H5 converter

Where H5 comes from

The US National Center for Supercomputing Applications introduced HDF in the late 1980s for scientific data that did not fit a flat table. HDF5, first released in 1998, redesigned the storage model around groups and datasets and became a long-lived standard across earth observation, particle physics, microscopy, engineering and machine learning. The HDF Group now maintains the specification and reference C library; h5wasm compiles that library for bounded browser-local access.

How H5 works

• The eight-byte signature may begin at byte zero or after a power-of-two user block, which lets an application prepend its own metadata without changing the HDF5 superblock.
• Groups form a path-addressed hierarchy and datasets carry an explicit shape and datatype, so a 32-by-8 detector array remains distinct from a flat list of 256 values.
• Attributes attach units, provenance and conventions to the root, groups or datasets; they are read alongside values instead of discarded during discovery.
• Datasets may be chunked and filtered with compression or checksums. A reader needs the corresponding compiled filter, so unsupported plugins must fail explicitly rather than yield partial values.
• Virtual datasets and external links can point outside the current file. The local converter rejects those references because a single uploaded file cannot resolve them safely or deterministically.

When to use H5

• Inspecting instrument or simulation output without uploading research data
• Listing dataset paths, shapes and units before handing a file to another pipeline
• Exporting bounded arrays into a workbook or per-dataset CSV package
• Comparing deterministic scientific fixtures by hierarchy, metadata and values

Compatibility

The HDF5 reference tools, h5py, MATLAB, Julia, R, Mathematica, NetCDF-4 software and many domain applications read HDF5. Novus Convert uses a self-hosted h5wasm worker, applies separate constrained-device and decoded-value limits, rejects virtual or external resources, and independently reopens JSON, CSV, XLSX and ZIP outputs before offering them for download.

Root and dataset attributes, paths, shapes, datatypes and filter declarations are retained in JSON and the ZIP structural manifest. CSV is deliberately a dataset inventory, while XLSX puts each bounded dataset on its own worksheet and summarizes hierarchy separately.