Before any thermal conductivity number: the phonon stability gate, and the inputs to test it on

Our public phonon dispersion route has a failure mode worth putting on the record. Fed a free-standing MoS2 monolayer, it returns deep spurious imaginary branches under two independent machine-learned potentials. A thermal conductivity computed downstream of a dispersion like that would inherit the error without a word of complaint.

Cem Sevik's group at Antwerp benchmarks thermal conductivity limits in MoS2 and MoSe2 (arXiv:2509.13798), and their protocol checks dynamical stability on the full grid, with the phonon density of states, before any conductivity number is computed. He has offered to share the settings. This post pins down our half of that exchange so the check can be encoded as a fixed gate on the route. Everything below is already fixed except two numbers that come from their protocol.

The failing input

2H-MoS2 monolayer control (literature PBE geometry): 2H-MoS2 monolayer, a = 3.160 Å, Mo–S 2.41 Å from the literature PBE geometry, 18.5 Å of vacuum in a 22 Å cell.

The run settings

Finite displacement, supercell [3, 3, 3], Δ = 0.01 Å. Two potentials: Orb v3 conservative inf MPA and MACE-MPA medium. Dispersion along the route's standard high-symmetry path, with the element-projected phonon density of states.

What the runs produced

• Literature geometry, Orb v3: minimum frequency −35.06 THz, 1835 imaginary modes on the sampled grid. Dispersion, run receipt

Two boring explanations are already ruled out. The same failure appears for the input written in standard P-6m2 symmetry style, so it is not the P1-style CIF artifact we caught on this route earlier. Relaxing the cell first does not rescue it either, so it is not a strain artifact of the starting geometry.

What the runs do not settle is where the failure lives. Both potentials fail, while the same monolayer is dynamically stable in the DFT literature that this benchmark builds on. The shared cause could sit in the route's 2D handling (vacuum, periodic images, force-constant interpolation) rather than in either potential. That alternative is still open.

The clean control

SnP₃ bulk, R-3m, COD 1527463 with the same settings: no imaginary modes, minimum frequency −0.00 THz. Run receipt

The gate, as we propose to encode it

Before any conductivity call, the route would compute the full-grid frequencies and the density of states, and refuse to return a conductivity number if any mode falls below the imaginary-mode tolerance. Two values come from the Sevik protocol and are deliberately left blank here: the q-grid their stability check runs on, and where they draw the line on imaginary modes.

The pair above is the test for the gate itself. A setting that passes the failing monolayer, or flags the clean control, is mis-set.

Next step is small: their two numbers, then we run the pair on the route and publish the receipts with the protocol credited.