Software
I develop open-source, JAX-native scientific software for the systematic exploration of the string landscape. The StringJAX ecosystem brings together a family of interoperating, separately-versioned, separately-citable Python packages that make the geometry → effective theory → moduli-stabilisation pipeline fast, differentiable, and reproducible on CPU, GPU, or TPU.
The StringJAX ecosystem
Packages
-
stringjax (GitHub · docs) — the umbrella. A thin metapackage that pins compatible versions of the member packages and routes new users to the right entry point.
-
jaxvacua (GitHub · docs) — Type IIB flux-vacuum engine. Periods, Kähler potential, GVW superpotential, F-terms, ISD samplers and a linearised optimiser, Newton refinement, flux bounding, and conifold / coni-LCS limits.
-
stringforge (GitHub · docs) — data layer. Curated Calabi–Yau geometry databases (TDF / CICY / KKLT) plus the persistent vacua vault, served from HuggingFace and cached locally. Solver-light: a single
LCSDatabase(...).load_model(...)call hands you a JAXVacua-ready model. -
jaxpolylog (GitHub · docs) — JAX-native polylogarithms with custom JVPs that survive arbitrary-order autodiff down to $|z| \approx 10^{-30}$. Used by
jaxvacuafor worldsheet-instanton sums; standalone-useful for anything needing differentiable $\mathrm{Li}_s(z)$.
Datasets (HuggingFace)
- cy-database
— lazy parquet shards of precomputed Calabi–Yau threefold
data, organised as sub-datasets:
tdf/— all FRST phases of all double-favourable, trilayer polytopes in KS up to $h^{1,1} = 11$, plus GV data and $h^{1,2} = 2$ conifolds.cicy/— complete-intersection Calabi–Yau threefolds.kklt/— curated index over TDF matching the working IDs of our current de Sitter constructions.
Catalogues are ~10 MB; individual model shards pull on demand via
stringforgeor plainpandas+huggingface_hub. - vacua_vault
— community vault of designated flux-vacuum records with
provenance. Currently scaffolding only; will be populated as members of
the group submit vetted vacua. Pushed and fetched via
stringforge.VacuaWriter/LCSDatabase.push_vacua_to_hub; readable as plain parquet for analysis or ML training.
Status
All four packages are pip install-able, GPL-3.0, Python ≥ 3.12,
run on CPU and GPU, and ship with quickstart notebooks in the docs. Happy
to demo on a call — get in touch.