HDF5 Schema for ChemGP Plots

HDF5 Layout

ChemGP Rust examples write HDF5 files with a fixed layout. The plt-gp subcommands each read specific groups and attributes from this schema.

Top-Level Groups

Group	Contents
grids/<name>	2D arrays with attrs x_range, y_range, x_length, y_length
table/<name>	Group of same-length 1D arrays (DataFrame columns)
paths/<name>	Ordered point sequences (x, y or rAB, rBC)
points/<name>	Point sets (x, y or pc1, pc2)
Root attrs	Metadata scalars (conv_tol, gp_e_mae, etc.)

Group Details

grids

Each grid is a 2D float64 dataset. Attributes store axis info:

x_range

[x_min, x_max] (float64[2])

y_range

[y_min, y_max] (float64[2])

x_length

Number of x grid points (int)

y_length

Number of y grid points (int)

Named grids: energy (true PES), gp_mean_N (GP at N training points), variance, nll, gradient_norm.

table

A group whose children are same-length 1D arrays, analogous to DataFrame columns. Common columns:

oracle_calls

Cumulative oracle evaluations

max_fatom, max_force, force_norm, ci_force

Force convergence metrics

energy

Total energy at each step

method

String array identifying the optimizer

d_rff, energy_mae, gradient_mae

RFF sweep data

paths

Each path is a group with x and y (or rAB and rBC) arrays tracing a reaction path on the 2D surface.

points

Each point set is a group with coordinate arrays:

x, y

Cartesian coordinates

pc1, pc2

PCA projections (for FPS scatter)

Named sets: minima, saddles, endpoints, training, selected, pruned.

Schema per Plot Type

Subcommand	Required groups/attrs
convergence	table/convergence (oraclecalls, force metric, method); root conv_tol
surface	grids/energy; optional paths/*, points/*
quality	grids/gp_mean_N* + points/training_N*; grids/energy
rff	table/rff (drff, energymae, gradientmae); root gp_e_mae, gp_g_mae
nll	grids/nll; optional grids/gradient_norm; root log_sigma2, log_theta
sensitivity	table/slice (x), table/true_surface (E); 9x table/gp_ls{j}_sv{i} (Epred, Estd)
trust	table/slice (x, Etrue, Epred, Estd, intrust); optional points/training
variance	grids/energy, grids/variance; optional points/training, stationary root attrs
fps	points/selected (pc1, pc2), points/pruned (pc1, pc2)
profile	table/profile (image, energy, method)
landscape	.con files in source_dir (not HDF5; delegates to plt-neb)

Filename Conventions

The surface and quality subcommands auto-detect energy clamping from the input filename:

Pattern in filename	clamp_lo	clamp_hi
mb	-200	50
leps	-5	5
(other)	data min	data max