Supplementary Figure 1: Errors in vibrational energy due to discrete sampling.

point phonon modes become a better approximation of the full vibrational spectrum as

the computational unit cell becomes larger. Here a 6-atom cell and a 24-atom supercell-

based free energy difference are compared to a 54-atom supercell as a reference case.

From this plot, it is evident that the 6-atom vibrations already predict a fairly accurate

vibrational free energy difference with reference to the more finely sampled 54-atom

spectrum. For each phase, we construct rectangular 6-atom [1x1], 24-atom [2x2] and a

54-atom [3x3] supercells with commensurately coarsened k-point grids in the electronic

Brillouin zone. These cells have different sets of lattice constants, each derived from the

equilibrium lattice constant of the phase the cell contains. The 1T’-2H’ vibrational free

energy difference is computed as a function of temperature for the 1x1, 2x2 and 3x3

supercells based on 15, 69 and 159 phonon mode frequencies, respectively. The error

made by using only 15 frequencies compared to 159 is about 2.5 meV at room

temperature. This can be compared with the 1T’-2H’ energy difference of 43 meV.

1
 Supplementary Figure 2: An alternative to Lagrange interpolation. Rather

than apply full two-dimensional interpolation as before, DFT energies can be fitted in a

more conservative procedure. Starting from the first two plots, 2H and 1T’ Helmholtz

free energies from PBE/HSE06 calculations are fitted to cubic polynomials ( )

(shown as solid lines) as a function a for each discrete row of b-values. The minima of

each fitted curve (empty squares) is fed back into a cubic fit for both phases, this time as

a function of b. Common tangents and transition strains in the uniaxial-load case can thus

be determined without two-dimensional Lagrange interpolation.

2
Supplementary Figure 3: Estimate of sampling and interpolation errors. Lagrange-

interpolated results shown in the main paper are compared to two numerical alternatives:

one-dimensional fitting on a ‘fine’ 10x10 grid, and cubic one-dimensional fitting on a

‘coarse’ 7x7 grid. The choice between these three avenues for data refinement and

numerical processing does not significantly affect the predicted transition strains beyond

which 1T’ is stable under uniaxial loading. In the case of the HSE06 functional, the

predicted strains at room temperature show a spread of 0.3 percentage points. The PBE

functional predicts higher strains in general, and these have a spread of 0.5 percentage

points. While some numerical discrepancies are observed, these results indicate that the

three chosen methods closely approximate the result that would be predicted in the

limiting case of an infinitely fine grid. The temperature-dependent trends are the same

regardless of which underlying data grid or interpolation/fitting is chosen.

3
Supplementary Figure 4: Phase transitions and coexistence under macroscopic

uniaxial strain. In addition to uniaxial load as a thermodynamic constraint, a scenario

where both phases’ lattice constant a is fixed to some value can also be envisioned. In

this case, the forces along the a-axis are not necessarily zero (as in the uniaxial-load

case). The refined 10x10 grid of DFT energies and phonons can be easily adapted to this

constraint by ignoring the minimum of all ( ) -curves and simply selecting a

particular value for a for further fitting into a cubic ( ) master curve. The ( ) energy

curves are constrained to follow a path that does not change in a. The fitted ( )-

curves are all sampled at a = 3.550 Å (represented by empty squares).

4
Supplementary Figure 5: summary of uniaxial-load and macroscopic uniaxial strain

transition strains. These results derive from cubic one-dimensional fitting of the fine

10x10 data grid. Both PBE and HSE06 functionals are used, along with three different

treatments of vibrational free energy (none, zero-point and 300 K).

5
Supplementary Table 1: PBE-calculated equilibrium lattice constants and

enthalpies at zero stress. This table lists the three phases’ (2H, 1T’, 1T) equilibrium

energy differences and lattice constants, as calculated using PBE-DFT and the Lagrange

interpolation scheme described in the main paper. The enthalpy (equivalent to energy at

zero stress) of the lowest-enthalpy phase is marked in boldface.

( ) (eV per MX2) Lattice constants [a, b] (Å)

Material 2H 1T’ 1T 2H 1T’ 1T
MoS2 0 0.55 0.84 [3.183, 5.513] [3.179, 5.717] [3.176, 5.500]
MoSe2 0 0.33 0.71 [3.318, 5.747] [3.280, 5.971] [3.278, 5.674]
MoTe2 0 0.04 0.52 [3.550, 6.149] [3.455, 6.380] [3.493, 6.054]
WS2 0 0.54 0.89 [3.182, 5.511] [3.197, 5.707] [3.197, 5.511]
WSe2 0 0.27 0.78 [3.315, 5.744] [3.300, 5.944] [3.298, 5.692]
WTe2 0 -0.09 0.57 [3.552, 6.154] [3.491, 6.320] [3.491, 6.320]