without conditioning (filled) and at fixed (open).
Above 200 K the classical largely mediates the anti-correlation (partial correlations to ). At 50–100 K it mediates only part: the partial correlation is to . Zero-point motion does not account for the remainder. The quantum amplitude, , ranks sites almost as the classical one does (Spearman 0.97 at 50 K, 0.996 at 100 K), and conditioning on it leaves partial correlations of to . Nor is the residual carried by the heat-capacity channel of the eigenvalue term (heavier atoms lower frequencies and raise the populations of modes with of a few ): this channel is 31% of at 50 K, but it is positively correlated with , and removing it from leaves the residual unchanged (−0.23). At 100 K, conditioning jointly on and on the local weight in the thermally active band () reduces the residual to −0.08; at 50 K it does not (−0.23).
Figure 5. Origin of the mixing window, si4096. (a) Cumulative share of the covariance between and (4 Å averages) carried by pairs of 10–30 meV modes closer than . (b) Co-localization and local overlap strength of the same mode pairs, normalized to 1 at small separation and 0 at 5–20 meV.
The residual is not removed by adding longer-wavelength modes. We repeated the analysis on the 32 768-atom replica (Supplemental Material, Sec. S13 and Table S8) of si4096, computed as eight independent -blocks (the current operator does not couple different ), whose lowest modes reach down to 2.0 meV instead of 3.9 meV. At 300 K the replica confirms the main result: the anti-correlation is stronger (−0.56 against
| 50 K | 100 K | 300 K | ||||
| cell | replica | cell | replica | cell | replica | |
| (W/(m K)) | 0.378 | 0.397 | 0.764 | 0.783 | 1.218 | 1.237 |
| lowest mode (meV) | 3.9 | 2.0 | 3.9 | 2.0 | 3.9 | 2.0 |
Table S8. Replica against cell (ACE, K, 4 Å averages; 95% block-bootstrap intervals).
−0.41), and at fixed it again largely disappears (−0.05 [−0.12, 0.00]). At low temperature the residual grows when the longer-wavelength modes are included: at fixed it is −0.38 at 100 K (cell −0.17) and −0.49 [−0.55, −0.41] at 50 K (cell −0.23). Thus the additional long-wavelength modes, which carry much of the heat when the thermal energy is only a few meV, contribute an anti-correlation that local softness does not capture. The picture is easily stated: in the diffuson regime the response map is local and is set by the soft regions; when propagons carry the heat it acquires a component that local softness does not capture. Three simple explanations fail. Restricting the AF weight to pairs of propagons (both below 10 meV) splits exactly; at 50 K these pairs carry 54% of and 78% of the covariance of with , but the part not explained by is about the same in the propagons and the remaining contributions (−0.40 and −0.36). Nor is it softness on a longer scale: averaging over spheres of 6, 8 and 12 Å makes the residual larger (−0.56, −0.63, −0.68), so softness on the larger scale is not the missing mediator; conditioning jointly on at 4 and 12 Å leaves −0.49 at 50 K, so the residual is not carried by softness on either scale. Nor is it an inconsistency between quantum heat capacities in the maps and a classical amplitude: with the quantum amplitude, including zero-point motion, the replica residual is larger still (−0.60 at 50 K and −0.44 at 100 K), while at 300 K the two amplitudes give the same result. Conditioning on both amplitudes together, which are different moments of the same local mode weights, roughly halves the residual (−0.29 at 50 K, −0.19 at 100 K), far more than a second field with the same spatial structure but displaced at random does (Supplemental Material, Sec. S13); this suggests that the shape of the local vibrational spectrum, not a single moment of it, mediates the correlation. The component lies in the rotation term: at fixed , keeps a partial correlation of −0.49 with at 50 K but only −0.10 at 300 K, while the eigenvalue term is unchanged (−0.28 and −0.29). Thus at low temperature the mixing of modes is no longer set by local softness alone: nearly half of the residual follows the shape of the local vibrational spectrum, and the rest remains open. Crystallinity alone. A plausible reason is that the grains of R1 are large enough to carry propagon-like modes, which respond to a mass change differently from diffusons. The negative sensitivity of grain interiors is a first-order statement only. Exact recalculation shows that is nearly stationary with respect to grain-interior masses, as for the crystal at the interface (Sec. IV G): 5% heavier raises of R3 by 0.3% (linear 0.5%), but 20% heavier lowers it by 0.5% (linear +1.9%). The same relation of softness to heat flow and to sensitivity that holds within a-Si holds here between the two phases, which gives a thermal signature of paracrystallinity: it is the disordered fraction, not the crystalline one, that sets how responds to change.

Figure 7. The paracrystalline model R3 (ACE): a 9 Å slab through the centre of its single grain (236 atoms, 24% of the cell). (a) Atoms classified by diamond-structure identification (OVITO): crystalline (cubic or hexagonal diamond, dark), grain edge (first and second neighbours of crystalline atoms, light) and matrix (grey). (b) Heat-flow map and (c) mass sensitivity on a diverging scale centred at zero (blue negative, orange positive), both averaged over 4 Å spheres, K. The grain carries more heat and has negative ; the matrix carries less and has positive .
Paracrystalline silicon
Paracrystalline silicon, amorphous silicon containing crystalline grains of nanometre size [56, 57], has been inferred from fluctuation electron microscopy [58], and the models R1–R3 of Rosset et al. [38] make it concrete. Identifying cubic and hexagonal diamond environments (with OVITO), we find crystalline fractions of 18% (R2), 24% (R3) and 93% (R1, essentially polycrystalline), against 0.1% in the DTW and Barkema–Mousseau networks. The two paracrystalline models conduct about 40% better than a-Si (1.72 and 1.74 W/(m K) for R2 and R3 against 1.22 W/(m K), K), but with two models the trend with crystalline fraction cannot be resolved; R1 (8.15 W/(m K)) is limited by the broadening and is not comparable. The maps show where the difference arises [Fig. 7]; profiles and regional averages in the Supplemental Material, Figure S2 and Table S9.

Figure S2. R2 and R3 (ACE, K, single sites averaged in 1.5 Å bins of the signed distance to the grain boundary; negative: amorphous matrix, positive: grains). (a) Local heat flow . (b) Mass sensitivity . (c) Classical vibrational amplitude (300 K).
| model | crystalline | , / 10 K | region | / | |
| si4096, BM10k | 0.1% | 1.22–1.23 / 0.95 | |||
| R2 | 18.1% | 1.72 / 1.28 | grain interior ( Å) | 2.05 / −3.79 | 0.1228 |
| R2 | 18.1% | 1.72 / 1.28 | matrix ( Å) | 1.62 / +3.14 | 0.1281 |
| R3 | 24.3% | 1.74 / 1.32 | grain interior | 2.06 / −3.82 | 0.1225 |
| R3 | 24.3% | 1.74 / 1.32 | matrix | 1.60 / +3.05 | 0.1286 |
| R1 | 92.7% | 8.15 / 3.29 | grain interior | 8.42 / +2.12 | 0.1220 |
| R1 | 92.7% | 8.15 / 3.29 | matrix | 4.69 / +108.5 | 0.1487 |
Table S9. Crystalline fraction and conductivity; and , (W/(m K)) and (Å) by region.
Moving from the amorphous matrix into a grain, the local heat flow rises steadily (for R3, from 1.60 in the matrix to 2.06 W/(m K) in grain interiors), the vibrational amplitude falls (0.1286 to 0.1225 Å), and the mass sensitivity changes sign (from +3.1 to −3.8 W/(m K)). Thus the grains conduct about 30% more heat per atom than the matrix, but, at 24% of the atoms, they do not carry most of it; the grain boundaries and the amorphous matrix control the response. The nearly polycrystalline R1 is different: its grain interiors keep a positive (+2.1, against +108 in its thin matrix), so the change of sign belongs to grains embedded in a matrix, not to crystallinity alone.
Material design: additivity and substituent placement
The sensitivity map is a first-order response. We tested how far it predicts finite changes by exact recalculation of for si4096 after each perturbation, at and 10 K; where sites are chosen from the map, the K map is used. Details and the full results are in the Supplemental Material, Sec. S15 and Table S10.
| , fraction | most sensitive | random | least sensitive | (top−random)/sd | random, unconstrained |
| 2 K, 1% | −7.96 | −4.64 ± 0.25 | −3.36 | 13 | −4.76 ± 0.19 |
| 2 K, 3% | −16.82 | −11.84 ± 0.31 | −9.10 | 16 | −12.42 ± 0.18 |
| 2 K, 5% | −22.36 | −17.15 ± 0.23 | −14.09 | 23 | −18.69 ± 0.17 |
| 10 K, 1% | −5.12 | −3.41 ± 0.11 | −3.04 | 15 | −3.39 ± 0.11 |
| 10 K, 3% | −12.21 | −9.23 ± 0.21 | −8.34 | 14 | −9.26 ± 0.26 |
| 10 K, 5% | −17.17 | −14.11 ± 0.16 | −13.04 | 19 | −14.27 ± 0.13 |
Table S10. Dispersed Ge (sites at least 5 Å apart): (%). Random: mean ± standard deviation over five sets. Last column: random sets drawn without the separation constraint.
Isotopes and bonds. Isotope substitution () and bond-stiffness changes of ±5% are nearly linear over regions of up to 46 atoms (6 Å radius): the ratio of the exact change to the linear prediction is 1.10–1.21 for isotopes and 0.93–1.06 for bonds. Bond changes of ±20% are not (0.67 and 1.17). Pairs of substituents act additively to within 2% beyond 5 Å; nearest neighbours interact at the 7–9% level.
Heavy substituents. We model germanium by its mass alone (, Si force constants, no relaxation): chemical bonding, structural relaxation and species-dependent force constants are excluded, so the results are not predictions for real Si–Ge alloys. This massonly Ge is far outside the linear regime: in compact regions the exact change is 2.3–4.0 times the prediction at K, and every Ge substitution lowers , including where the map predicts a rise. In compact clusters the effect does not depend on where the cluster sits: at K a 46-atom Ge region lowers by 2.33, 2.38 and 2.42% when centred on the sites of largest positive , random sites, and the sites of most negative . A compact cluster scatters as a single heavy defect. For dispersed substituents the map is a useful guide (Figure 8). We placed Ge on 1, 3 and 5% of the sites of si4096, selected by the signed rank of in the K map, subject to a minimum pair separation of 5 Å: the sites with the largest positive (where a heavier atom is predicted to lower most), those with the most negative , or random sites (five sets). At 5% and K, falls by 22.4% for the most positive sites, 17.2 ± 0.2% for random sites and 14.1% for the most negative sites: guided placement lies far outside the spread of the random sets (23 times their standard deviation, from five sets), and the two guided placements span 8 percentage points of . The ordering holds at every concentration and at K (17.2, 14.1 ± 0.2 and 13.0% at 5%). The separation constraint matters: random sets without it, where Ge atoms can cluster, lower by 18.7% at 5%. Thus, for dispersed heavy substituents the map is a useful ranking tool, while for compact clusters only the number of substituted atoms matters.
Figure 8. Exact change of when Ge replaces 1, 3 or 5% of the Si atoms of si4096 (ACE), at sites at least 5 Å apart chosen as the sites of largest positive , at random (mean and standard deviation of five sets), or as the sites of most negative ( K map; mass-only Ge).
An amorphous/crystalline interface
As an application we consider heat flow across an amorphous/crystalline interface: a 2022-atom periodic cell with a crystalline layer ( Å, 650 atoms) between amorphous regions (1372 atoms), so that the cell contains two interfaces. The model is that of Ref. [39], relaxed with GAP in Ref. [17]; relaxing it again with ACE moves atoms by 0.12 Å rms and leaves the crystal essentially unchanged (0.04 Å rms, 0.11 Å at most). The crystal is defect free; 3.6% of the amorphous atoms are three- or five-fold, more than in the bulk models, so this is a model interface rather than an annealed one. The gauge uncertainty of is 1% of its spread at 4 Å (4% at single sites), small here because the spread includes the contrast between crystal and amorphous regions; within the amorphous region more than 4 Å from the interfaces it is 7% at 4 Å, close to the bulk value.
Where heat flows. The in-plane conductivity is W/(m K) and the cross-plane conductivity W/(m K) ( K); directional conductivities are three times the term of Eq. (1). The in-plane local conductivity is about 5 W/(m K) in the crystal (limited by the cell and by ) and 1.0–1.2 W/(m K) in the amorphous region, the bulk a-Si value, with transitions about 5 Å wide (Figure 9(a)). The cross-plane profile of is nearly flat (1.4–1.8 W/(m K)). This is just what one expects for layers conducting in series. In a steady state the same cross-plane flux passes through every plane, so the share of carried by a slab of thickness is close to wherever the resistance lies, up to terms from bonds that cross the faces of the slab. As for resistors in series, which carry the same current, the current map cannot locate the region that limits cross-plane transport. Modal analysis of interface conductance [59] decomposes the conductance over pairs of modes; the question here is different: where a change would alter it.
Figure 9. The a-Si/c-Si interface model (2022 atoms, ACE). (a) Local in-plane and cross-plane conductivity ( times the mean in 2.3 Å slices; K); shaded: crystal. (b) Mass sensitivity of the cross-plane and in-plane conductivity ( K). (c) Exact change of when all crystal or all amorphous masses are scaled by (solid), against the linear prediction from the map (dashed).
Where to change it. The sensitivity map for can: it identifies the region where local perturbations most affect cross-plane conductivity. Its mass sensitivity peaks about 4–6 Å into the amorphous side of both interfaces, is moderate in the amorphous interior, and is small in the crystal (Figure 9(b)). Of the 40 bonds whose stiffening most raises , none lies in the crystal and 25 lie within 8 Å of an interface on the amorphous side (against 11 expected for bonds chosen at random). Low diffusons straddle the interfaces, low-frequency propagons lean toward the amorphous side, and there is no band of interface-localized modes; within each region the rattler statistics of the bulk reappear (Supplemental Material, Sec. S16).
Exact tests. We scaled all crystal masses, or all amorphous masses, uniformly and recomputed exactly (Table III). The amorphous side behaves as the map predicts: lighter amorphous atoms raise (by 2.5% for a 5% mass reduction, against 2.7% predicted), and the linear prediction holds to within 20–25% even for a 50% mass increase (−26% against −21%). The crystal response is dominated by second-order terms: is stationary near the natural crystal mass, so changes of either sign lower it, and a heavier crystal does not improve acoustic matching (−12% for 50% heavier). A plausible reason is that the crystal layer has near-degenerate modes: the first-order map is exact, but its range of validity is set by their splittings, which a uniform mass change of a few per cent already exceeds. The crystal is also where a uniform broadening is least physical, since its local conductivity is limited by the cell and by , so the curvature may be in part an artifact of the broadening. The cross-plane conductivity is controlled from the amorphous side, and the response map locates that region. Two caveats apply: this is a model interface with a defective amorphous region, and the crystal is treated at the point of a periodic cell.
| region, | linear | linear | ||
| crystal, 0.95 | −0.36% | −0.12% | +0.97% | +1.11% |
| crystal, 1.05 | −0.12% | +0.12% | −1.21% | −1.06% |
| crystal, 1.2 | −2.86% | +0.44% | −5.69% | −3.96% |
| crystal, 1.5 | −12.3% | +0.98% | −14.6% | −8.80% |
| amorphous, 0.95 | +2.45% | +2.68% | +1.61% | +1.77% |
| amorphous, 1.05 | −2.71% | −2.55% | −1.80% | −1.68% |
| amorphous, 1.2 | −11.3% | −9.51% | −7.22% | −6.29% |
| amorphous, 1.5 | −26.1% | −21.2% | −15.5% | −14.0% |
Table 3. Exact change of and ( K) of the interface model when all crystal or all amorphous masses are scaled by , with the linear prediction from the direction-resolved mass map.
Discussion
Two complementary local views. The sensitivity maps avoid the partition altogether, in the way forces do for the energy: the energy of an atom depends on the gauge, but the derivative of the total energy with respect to its position, the Hellmann–Feynman force, is unique. Here the share of belonging to an atom is ambiguous, but the derivative of with respect to its mass or its bonds is not; “alchemical” derivatives of the energy with respect to nuclear charge follow the same logic [60]. Differentiable transport exists at the continuum level, through adjoint and automatic-differentiation solvers of the phonon Boltzmann equation [61, 62], and bond-level response has been used to tune elastic moduli [63]; interface conductance modal analysis [59] resolves the conductance of an interface over pairs of modes. To our knowledge, none of these gives the exact response of an atomistic thermal conductivity to local masses and bonds. The heat-flow map answers “where does heat flow?”: through regions of small vibrational amplitude, that is, of low local compliance, whose bonds are on average slightly longer; the low compliance is a property of the network, not of its individual bonds. This description is robust at the 4 Å scale despite the gauge freedom of the local energy and across the tested interatomic potentials. The sensitivity map answers “where would a change matter?” Its answer is largely the complementary region: in soft, large-amplitude regions that carry little heat, is large and positive. The exact decomposition, Eq. (9), makes the link explicit. The current term ties positively to ; the rotation term reverses the sign by redistributing sensitivity through the mixing of nearby modes; and the local vibrational amplitude, a low-frequency moment of the same local mode weights, predicts where that redistribution goes.
Which map to use. In our view the two maps answer different questions, at comparable cost when computed densely (Sec. III). The site projection also has an route, through the Chebyshev expansion, for and for any fixed set of sites, which we used for BM10k; for the sensitivities we have not implemented one, but a route exists: is the gradient of a single scalar, so reverse-mode (adjoint) differentiation of the stochastic Chebyshev trace for yields the whole map at a cost comparable to a few evaluations of , with a stochastic error at each site that we have not assessed. Meanwhile , from sparse linear solves, is a cheap proxy. Thus the site projection is the tool for asking where heat flows: screening large models, finding channels and rattlers, and transport along an interface or a filament, where the current flows in parallel. It needs regional averages because of its gauge. The sensitivity maps are the tool for asking what to change: design, single sites, and geometries in which the layers conduct in series, such as the interface of Sec. IV G, where the flux is the same in every layer and the site projection is nearly uniform whatever the resistance. The practice we would recommend is to use both: the site projection first, to see where heat flows, and the sensitivity map where a decision about what to change must be made.
Relation to soft spots and elastic heterogeneity. The large- regions are the vibrational analogue of the soft spots of sheared amorphous packings, identified from quasi-localized low-frequency modes [24, 26], and of the regions of low local modulus that Mizuno, Mossa and Barrat linked to the vibrational anomalies and the thermal conductivity of model glasses [27, 28]. Here the same regions are where little heat flows and where the conductivity is most easily changed. Gautam et al. found, in five amorphous, paracrystalline and polycrystalline silicon models, that sites with large thermal displacements (from molecular dynamics and in the harmonic approximation) have low normalized SPTC, and described them as rattlers [18]. Our rank statistics, block-bootstrap intervals and partial correlations, over six networks and three potentials, confirm this relation and extend it from the heat-flow map to the response of .
Limitations. The calculations are harmonic, at the point, in cells of up to 10 000 atoms. (i) The propagon contribution is cell-size limited and is not the subject of this work; it requires a treatment beyond the harmonic AF picture [7, 8] or modal analysis of molecular dynamics [64, 65]; the diffuson conductivity is converged (0.944–0.963 W/(m K) in the 4096-atom cell and in a 32 768-atom replica; Supplemental Material, Sec. S2). (ii) The maps use a uniform broadening, and the physical lifetimes justify it for . Three-phonon widths from finite-difference cubic force constants (median half-width 0.10 meV at 300 K) give W/(m K) for si4096 and 1.270 W/(m K) for the 32 768-atom replica, within 3% of the K values (1.218 and 1.237). The maps are another matter. Many physical widths fall below the level spacing of a 4096-atom cell; the mass map then becomes large and noisy (the Euler rule still closes, to 10–7, but the largest is 7 times its value at K and 19% of the sites exceed ten times the spread of the K map), and the anti-correlation with is lost: Spearman −0.05 at 4 Å with raw widths and −0.03 to −0.17 with widths floored at one to four local level spacings, against −0.41 at K (Supplemental Material, Table S2). The anti-correlation returns steadily as the floor rises, so in our view this is a finite-cell effect, not physics, and it is reasonable to conjecture that a map with physical widths needs cells in which the widths exceed the level spacing. Until then, region-averaged maps at –10 K are the practical choice, and the anti-correlation is a property of those maps. (iii) Mass maps from two potentials agree well above chance but only moderately: between ACE and GAP-18 for si4096 the Spearman correlation is 0.47 at single sites and 0.51 at 4 Å at K, and 0.63 and 0.71 at 10 K (shift nulls 0.05, 0.15, 0.03 and 0.08), against 0.89–0.94 for . The heat-flow map is thus more transferable between potentials than the response map. (iv) A per-atom Green–Kubo check with fully anharmonic molecular dynamics (si512) support the rattler picture qualitatively (Supplemental Material, Sec. S17 and Table S11).
Bonds. The bond sensitivities (Supplemental Material, Table S12) are dominated by bending for all three potentials, and stiffening the bonds around heat-carrying atoms raises in every case. Their dependence on bond length is potential specific: for ACE and GAP-18 long bonds are the most sensitive, through stretching, whereas the Keating form, whose stretching constant grows with bond length, reverses the trend.
Design. For isotopes and small changes of bond stiffness the linear map is quantitative. For heavy substituents it is a ranking tool when the substituents are dispersed, and it does not distinguish the compact-cluster placements tested here. The connection to also gives a cheap screening variable: needs only sparse linear solves and predicts the sensitivity map with Spearman 0.3–0.9 at 4 Å.
Conclusions
The site-projected thermal conductivity of a-Si depends moderately on the choice of local energy at single sites but is robust over regions and across the tested potentials. Its gauge-free counterpart, the exact response of to local masses and bonds, obeys an Euler sum rule and contains the heat-flow map as one of three exact terms. With ACE, GAP-18 and Keating force constants the two maps are anti-correlated; the anti-correlation strengthens with cell size and when longer-wavelength modes are included, and at room temperature the local vibrational amplitude largely explains it: soft regions carry little heat, and there a heavier atom lowers most, through the mixing of modes about 1–2 meV apart. In paracrystalline silicon the grains conduct more heat per atom and the matrix controls the response. The response map guides the placement of dispersed substituents and locates the region, on the amorphous side of an amorphous/crystalline interface, that controls cross-plane transport.
More broadly, the gauge question that has shadowed local descriptions of heat transport has a constructive answer: ask how responds, not how it is shared. The construction needs only a harmonic network and applies to any transport coefficient written as a trace over eigenstates, the Kubo–Greenwood electrical conductivity among them; we expect the distinction between where a current flows and where it can be changed to be general. Natural applications are the amorphous dielectrics and interfaces of nanoscale devices, and thermoelectrics, where the electrical analogue would locate the regions that control both currents.
Acknowledgments
We thank Dr. N. Bernstein for providing the amorphous/crystalline interface model, and L. A. M. Rosset and V. L. Deringer for the paracrystalline models R1–R3 and the SiH-ACE-25 potential. We thank A. Gautam for his earlier work on site-projected thermal conductivity and on rattling sites in silicon networks, on which this study builds. We acknowledge support from the U.S. National Science Foundation (Grant No. MRI-2320493) and the U.S. Office of Naval Research (Grant No. N00014-23-1-2773). C.U. acknowledges support from the Laboratory Directed Research and Development program of Los Alamos National Laboratory under the Director’s Postdoctoral Fellowship Program (Project No. 20240877PRD4). Los Alamos National Laboratory is operated by Triad National Security, LLC, for the National Nuclear Security Administration of the U.S. Department of Energy (Contract No. 89233218CNA000001). Computations were performed on a single workstation. Calculations, analysis code and drafting of this manuscript were assisted by Claude Code (Anthropic), an AI system, in accordance with the APS policy on the use of AI tools; the authors take full responsibility for the content.
Data availability. The structures, dynamical matrices, sensitivity maps and analysis scripts supporting this study are openly available in Zenodo at https://doi.org/10.5281/zenodo.23620133.
Remarks on the original formulation
The following points correct or qualify Refs. [16, 17]; none changes their conclusions about .
Asymmetric weight. The AF weight is not symmetric in . This leaves unchanged but makes site values depend on whether a bond current is anchored at its row or its column atom; the symmetrized weight of Sec. II removes the dependence (site values change by about 0.5%).
Locality. The decay of the site-pair matrix [16, 17, 19], whose row sums give , is inherited largely from the range of the force constants and is a power law, so the analogy with electronic nearsightedness is weaker than stated. At small broadening the response is non-local over several nanometres (Supplemental Material, Sec. S6).
Negative site values at localized modes are at least partly a property of the gauge-dependent partition, not numerical artefacts.
Gauge. The non-uniqueness noted in Ref. [17] is quantified in Secs. II B and IV A and complemented by the gauge-free maps of this work.
The Euler rule
Under for all , and while the eigenvectors are unchanged, so . The AF weight has the form
with the mode heat capacity, a dimensionless ratio and a normalized Lorentzian. If and are scaled with the frequencies, and , then and are unchanged, and , so and . Differentiating at gives
which with is Eq. (6). Here the temperature derivative acts through the heat capacities only. The analogous rule for bonds follows from .
Derivation of the gradient
Write Eq. (1) as with and , and consider a symmetric perturbation . Three quantities change: the operators , the eigenvalues, and the eigenvectors.
Explicit part. Since is linear in , at fixed , and , which gives the second term of Eq. (7).
Eigenvalues. First-order perturbation theory gives , so the eigenvalue part is , with the derivative at fixed eigenvectors, acting on only.
Eigenvectors. For non-degenerate ,
with . Then
Inserting into , using the symmetry of and the antisymmetry of (so that ) and collecting the coefficient of , gives with and of Eq. (8). is antisymmetric, so is symmetric, and Eq. (7) follows. This is the Daleckii–Krein formula [32, 33]: the divided differences in tend to as , which gives the near-degenerate form quoted in the main text.
Mass sensitivity. With , where projects on the coordinates of atom , . The explicit part contracted with gives
and for the spectral part,
The diagonal () gives ; symmetrizing the off-diagonal sum over gives
Hence , and is Eq. (9).
References
Supplemental Material for “Where heat flows and where to change it: gauge-free sensitivity maps of thermal conductivity in disordered solids”
D. A. Drabold; C. Ugwumadu
SM.pdf
This Supplemental Material collects the checks, the secondary results and the “nuts and bolts” that support the main text but would interrupt it. The sections appear in the order in which the main text first calls on them. Most are short, and each says at the end what we take from it. A reader who wants only the physics may skip Secs. S1, S3, S4 and § 6.0.6, which concern definitions, numerics and models. Validation tests are collected in Sec. S3.
Bond compensation
Scaling one bond of a many-body potential raises a small but real question of definition: the off-diagonal blocks are not symmetric, so no change of a single bond keeps both the symmetry of the dynamical matrix and the acoustic sum rule exactly (Sec. II of the main text). Table S1 compares our choice with the obvious alternative.
The two definitions give nearly the same bond maps, but only ours satisfies the Euler rule. Thus the choice matters for the sum rule and hardly at all for the physics.
Broadening, widths and heat capacities
The broadening stands in for the true widths of the modes, and it is fair to ask what happens if one uses the true widths instead. Table S2 gives the answer for si4096.
| widths | (W/(m K)) | Euler | sd() | max | Spearman site / 4 Å | Pearson site / 4 Å | > 10 sd (%) |
| K | 1.218 | 0.024 | −0.31 / −0.41 | −0.31 / −0.44 | 0.0 | ||
| K* | 1.067 | 0.003 | −0.34 / −0.48 | −0.33 / −0.51 | 0.0 | ||
| K | 0.948 | 0.002 | −0.31 / −0.41 | −0.29 / −0.43 | 0.0 | ||
| K* | 0.844 | 0.001 | −0.16 / −0.19 | −0.14 / −0.20 | 0.0 | ||
| K, classical C* | 1.331 | 0.025 | −0.26 / −0.38 | −0.27 / −0.40 | 0.0 | ||
| anharmonic, raw | 1.205 | 0.182 | −0.04 / −0.05 | +0.01 / −0.02 | 19.0 | ||
| floor | 1.262 | 0.098 | −0.07 / −0.03 | −0.06 / −0.03 | 0.3 | ||
| floor | 1.241 | 0.058 | −0.10 / −0.09 | −0.12 / −0.09 | 0.1 | ||
| floor | 1.199 | 0.015 | −0.14 / −0.17 | −0.15 / −0.16 | 0.1 |
Table S2. TABLE S2. si4096 (ACE): conductivity, Euler closure and mass-map statistics for different broadenings and width models ( K, quantum heat capacities unless noted). Floor: widths with the local level spacing, the mean spacing of the 20 modes nearest in frequency (ten on each side in the sorted spectrum). Anharmonic widths are extrapolated as below the lowest resolved mode; with any floor the floor sets the widths there, so the floored rows do not depend on the extrapolation. Correlations are with from the same run (rows marked *: at K); the 95% block-bootstrap intervals of the Spearman values are ±0.03–0.05 at single sites and ±0.05–0.09 at 4 Å. Last column: percentage of sites with above ten times sd() of the K map. The raw-width (1.205) comes from the map code; the conductivity code used below gives 1.208.
Three-phonon widths were obtained from finite-difference cubic force constants of si512 and Fermi’s golden rule (median HWHM 0.10 meV at 300 K), extrapolated as below the lowest resolved mode. Without a floor, many widths fall below the local level spacing, and the mass map becomes large and noisy: the Euler rule still closes, but sd is 9 times its K value and 19% of the sites lie beyond ten times that spread. A floor of a few times the spacing shrinks the map and brings back part of the anti-correlation with . The lesson is simple: a mode narrower than the level spacing of the cell is not resolved by the cell, and the map notices.
Diffusons and propagons. A 32 768-atom replica ( copies of si4096, eight real -points) with force constants truncated at 11 Å and the sum rule restored gives the diffuson conductivity (10–56.5 meV) 0.944 W/(m K), compared with 0.948–0.963 in the 4096-atom cell for all width models. The diffuson part is converged. The propagon part (< 10 meV) is not. With physical widths (three-phonon widths extrapolated as below the lowest resolved mode) it is 0.241 W/(m K) (4096; equal, by coincidence, to the K value) and 0.306 W/(m K) (replica), and total is 1.208 and 1.270 W/(m K), within 3% of the K values (1.218, 1.237). Three-phonon widths are resolved only down to the lowest mode of si512 (about 7.5 meV) and must be extrapolated below it. An earlier extrapolation held them constant there, which over-damps the lowest modes (the physical widths fall roughly as ). In a periodic cell larger widths raise the AF propagon value, so that extrapolation gave propagon contributions that are too high, 0.310 (4096) and 0.402 W/(m K) (replica); it is superseded by the extrapolation quoted above. In a periodic cell the AF propagon conductivity decreases when the widths fall below the level spacing, so finite cells cannot represent long-mean-free-path propagons. Thus the diffusions, which dominate and the maps, are in good hands; the propagons are not, and we make no claim about them.
Validation
An exact gradient is only as good as its implementation. The checks below all pass by wide margins.
Analytic gradient against central finite differences from an independent code (si512, ACE, K): median relative error over all 512 mass sensitivities, and – over 48 bond sensitivities. The bond check requires the symmetrized self-block compensation , because the off-diagonal blocks of a many-body potential are not symmetric.
Euler rule, Eq. (6) of the main text, closes to in every -broadened run (all models, potentials, temperatures and heat-capacity models), and to with physical widths, raw or floored (Table S2).
Decomposition: to round-off; the three terms reconstruct to ; in every run; to .
The compact pair formula for [Eq. (10) of the main text], with built directly from the eigenvectors, reproduces the code to maximum difference of W/(m K), against a largest of W/(m K) (si512, K). is antisymmetric to round-off.
Conjugate-gradient against the dense pseudo-inverse of the same sparse matrix: maximum relative difference (si512) and (si4096); truncating the matrix at 11 Å changes by at most 0.4%.
Potentials and structures
We use three descriptions of the forces on purpose: two machine-learned potentials of near first-principles quality, and the Keating model, which is simple, transparent and in some respects wrong. A result that survives all three is unlikely to be an artifact of one. Table S3 lists the models.
TABLE S3. Models. Coordination defects counted with a 2.8 Å cutoff after relaxation with the potential used for the maps. A 216-site diamond cell ( Å) with 20 sites removed at random and Gaussian displacements of 0.3 Å, relaxed with the original Stillinger–Weber parameters; used only for the exact gauge comparisons of Sec. II B of the main text.
| model | relaxation | 3/5/6-fold | defects | |
| si4096 (DTW [1]), ACE | 4096 | fixed cell (43.42 Å) | 5/23/0 | 0.68% |
| si4096, Keating | 4096 | fixed cell, bond topology of the input | (all four-bonded) | — |
| si4096, GAP-18 | 4096 | fixed cell (43.42 Å) | 2/38/1 | 1.0% |
| si512, ACE and GAP-18 | 512 | fixed cell | 0/0/0 | 0 |
| BM10k relaxed | 10 000 | ACE, zero pressure | 16/36/4 | 0.56% |
| BM10k, 800 K anneal | 10 000 | ACE MD, then zero pressure | 51/73/0 | 1.24% |
| BM10k, 1000 K anneal | 10 000 | ACE MD, then zero pressure | 63/87/0 (one 2-fold) | 1.51% |
| R1 / R2 / R3 | 1000 | ACE, fixed cell | 10/0/0, 5/3/0, 3/3/0 | 1.0, 0.8, 0.6% |
| interface | 2022 | GAP, then ACE | 19/31/0 (all amorphous) | 2.5% (3.6% of a-Si) |
| SW gauge tests | 196 | SW, fixed cell (16.28 Å) | 33/1/0 (one 2-fold) | 18% |
Table S3.
Keating. with eV/Å (47.5 N/m), (13.5 N/m), Å, on the bond list of the input network. With a 2.8 Å distance criterion the relaxed Keating structure shows 16 five- and 2 six-fold atoms, but these are geometric contacts; the energy keeps the four-bonded topology. The stretching constant of this form, , rises with bond length, whereas Si bonds soften when stretched; this is why the bond sensitivity of Keating has the opposite length dependence to ACE and GAP-18 (Sec. S18).
GAP-18. The gp_iter6_sparse9k.xml potential [2], run in LAMMPS through QUIP. Diamond Si has its minimum at Å; per-atom energies sum to the total. For si512 GAP-18 gives W/(m K) (ACE 1.091) at K and an optic peak at 57.5 meV (ACE 59.8 meV). For si4096 the force constants come from sparse finite differences with graph colouring, truncated at 10 Å, symmetrized and with the acoustic sum rule restored; applied to si512, the same procedure reproduces the dense matrix to (largest relative difference of a block; asymmetry before symmetrization ). It gives W/(m K) (ACE 1.218) at K and 0.867 (ACE 0.948) at 10 K.
Vibrational densities of states
Experimental comparison. The measured spectrum was digitized from Figure 3(a) of Ref. [3] (58 data points, the crosses; the energy axis was calibrated on the five energy values marking the cuts in that figure, with residuals below 0.13 meV). The reference does not report the instrumental resolution (time-of-flight data on IN4 with incident energy 12 meV, so that most of the spectrum is measured in neutron energy gain, where the resolution varies with energy), so we broadened the calculations with a Gaussian whose width minimizes the rms difference from the data; this constant width is an effective resolution: FWHM 4.5 meV for ACE (si4096) and 3.5 meV for GAP-18 (si512), with misfits of 9 and 10% of the peak height. Panel (a) uses 4 meV for all models. ACE and GAP-18 reproduce the TA-like band, the minimum near 30 meV and the TO-like band; the paracrystalline models R1–R3 have sharper peaks than the amorphous measurement. The Keating spectrum has its TO-like peak near 48 meV (ACE: 59.6 meV), about 10 meV below the measured band, and its TA-like maximum near 28 meV (ACE: 21.5 meV); its TA-like band extends too high and fills the minimum near 30 meV ( at 30 meV is 0.99 of the TA-like maximum, against 0.22 for ACE). Kamitakahara et al. found the TA-like band of their own Keating model too high as well (26 against 20 meV).
Local spectra. The weight below 12.4 meV is 0.034 averaged over all sites, 0.027 and 0.051 for the 10% of sites with the largest and smallest , and 0.054 and 0.027 for the 10% with the largest and smallest . High- sites also have a slightly softer optic band (mean frequency 39.8 against 40.4 meV for low- sites; weight above 50 meV 0.38 against 0.41), consistent with their longer bonds. Thus the regions that carry heat are, in a modest way, spectrally different from those that do not, which is the local softness of the main text seen from the side of the spectrum.
evaluation
Dense diagonalization stops at a few thousand atoms. For larger systems one needs a route that scales linearly with , and this section describes the one we used for BM10k. Writing Eq. (3) of the main text as the diagonal of , with Chebyshev polynomials of the rescaled , two-dimensional Chebyshev coefficients of the AF weight and no damping kernel, reproduces with a relative error of 0.03% and to 0.9% of its spread with (si4096, K); the three acoustic zero modes are projected out exactly. is computed as a stochastic trace (32 random-sign vectors). Per-site would need random vectors and is computed instead with unit vectors on each requested site. The cost is therefore for and for a fixed number of sites (each site costs ), but for the complete map of all sites. The GPU implementation (CuPy) has memory linear in and reproduces exact results at 4096 atoms (: Pearson 0.9999, rms error 1.5% of the spread) and on the 32,768-atom replica ( 1.237 0.008 against 1.237). is not local at small broadening: spheres of 25 Å reproduce 93% of its spatial scale at K but only 40% at K, so spatial partitioning is not a route to large systems. In plain terms: at small broadening the heat flow at a site knows about the structure several nanometres away.
The sparse ACE dynamical matrix of BM10k was obtained from LAMMPS forces with graph colouring: atoms farther apart than Å are displaced in the same configuration, and blocks are kept for Å, with the acoustic sum rule restored on the self blocks. It reproduces the dense matrices of si512 and si4096 to and .
Defects in BM10k. The three- and five-fold sites of BM10k are soft: – Å at three-fold and Å at five-fold sites, against – Å at four-fold sites (), and they carry 13–37% less heat. Restricted to four-fold sites, is still to (Sec. IV C of the main text), so the relation is not carried by the defects.
Local softness as a diagonal of the pseudo-inverse
In the classical harmonic approximation a mass-weighted normal coordinate has , and the Cartesian displacement of coordinate is . Hence , where the sum and the pseudo-inverse exclude the three uniform translations. Summing over the three coordinates of gives Eq. (11) of the main text. To evaluate without eigenvectors we solve , where removes the mass-weighted translations , by conjugate gradients with a block-Jacobi preconditioner (the inverse of each atom’s diagonal block), reapplying at every step so that and . A relative residual of takes about 210 iterations for si4096. The quantum generalization replaces by and requires the spectrum.
Site currents in an arbitrary gauge
With and in , and , written with creation and annihilation operators, the coefficient of is proportional to . Normalizing to the AF form gives Eq. (4) of the main text; the terms and do not contribute at zero frequency. For the AF gauge, is antisymmetric and . The site-gauge map replaces one factor in Eq. (1) of the main text by the current of site alone, built from ; its evaluation with per-atom energies only is described in Sec. S9.
Gauge dependence of the site projection
To measure the gauge dependence of we need the map in a second gauge, and the natural second gauge is the one the potential itself supplies through its atomic energies. Here we describe how we computed it without ever forming the per-atom Hessians. In the site-energy gauge the local energy is with , where is the energy the potential assigns to atom . Then
where is symmetrized, , because only its symmetric part contributes to the contraction with the symmetric , and with , the matrix of Eq. (3) of the main text. is taken from the total (AF) current: the gauge changes the total current only off shell, which costs an error in . is never formed. We diagonalize the small matrix (with dimension equal to three times the number of atoms within the site-energy cutoff of ), and evaluate with , which needs only per-atom energies; the displacement step is Å. Sites farther apart than twice the cutoff are probed in the same configuration. For Stillinger–Weber, where the exact site-gauge result is available, the probed agrees with it to 0.3%.
For si4096 with ACE, the AF and site-energy maps have Pearson and Spearman correlation 0.92 at single sites; 819 of the 1024 top-quartile sites and 828 of the 1024 bottom-quartile sites agree in both gauges. The two maps differ by a uniform offset: in the site-energy gauge is 0.78% lower than in the AF gauge ( W/(m K) per site), the difference of result (i) of the main text. No averaging removes it, and left in it makes the ratio below rise again beyond 6 Å (0.157 at 8 Å, 0.192 at 10 Å). It is a global offset, not a local one, so we quote it once and remove it. The local gauge uncertainty (standard deviation over gauges, offset removed, relative to the spatial spread of the mean) is then 0.227 at single sites, 0.124 at 2.6 Å, 0.095 at 4 Å, 0.077 at 5 Å, 0.070 at 6 Å, 0.055 at 8 Å and 0.046 at 10 Å (Spearman 0.917, 0.975, 0.984, 0.989, 0.990, 0.993, 0.995). Beyond 4 Å it is fitted well by Å/; it halves between 4 and 10 Å, against a factor 2.5 for pure . For the interface model it is 0.04. Thus the local gauge ambiguity falls roughly as beyond about 4 Å ( Å for ACE), as expected for a surface term, and the ranking of sites is nearly gauge-invariant.
Anti-correlation and mediation at K
The anti-correlation and its mediation by local softness do not depend on the choice of broadening. Table S4 repeats the three si4096 rows of Table II of the main text at K; the Keating anti-correlation is, if anything, stronger.
Sign and magnitude of the mass map
Since takes both signs, “anti-correlated with the heat flow” could mean two things: that the sign of the response follows the heat flow, or that its size does. Table S5 separates the two. The sign follows reliably; the size follows only on average and only after spatial averaging.
Decomposition of the covariance and pair-resolved rotation term
The exact decomposition lets us ask which term carries the anti-correlation and, for the rotation term, which pairs of modes carry it. Table S6 gives the answer: the rotation term, and pairs of modes a few apart in the 10–30 meV band. In every separation bin up to the binned rotation term is negative at high- sites (top quartile) and positive at low- sites (bottom quartile).
What sets the width of this “mixing window”? We compared several measures of how quickly correlations between pairs of modes decay with their separation in frequency (Table S7).
The local overlap measures decay smoothly and identically at both broadenings. The current-correlation measures either depend strongly on (, which contains the AF weight) or differ between the potentials without a common scale; they do not explain the window. The rebuilt used for this analysis agrees with the sensitivity code to (near-degenerate pairs omitted). Thus the window is set by where modes live, not by how they carry current.
Local return probability. For the 10–30 meV band we computed the site-averaged return probability of vibrational weight, with , which is the Fourier partner of . Weight leaves its atom quickly: falls to of its initial excess over at 0.146 ps (ACE) and 0.136 ps (Keating), i.e. and 4.8 meV, set by the band width rather than by the window. Between 0.2 and 2 ps, falls with local logarithmic slopes between and (ACE) and oscillates for Keating, so the tail is not the law of three-dimensional diffusion in this cell; the diffusion constants implied by (11–33 Å/ps) are smaller than the AF values by factors of 3–9. AF mode diffusivities (classical heat capacities, K), averaged over 10–30 meV: 94 Å/ps (0.94 mm/s, ACE) and 160 Å/ps (1.60 mm/s, Keating); the classical rebuilt from them, 1.331 and 1.862 W/(m K), reproduces the direct values. We conclude that the time domain is not the place to see the window in a cell of this size; the frequency domain is.
The 32768-atom replica
A larger cell supports longer-wavelength modes, and it is natural to ask whether the results of the main text survive them. Rather than build a new network, we used a replica of si4096, which adds the long waves while keeping the local structure fixed. The replica of si4096 has wave vectors , , at which the Bloch phases are , so each block is a real matrix corresponding to 4096 atoms. The force constants are truncated at 11 Å, with the acoustic sum rule restored; no pair has two periodic images within this range, so the construction is exact. Because is translation invariant on the replica, and are block diagonal in , and every quantity in Eqs. (1)–(10) of the main text is a sum over blocks. In particular, . A block calculation changes atom in every one of its eight images at once, a perturbation with the periodicity of the 4096-atom cell, which is why the problem stays block diagonal: the derivative of for the eight images together is , and since the images are equivalent, the derivative for a single replica atom is one eighth of that, . The same holds for , while is a property of each image. The constant factors do not affect any rank statistic, and the Euler sums are consistent: summing the single-atom derivatives over all 32 768 atoms gives . Each block was run through the same code as a cell. The Euler rule closes to within W/(m K) in all 24 block runs (8 blocks, 3 temperatures), and the block reproduces the 4096-atom cell ( against 0.37831 W/(m K) at 50 K).
The low-temperature residual (replica, 50 K). Splitting by restricting to pairs with both modes below 10 meV (exact, since and its gradient are linear in ; the masked runs close the Euler rule in every block): the propagon-pair part carries of 0.397 W/(m K) and 78% of . Spearman with : (propagon pairs), (rest); partial at fixed : and . Partial of the full and at fixed averaged over 4, 6, 8 and 12 Å: , , , ; over the same radii falls from to . With the quantum amplitude averaged over the eight -blocks, the partial of and is at 50 K, at 100 K and at 300 K, against , and with the classical amplitude; the classical and quantum amplitudes produce similar site rankings (Spearman 0.93, 0.99 and 1.00). By term (replica, 4 Å): at 50 K the rotation term carries 85% of and keeps a partial correlation with of at fixed ; the eigenvalue term carries 25% with a partial of . At 300 K the corresponding values are 108% and (rotation), 25% and (eigenvalue). Thus at room temperature local softness accounts for the anti-correlation, in the replica as in the cell; at low temperature the long waves add a component that it does not capture, and that component lives chiefly in the rotation term.
Joint conditioning on spectral moments. The low- residual of in the replica (4 Å) is at 50 K and at 100 K. Adding the quantum amplitude as a second condition gives and . The two amplitudes are strongly collinear (rank correlation 0.97 at 50 K, 0.996 at 100 K; variance inflation factors 15 and 130), so we test the drop against a null in which the second field is translated by a random vector of the periodic cell, which keeps its distribution and spatial correlation but breaks its alignment with the maps: 200 such surrogates leave the partial at and (). The local moments tell the same story: is the classical amplitude (rank correlation 1.000), and adding gives and ; three and five moments give and at 50 K. Band weights below 10 meV recover less () than band weights over the whole spectrum (). At 300 K the two amplitudes coincide and nothing is gained. Thus the residual depends on the shape of the local spectrum, not only on one of its moments, and a part ( with five moments at 50 K) is not captured by the local spectrum at all.
Paracrystalline models
Paracrystalline silicon, with small grains embedded in an amorphous matrix, is a natural test of the picture: the grains are stiff and the matrix is soft. Atoms were classified with the “identify diamond structure” modifier of OVITO (version 3.16.1); crystalline atoms are those in cubic or hexagonal diamond environments. The classification gives 100% for perfect diamond and 0.1% (si4096) and 0.07% (BM10k) for the continuous random networks, and 18.1% (R2) and 24.3% (R3), the fractions quoted in Ref. [4]. The signed distance to the grain boundary is the distance from a crystalline atom to the nearest non-crystalline atom (positive), or minus the distance from a non-crystalline atom to the nearest crystalline one; no atom lies within about a bond length of the other region, hence the gap at zero in Figure S2.
Exact test of the grain-interior response. Scaling the masses of all grain-interior atoms ( Å; 59 in R2, 97 in R3) by and recomputing exactly ( / 10 K), against the linear prediction in parentheses: R2, : −0.47 (−0.33) / −0.18 (−0.09)%; : +0.19 (+0.32) / −0.01 (+0.08)%; : −0.73 (+1.18) / −0.99 (+0.30)%. R3, : −0.72 (−0.54) / −0.22 (−0.07)%; : +0.34 (+0.52) / −0.07 (+0.07)%; : −0.51 (+1.93) / −1.66 (+0.26)%. The response is dominated by second-order terms beyond a few per cent, as for the crystal layer of the interface model. In other words, the linear map is a reliable guide to small changes of the grains, and only a rough one to large changes.
Design tests
A sensitivity map is a first-order statement, and a designer will want to know how far it can be trusted for finite changes. Here are the details of the exact tests summarized in the main text. All tests use si4096 (ACE), exact from one diagonalization per perturbed structure, K, quantum heat capacities. Linear predictions: masses ; bonds over bonds with both atoms in the region.
Regions. We changed compact regions around five centres per class, with radii of 2.6, 4 and 6 Å (about 5, 14 and 46 atoms). The ratio of the exact to the linear change for regions centred on the most sensitive sites is: 1.10, 1.10, 1.21; bonds ±5% (4 Å) 0.93 and 1.06; bonds ±20% 0.67 and 1.17; Ge ( K) 2.3, 3.0, 4.0. For Ge at K, for top / random / bottom centres is −0.41 / −0.42 / −0.40% (5 atoms), −0.98 / −0.99 / −0.93% (14 atoms) and −2.33/−2.38/−2.42% (46 atoms). Thus isotopes and modest bond changes are nearly linear, while a compact cluster of heavy atoms is not, and does not care where it sits.
Pairs. For pairs of Ge atoms we define the interaction as the exact change for the pair minus the sum of the two single changes, relative to that sum. si4096: 6.5–8.7% at 2.35 Å, 1.5–1.9% at 5 Å, 0.5–2.0% at 10 Å (45 evaluations). In si512 (600 evaluations) the interaction is nearly independent of distance at K, a small-cell effect. So substituents more than a few Å apart act nearly independently, and Table S10 shows what the map can do with that.
Interface
The interface model has a crystalline layer between two amorphous regions, and one may wonder whether it has vibrations of its own. It does not, in any strong sense. The fraction of modes with 20–80% of their weight in the crystalline layer ( Å, 32% of the atoms), and in parentheses the mean weight in the crystal: below 10 meV 0.42 (0.19); 10–30 meV 0.89 (0.32); 30–45 meV 0.73 (0.27); 45–56 meV 0.69 (0.24); above 56 meV 0.84 (0.42). The mean weight in quasi-localized soft modes (modes below 12.4 meV in the lowest 10% of participation ratio in that band) is 0.004 in the crystal, 0.005 within 4 Å of the interfaces and 0.011 in the amorphous region; the weight below 12.4 meV is 0.048, 0.060 and 0.104. Within regions (single sites, K; interfaces: atoms within 4 Å of or Å), and (amorphous, interfaces, crystal). Thus the soft modes live on the amorphous side, and within each region the rattler statistics of the bulk reappear.
Fully anharmonic check: per-atom Green–Kubo
Everything in the main text is harmonic. A fully anharmonic check is possible in principle, with Green–Kubo molecular dynamics and a per-atom heat current, but it is noisy and expensive; we did it for the 512-atom model only, and read it qualitatively.
Method. The per-atom heat current is , with the centroid form of the many-body virial [5]. We added centroid-stress support to the pair-pair style of LAMMPS (CPU version), accumulating on the atom that owns the energy; the patch agrees with finite differences to about , and equals the total heat flux of LAMMPS exactly. For si512 we ran four independent NVE trajectories of 1 ns (time step 2 fs, after 20 ps NVT) at 300 K and at 125 K. The local map with sums to exactly. The harmonic reference uses classical heat capacities in the ACE energy gauge.
The plateau values of exceed the harmonic AF value of the same cell. We do not read this as an anharmonic enhancement: the plateau was chosen by eye in a noisy integral, as the first broad maximum of the mean over four trajectories (10 ps at 300 K, 20 ps at 125 K), and Table S11 gives the maps at short times and at the plateau so that their dependence on the window can be judged; we have not attempted an objective criterion (at 300 K the integral falls to at 20 ps), and in a 512-atom periodic cell the few low-frequency modes are long lived in molecular dynamics but are represented in the AF value only through the broadening . We use the Green–Kubo runs only for the map comparisons below, not for the magnitude of .
| 300 K | 125 K | |
| at the plateau (W/(m K)) | 1.57 0.09 ( ps) | 1.67 0.07 ( ps) |
| harmonic AF, classical , K | 1.21 | 1.21 |
| at ps | 1.06 0.15 | 1.67 0.07 |
| map reproducibility, 4 Å (split-half Pearson) | 0.84–0.98 | 0.86–0.97 |
| MD map vs harmonic , 4 Å: –2 ps / plateau | −0.43 to −0.52 / +0.45 | −0.41 to −0.50 / +0.45 |
| MD map vs : –2 ps / plateau | +0.26 to +0.30 / −0.29 | +0.23 to +0.25 / −0.26 |
| from MD vs harmonic: Pearson, slope | 0.83, 1.57 | 0.87, 1.39 |
Table S11. Green–Kubo results for si512 (ACE; mean standard error over four trajectories).
At short integration times the local Green–Kubo map follows current fluctuations, which are large at the rattlers; transport information appears near the plateau, later at lower . At 300 K the integral turns down beyond 10 ps because of a noisy tail (). The comparison is qualitative: the 512-atom system and 1 ns trajectories provide limited spatial and temporal sampling for a converged local map. Still, the change of sign of the map between short times and the plateau is just what the rattler picture leads one to expect: rattlers dominate the fluctuations of the current, not its transport.
Bonds
The main text concentrates on masses. The bond map is its companion, and here we ask what kind of bond motion it responds to. Each nearest-neighbour block is split into its central part (stretching) and the remainder (bending and non-central terms), with the same symmetric self-block compensation as ; the two parts add to to . Bending dominates the total bond sensitivity for all potentials. For the two machine-learned
| ACE si512 2 K | ACE si512 10 K | GAP si512 2 K | GAP si512 10 K | ACE si4096 2/10 K | Keating si4096 2 K | |
| Spearman(, length) | +0.07 | +0.45 | +0.16 | +0.43 | +0.30 / +0.47 | −0.25 |
| stretching part vs length | +0.08 | +0.48 | +0.14 | +0.44 | +0.36 / +0.52 | −0.17 |
| bending share of | 79% | 64% | 69% | 52% | 71% / 56% | 85% |
| site stiffness vs (Pearson) | +0.12 | +0.38 | +0.05 | +0.29 | +0.18 / +0.38 | +0.09 |
Table S12. Bond sensitivities: rank correlation with bond length, its stretching part, the bending share of the summed sensitivity, and the correlation of site stiffness with .
potentials the length dependence is driven by stretching of long bonds; Keating reverses it (Sec. S4). Stiffening the bonds of heat-carrying atoms raises for all three potentials. Thus, as one would guess for a tetrahedral network, the angles matter more than the bond lengths.
References
References
- [1]B. R. Djordjević, M. F. Thorpe, and F. Wooten, Computer model of tetrahedral amorphous diamond, Phys. Rev. B 52, 5685 (1995).pubmed.ncbi.nlm.nih.gov/9981754
- [2]A. P. Bartók, J. Kermode, N. Bernstein, and G. Csányi, Machine learning a general-purpose interatomic potential for silicon, Phys. Rev. X 8, 041048 (2018).DOI
- [3]W. A. Kamitakahara, C. M. Soukoulis, H. R. Shanks, U. Buchenau, and G. S. Grest, Vibrational spectrum of amorphous silicon: Experiment and computer simulation, Phys. Rev. B 36, 6539 (1987).pubmed.ncbi.nlm.nih.gov/9942364
- [4]A. Gautam, C. Ugwumadu, L. A. M. Rosset, R. M. Tutchton, S. Nakhmanson, and D. A. Drabold, Lattice dynamics of disordered phases of silicon and spatially resolved computations of thermal conductivity (2026), arXiv:2610.10165 [cond-mat.mtrl-sci].arxiv.org/abs/2610.10165
- [5]Z. Fan, L. F. C. Pereira, H.-Q. Wang, J.-C. Zheng, D. Donadio, and A. Harju, Force and heat current formulas for many-body potentials in molecular dynamics simulations with applications to thermal conductivity calculations, Phys. Rev. B 92, 094301 (2015).DOI







