austenite 0.1.2

Bayesian end-to-end materials engineering — chemistry to compliance, with citations.

austenite

Bayesian end-to-end materials engineering — chemistry to compliance, with citations.

🚀 New to this? Start with GETTING_STARTED.md — a beginner-friendly tutorial with copy-paste examples and explained results. 📖 Full capability reference: USAGE.md — every namespace (au.*) with runnable examples.

austenite is the Python client for the Austenite API. One function call takes you from composition to a code-stamped fatigue allowable, with Bayesian posterior bands and the literature citations the engine used.

import austenite as au

result = au.design_path(
    alloy="AISI 4140",
    cooling_rate=10,
    spec="ASME-VIII-Div2",
    sigma_app_MPa=600,
    posterior=True,
)

print(result.properties.sigma_y_MPa)        # 1535
print(result.compliance.pass_)               # True
print(result.posterior["sigma_y_MPa"].p05)   # 1500
for c in result.citations:
    print(c.source, c.page)

---

Install

pip install austenite

For the notebook examples:

pip install "austenite[notebook]"

For batch MTC parsing with PDF + OCR support (requires tesseract binary):

pip install "austenite[ocr,xlsx]"
# also install tesseract: brew install tesseract  / apt install tesseract-ocr
# language packs: tesseract-ocr-deu, tesseract-ocr-jpn, tesseract-ocr-kor, ...

For the JAX-differentiable backend (end-to-end gradients through the composition -> hardness -> strength -> fracture -> fatigue chain, suitable for jax.grad / jax.jit / jax.vmap and gradient-based inverse design):

pip install "austenite[jax]"

import austenite as au
import jax
import jax.numpy as jnp

def objective(comp):
    # comp = [C, Mn, Cr, Mo, Ni, Si, V] in wt%
    r = au.design_path_jax(comp, cooling_rate=10.0, sigma_app_MPa=500.0)
    return r.fatigue.utilisation_pct

gradient = jax.grad(objective)(jnp.array([0.40, 0.85, 0.95, 0.20, 0.0]))
# gradient[i] = sensitivity of fatigue utilisation to element i (wt%)

See examples/08_jax_inverse_design.py for a full inverse-design loop. The differentiable engines (Maynier-Dollet HV30, Pavlina-Tyne 2008, Barsom-Rolfe 1972, Hall-Petch, Goodman) are calibrated against the same references as the production HTTP backend. The conceptual basis for end-to-end differentiable matsci is Liu & Strachan (2025), arXiv:2505.01585 -- austenite ships the first production implementation.

For development:

pip install -e ".[dev]"

Python 3.10+.

---

What's in the box

Function	What it does
au.design_path(...)	Full ICME chain: composition → CALPHAD → KWN → Orowan → Sy(t) → Goodman
au.am.qualification(...)	Predict porosity, regime, and properties for LPBF / DED / EBM
au.am.sweep(...)	Grid sweep over laser power × scan speed
au.mtc.parse(pdf)	Parse a Material Test Certificate (PDF, URL, or raw text)
au.mtc.batch_parse(folder)	Folder of 1000 PDFs in 7 languages → pandas DataFrame (replaces 2-4 QC FTEs)
au.mtc.audit_lot(df, spec=...)	Compare a lot of heats against the spec library → Excel report
au.mtc.compare(a, b)	Element-by-element diff between two MTCs (>0.05 wt% flag)
au.mtc.detect_grade(comp)	Identify the most likely grade for a composition map
au.ttt.compute(...)	Isothermal TTT curves + M_s / B_s
au.ttt.cct_sweep(...)	CCT curves across a list of cooling rates
au.calphad.equilibrium(...)	Phase fractions at (T, P, composition)
au.calphad.ternary(...)	Sampled ternary isothermal section
au.bayesian.calphad(...)	Posterior P(phase ≤ threshold) — the differentiator
au.audit.record / replay(...)	Server-side audit log + replay

---

Configure

import austenite as au

au.set_api_url("https://api.austenite.org")   # default
au.set_timeout(30)                            # seconds
au.set_api_key("sk-...")                      # bearer token
au.set_max_retries(3)

Or set per-call:

client = au.Client(api_url="http://localhost:8000", timeout=10)

---

Async

import asyncio
import austenite as au

async def main():
    async with au.AsyncClient() as client:
        resp = await client.post("/v1/design-path", json={"alloy": "AISI 4140"})
        print(resp.json()["properties"]["sigma_y_MPa"])

asyncio.run(main())

---

Errors

import austenite as au

try:
    au.design_path(alloy="bogus")
except au.ValidationError as e:
    print(f"Bad input: {e}")          # HTTP 4xx
except au.RateLimitError as e:
    print(f"Slow down: {e}")          # HTTP 429
except au.EngineError as e:
    print(f"Server failed: {e}")      # HTTP 5xx
except au.NetworkError as e:
    print(f"Transport: {e}")          # timeout / DNS / reset

All transient failures (5xx, network) are retried automatically with exponential backoff (default 3 retries).

---

Examples

See examples/ for runnable scripts:

1. 01_design_path_quickstart.py — AISI 4140 end-to-end
2. 02_am_qualification.py — IN718 LPBF parameter sweep
3. 03_batch_mtc_parse.py — folder of MTCs → pandas DataFrame
4. 04_ttt_steel_grade.py — chemistry → CCT curves
5. 05_calphad_phases.py — Fe-Cr-Ni equilibrium + Bayesian σ-phase posterior
6. 07_batch_mtc_workflow.py — full 7-language batch MTC + audit + Excel export

---

Why this client?

• Typed inputs and outputs — pydantic v2 models, so your IDE knows the shape.
• Forward-compatible — server-side schema additions never break old clients.
• Production-ready transport — automatic retries, idempotency keys, timeouts, bearer-token auth.
• Sync + async in a single API.
• Citations on every call — every prediction carries the literature it stands on.

---

Links

• API: https://api.austenite.org/v1
• Docs: https://docs.austenite.org
• Issues: https://github.com/jvnshirr-collab/austenite/issues
• License: MIT

---

Built by Javanshir Hasanov.