Phidelity Sci · scientific data
Your simulations, lighter, with a guaranteed error.
Phidelity Sci compresses simulation and measurement fields in single or double precision, in 1 to 3 dimensions. You set the maximum error: every value read back respects it. One encoder for every kind of data: it chooses on its own, field by field. Three contracts: absolute error, relative error, or bit-exact lossless. Version 1.0, frozen format: C library, HDF5, NetCDF-4 and Zarr modules.
Ordinary matter density (NYX), at equal quality; 17 to 58 % on the six fields of the dataset.
Hurricane Isabel: 15 to 26 % more compact; CESM atmosphere model: 13 to 19 %, at equal quality.
The other two are a tie. None is lost.
In every case measured, including a test campaign on three series of public datasets, with verdicts computed value by value. SZ2 and SPERR exceeded it on some fields.
At equal quality
Average size difference at equal PSNR (BD-rate), over nine error levels from 1 % to 0.001 % of the field range, against the best setting of SZ3.1, QoZ 1.1 and HPEZ chosen separately for each field — which favours them. Public datasets from the SDRBench benchmark; quantum computer states simulated by us.
| Data | Dataset | Files |
|---|---|---|
| Cosmology | NYX, 6 3D fields | −17 to −58 % |
| Weather | hurricane Isabel, 3D | −15 to −26 % |
| Climate | CESM atmosphere, 2D | −13 to −19 % |
| 1D signals | Brownian-type series, 64-bit | −14 to −23 % |
| X-ray imaging | EXAFEL detector | −10 to −13 % |
| Quantum chemistry | QMCPack orbitals | −9 % |
| Simulated quantum computing | 24-qubit states | −5 to −20 % |
| Smooth turbulence | Miranda, 64-bit | tie |
| Molecular dynamics | EXAALT positions | tie |
At equal maximum error
Bits per value at 0.01 % of the range (smaller is better; uncompressed: 32 or 64 bits). Same maximum error for all, every reconstruction read back and checked.
| Dataset | Phidelity Sci | Best of SZ3.1 / QoZ / HPEZ | SZ3 | SZ2 | SPERR | ZFP |
|---|---|---|---|---|---|---|
| Hurricane Isabel (32-bit) | 1.13 | 1.49 | 1.54 | 1.65 * | 4.52 * | 4.06 |
| CESM climate (32-bit) | 2.21 | 2.62 | 2.74 | 2.92 | 4.57 * | 7.50 |
| Miranda turbulence (64-bit) | 0.45 | 0.50 | 0.68 | 1.15 | 2.81 | 2.24 |
* error bound exceeded on at least one field of the dataset.
Where it matters
Weather and climate
Model outputs and reanalyses: the most consistent lead, at every error level.
Computing centres and archives
Reference storage at very low error: 20 % less space, every year.
Cosmology and astrophysics
Large simulation volumes: up to 58 % smaller.
Numerical simulation
Physics, chemistry, plasmas: one tool for every field, in single or double precision.
Scientific imaging
X-ray detectors and instruments: measurement data with a controlled error.
Simulated quantum computing
Simulated computer states: up to 20 % smaller on a physical state.
“No data” values
Land in ocean models, unsurveyed areas, NaN or −9999: returned exactly, the rest within the bound. NOAA bathymetry to 10 cm: 300 times smaller than the original.
What we do not claim
- Format 1.0 frozen on 25 October 2026: every version must decode the reference files and reproduce them byte for byte. HDF5, NetCDF-4 and Zarr (v2 and v3) modules, tested with h5py, netCDF-C, xarray and zarr; Windows only for now.
- Lossless: on par with zstd and xz, no clear lead (0.4 to 6 % behind on particle positions and orbitals).
- At equal maximum error, our mean error is slightly higher (0.2 to 1.9 dB of PSNR): this is why we also compare at equal quality.
- Smooth turbulence and molecular dynamics: a tie, no lead.
- Speed (16 threads, whole fields): compression from 46 to 218 million values per second, reading back from 82 to 488. On large 3D fields HPEZ remains about 2.5 to 3 times faster, with clearly larger files. In chunked HDF5 files, our parallel writing and reading are faster than SZ3 and SPERR.
- On 64³ HDF5 chunks of very smooth 3D fields, SPERR produces files 27 to 46 % smaller than ours at equal mean quality. At equal maximum error we stay ahead in the vast majority of cases, and SPERR exceeded the bound on 16 of our 60 measurement points.
- Measurements from our own benchmarks on public datasets: to be confirmed on your data.
Try it on your data
- You send an extract of your files (HDF5, NetCDF or raw binary) and the error you can tolerate.
- You receive a report: size gain, fidelity (PSNR, gradients), bound checked on every value.
- Non-disclosure agreement if needed; no data is ever published.