# Restart a simulation

**You will learn:** how to resume an ACEMD simulation from its checkpoint after a crash, time-limit, or planned stop.

**Prerequisites:**

- A previous run in the directory that produced `restart.chk`.
- The same GPU **model** that produced the checkpoint (an `RTX 4090` checkpoint won't restart on a different card).

## Setup

ACEMD writes `restart.chk` every `trajectoryperiod` steps regardless of whether you plan to restart. To pick up where it left off, add `restart: true` to the input file:

```{code-block} yaml
:caption: input.yaml

structure: dhfr.psf
parameters: dhfr.prm
coordinates: dhfr.pdb
boxsize: [62.23, 62.23, 62.23]
thermostat: true
run: 100ns
restart: true
```

Then re-run:

```bash
acemd
```

## What the restart restores

`restart: true` overrides the on-disk initial state with the checkpoint contents:

- Atomic coordinates.
- Atomic velocities.
- Periodic box vectors.
- Thermostat and barostat state.

It also **skips minimization** even if `minimize: > 0` is set.

## What the restart does **not** do

- It does **not** roll back the trajectory file. New frames are appended to `output.xtc` from the checkpoint's step onward.
- It does **not** re-run already-completed steps. `run` is a **total** target step count, not an additional duration: with `run: 100ns` and a checkpoint at 60 ns, the restart adds 40 ns to reach the 100 ns total. If the checkpoint is already at or past `run`, the simulation immediately exits with nothing to do — bump `run` higher to extend a finished trajectory.

## Extend a finished simulation

A trajectory that reached its original `run` target can be extended in place: bump `run` to the new total and restart. Because `run` is the **total** step count, set it to the new endpoint, not the additional duration:

```{code-block} yaml
:caption: input.yaml

# Original run hit 100ns; extend to 200ns total
run: 200ns
restart: true
```

```bash
acemd
```

ACEMD picks up the checkpoint at the end of the first 100 ns and runs the integrator forward to 200 ns total. New frames append to `output.xtc`.

## Continue without a checkpoint

If `restart.chk` is missing or unusable (different GPU model, ACEMD version changed, accidentally deleted), you can pick up from the last successful frame by feeding ACEMD the run's final output files as the starting state — no `restart: true`:

```{code-block} yaml
:caption: input.yaml

structure: dhfr.psf
parameters: dhfr.prm
coordinates: output.coor   # final positions of previous run
velocities: output.vel     # final velocities
boxsize: output.xsc        # final box vectors
thermostat: true
run: 100ns
minimize: 0                # don't re-minimize already-equilibrated coordinates
```

```bash
acemd
```

Caveats:

- **Thermostat and barostat state are lost.** A new Langevin random seed is used; if you're sensitive to bitwise reproducibility this matters. For typical statistical-sampling work it does not.
- **Previous output files are renamed, not deleted.** Because this is a fresh run from ACEMD's point of view, any existing `output.csv`, `output.xtc`, `output.coor`, `output.vel`, `output.xsc` (and any user-set `trajforcefile` / `trajvelocityfile`) is moved aside to `output.1.<ext>` (then `output.2.<ext>`, …) before new ones are written. Your previous trajectory and final state are preserved, just under different filenames.

## Restart from Python

```python
from acemd import acemd

acemd(".", restart=True)
```

## Gotchas

- Restarting on a different GPU model fails. If you need to migrate hardware, run a fresh start from the last `output.coor`/`output.vel`/`output.xsc` instead.
- If the GPU drivers or ACEMD version changed since the checkpoint was written, the restart may fail. Re-running from `output.*` files is the safe fallback.

## See also

- `restart` option in [Input options](../reference/input-options.md)