Settings

Use Settings to adjust rendering, import and export behavior, the NEP backend, and plot styles. You do not need to tune everything at first. Change a setting when you encounter slow loading or rendering, slow prediction, an incompatible export format, or a need to switch between CPU and CUDA.

Choose the relevant setting

What you observe	Setting to check	Effect
Large structures render slowly	`Canvas Engine`	Switch the structure and scatter rendering backend among `Auto`, `PyQtGraph`, and `VisPy`
Force plots use an unfamiliar scale	`Force data format`	Switch between raw-force and normalized display
Opening a directory should not trigger automatic loading	`Auto loading`	Control whether related result files in the startup path are loaded automatically
Atom order makes card outputs difficult to inspect	`Sort atoms`	Sort atoms by element while processing structures
The card menu is crowded	`Use card group menu`	Group the Make Dataset menu by each card’s `group`
DeepMD export should preserve subdirectories	`Keep DeepMD subfolders`	Preserve the imported subdirectory structure when exporting `deepmd/npy`
Reopening the same outputs is slow	`Cache output files`	Cache `*.out` and `descriptor.out` files to avoid repeated parsing
Exported numeric precision is unsuitable	`Export significant digits`	Set the significant digits for per-atom values in `xyz` and `extxyz`
Source structures have no `Config_type`	`Default Config_type`	Assign a default provenance label when one is missing
Nonphysical-structure detection is too permissive or too strict	`Covalent radius coefficient`	Adjust the nearest-neighbor distance criterion

NEP backend

NEP Backend controls whether prediction uses CPU, CUDA, or automatic backend selection.

Auto: the default path. It uses CUDA when available and supported by the current model; otherwise it explains the decision before the task starts and uses CPU. It does not silently switch backend during execution.
CPU: use when no compatible GPU is available or when inspecting only a few structures.
CUDA: explicitly require the NVIDIA CUDA backend. If unavailable, the application reports the driver or installation problem instead of changing backend.

NEP Chunk Max Atoms sets the total atom count allowed in one prediction chunk, with the same meaning on CPU and CUDA. Reduce it when device or host memory is insufficient. If one structure alone exceeds the CUDA capacity, use CPU because a single structure cannot be split safely.

Data Precision controls the storage precision of imported DFT and structure data. Keep the default for normal work. Change it only to address clear memory pressure or a requirement to retain higher-precision values.

The runtime health section under NEP Settings reports NepTrainKit’s native helper modules, the nep-adapters version, and CPU/CUDA availability. Use NEP runtime updates to check PyPI manually for a compatible update. The application also checks once in the background after startup: no-update and network-failure results remain silent, while an available version opens an installation prompt. Before activation, NepTrainKit verifies the wheel SHA256 and uses an isolated process to confirm that it imports and that the CPU backend is usable. The update takes effect after restarting NepTrainKit; a failed validation leaves the existing version active.

For a pip installation, updated runtimes are stored in the user configuration directory. In a standalone Nuitka build, they are stored next to NepTrainKit.exe under runtime/nep-adapters/versions; the current and previous versions are retained for recovery.

Plots and structure rendering

Plot Settings affect presentation only; they do not alter the underlying data.

Scatter edge color, Scatter face color, and Face alpha: default scatter colors and opacity.
PyQtGraph scatter size, VisPy scatter size, and VisPy antialias: point size and antialiasing for each canvas backend.
Structure background and Lattice line color: background and cell-line colors in the structure view.
Selected color, Show color, Current marker color, and Current marker size: styles for selected points, highlighted points, and the current-structure marker.

When the goal is to spot anomalous structures, adjust point size and the current-structure marker first. Do not change the filtering workflow merely to compensate for plot colors.

NEP89 and software updates

About NEP89 checks for and updates the bundled nep89.txt. NEP89 can quickly prescreen candidates for some systems, but it is not reliable for every material. Use it to flag anomalies, never as a source of DFT labels.

If a NEP model trained for the current system is available, prefer it for prescreening.

Check for Updates under About checks for a newer NepTrainKit release. It only updates the version notification and never modifies training data automatically.

Guidelines for changing settings

First determine whether the problem is in rendering, input, prediction, export, or the card workflow, then change only the corresponding setting.
Change one important option at a time so its effect remains attributable.
After changing a data-cleaning threshold, inspect some of the selected structures rather than judging the change only by the count.