Output

output

Output: where a Simulation writes and how often, and the chunked writer (SPEC §7.2, §10.6).

Output is the record (plan task 1b-T01). ChunkedOutput (plan task 1b-T06) writes the thermo rows and trajectory frames of a run as chunks, thermo.<n>.csv and trajectory.<n>.traj for n = 0, 1, ..., with a chunk boundary exactly at every restart point (SPEC §10.6 "Output files at restart points"):

• restart_point() closes the current chunks, fsyncs them and their directory (plan task 2-T10; rows and frames in between are flushed, not fsynced) and opens the next pair; it returns the indices the restart file stores (io.restart.OUTPUT_KEYS): output_chunk, the last closed chunk, and output_rows and output_frames, the rows and frames in chunks 0 to output_chunk. Call it after draining the output queue (AsyncOutput.restart_point does both), then write the restart file. Every restart write is a chunk boundary, a stop request's included.
• Resume (resume_from= those indices) checks the chunks up to output_chunk (present, with the stored numbers of rows and frames) and that every later chunk can be deleted, before it touches anything; otherwise it raises RestartMismatch. Then it deletes the later chunks and opens chunk output_chunk + 1. Nothing is ever rewritten: the research engine's resume reread and rewrote md.traj (workflow.truncate_outputs, up to 770 MB on the NAS).

Rows and frames written after the last restart point (by the asynchronous writer, or before a crash) are thus in later chunks only. A crash between restart_point and the restart write leaves the older restart file, and resume from it deletes the newer chunks as well.

Every chunk is complete on its own: each thermo chunk starts with the header, and a chunk without rows or frames still has both files (an empty trajectory reads as no frames). read_thermo and read_frames concatenate the chunks in order; ase.io.read(path, index=":") reads one. This module imports neither torch nor ASE at import time (fm.Output is a light name).

Output dataclass

Output directory and cadences, in MD steps (defaults: the production settings of §7.7).

trajectory_every and restart_every may be None (no trajectory, no restart files); thermo rows are always written. directory is kept as given, as a string.

read_thermo

read_thermo(
    directory: str | Path,
) -> tuple[list[str], list[list[str]]]

The header and every thermo row over all chunks, in order (rows as strings).

read_frames

read_frames(directory: str | Path) -> list[Atoms]

Every trajectory frame over all chunks, in order.