Skip to content

Long-time dynamics, lightning-fast performance.

DYNAMITE is an efficient solver of DMFT equations with long memory. It features sublinear cost, GPU support and reusable checkpoints.

Install Browse usage Try the first run

Terminal snapshot

Why DYNAMITE

Purpose-built for aging and quench protocols

Non-stationary solvers

Track two-time correlators C(t, t') and responses R(t, t') deep into aging

Interpolated memory

Two-dimensional sparse interpolation drops the asymptotic cost from O(T³) to linear while controlling errors.

Adaptive integrators

RK54 ⇄ SSPRK104 switching, stability-aware steps, and sparsification-aware tolerances ensure stability to late times.

Smart hardware detection

Autoruns on compatible GPU when available, seamlessly falls back to CPU.

Checkpoints + reproducibility

Versioned HDF5 outputs and restartable checkpoints for every trajectory.

DMFT problem setting

We evolve correlation and response functions after a quench under closed dynamical equations of the form

\[ \partial_t C(t,t') = \mathcal{F}[C,R](t,t')\,,\qquad \partial_t R(t,t') = \mathcal{G}[C,R](t,t')\,,\quad t\ge t'\,. \]

The solver implements a numerical renormalization scheme with two-dimensional interpolation, reducing the asymptotic cost from cubic to linear in simulated time while controlling accuracy for aging observables.

When DMFE is the right tool

  • Your DMFT equations close on C and R with memory integrals over the past.
  • You need long-time, high-accuracy trajectories (aging, quenches, or other non-stationary protocols).
  • You rely on reproducible outputs, resumable checkpoints, and high performance.

Model assumptions

  • Single-site, causal effective dynamics laid out on the triangular domain t ≥ t'.
  • History terms expressible as integrals/convolutions evaluable on a 2D interpolation grid.
  • Model-specific closures supplied via an EOM module (see EOMs and observables).

Quickstart

Build

./build.sh

Short quench (L = 512)

./RG-Evo -L 512 -l 0.5 -m 1e4 -D false

Outputs

Versioned HDF5: QKv, QRv, dQKv, dQRv, t1grid, rvec, drvec. Binary + text summaries when HDF5 is unavailable.

Typical workflow

  1. Select the built-in EOMs (currently the mixed spherical p-spin model as in the accompanying paper) and set physical parameters.
  2. Choose an interpolation grid (sizes/layout) that balances accuracy and runtime.
  3. Select an integrator and tolerances; adaptive RK54 is the usual starting point.
  4. Run a short trajectory, inspect stability + error estimates, and adjust grids/tolerances.
  5. Launch production runs with checkpoints so you can resume to extend time horizons.

Outputs, performance, and accuracy

  • Sparse 2D interpolation of history terms yields sublinear cost in simulated time for long runs.
  • Stability-aware integrator switching maintains accuracy near stiff or transient regimes.
  • Recommended starting grids live in concepts/interpolation-grids.md; refine per observable tolerances.
  • CPU path is the reference; enable GPU when available for speed-ups after validating on your model.

Cite DYNAMITE & get help

  • Software: see Reference → Cite (powered by CITATION.cff).
  • Method paper: J. Lang, S. Sachdev, S. Diehl, “Numerical renormalization of glassy dynamics,” Phys. Rev. Lett. 135, 247101 (2025), doi:10.1103/z64g-nqs6.
  • License: Apache-2.0 (see LICENSE).
  • Support: follow Testing for issue templates or open a GitHub issue with your build info (compiler/CUDA, commit hash, minimal input).