phonon — GPU-accelerated phonon spectra (MLIP benchmark + DFT acceleration)

Goal: compute phonon spectra ≥50× faster than conventional DFPT/DFT, with controlled accuracy, toward a publishable high-throughput phonon framework.

Two research lines share one backend-agnostic phonon engine (src/phonon_accel/phonons.py):

Line	What	Speed	The science
A — MLIP benchmark + fine-tune	foundation MLIPs (MACE, MatterSim, SevenNet, ORB, eSEN, CHGNet) compute forces via phonopy	50–1000× "for free"	accuracy: fix softening / imaginary modes / ASR violations by fine-tuning
B — GPU-DFT workflow acceleration	Quantum ESPRESSO GPU finite-displacement, staged	~50× (B1), >100× (B2 stretch)	GPU SCF × cross-displacement density reuse × symmetry/non-diagonal supercells

Key reframe from the literature survey: pure GPU porting of DFT only gives ~5–15×; the 50×+ target is reached by MLIP surrogates (Line A) and by workflow-level acceleration of the finite-displacement DFT pipeline (Line B).

Status

• M0 — env + unified pipeline validated on Si (scripts/smoke_si.py)
• A2 — genuine DFPT reference (MDR, 10,034 materials) wired in; full-band benchmark harness + first real results (MatterSim, 14 crystals)
• A2 — scale to more models (own envs) + larger MDR sample
• M0b — remote GPU 100.105.21.7 probe (currently unreachable from dev box)
• B1 — QE-GPU workflow acceleration
• A3 — fine-tuning to fix failure modes

First results (Line A). Full-band MLIP-vs-DFPT (MDR/PhononDB reference):

model	freq MAE	mean softening	dyn. stable
MatterSim-v1	0.58 THz (14 crystals)	−7.2%	14/14
MACE-MP-0 (2023)	~1–4 THz (softening-dominated)	−26%	—

MatterSim tracks DFPT acoustic branches almost exactly; optical branches mildly softened. MACE-MP-0 softens ~4× more — the gap Line A's fine-tuning targets. See results/dfpt_mattersim_curated.csv, results/figures/, docs/findings.md.

Environments (per-model, to avoid e3nn conflicts)

• phonon — MatterSim + core stack (workhorse / fine-tuning)
• phonon-mace — MACE (pinned e3nn==0.4.4), isolated baseline

The pipeline code is env-agnostic (lazy model import); run the same scripts under whichever env has the target model.

Layout

src/phonon_accel/
  phonons.py      # core: structure -> displacements -> forces -> FC -> bands/DOS/thermal
  mlip_calc.py    # unified ASE-calculator factory for foundation MLIPs (lazy imports)
  structures.py   # built-in cells (Si, ...), Materials Project, file, relax()
  metrics.py      # freq MAE/RMSE, imaginary-mode count, ASR residual, thermal errors
  benchmark.py    # Line A: model x material -> metrics table
scripts/          # smoke_si.py, run_benchmark.py

Quickstart

conda env create -f environment.yml          # or: conda install -n phonon python=3.11 && pip install ...
conda run -n phonon python scripts/smoke_si.py --model mace --device cpu

On the GPU box use --device cuda (float64 is required for force constants; Apple MPS does not support float64, so the local dev box runs on CPU).

See docs//the plan file for the full research design and milestones.