• Amortized inference. Neural networks, trained once on millions of simulated samples, map the data straight to the posterior in seconds (Radev et al. 2022).
• Forward models. Fission-track annealing and lengths after Ketcham et al. (2007); helium diffusion with radiation damage (RDAAM, Flowers et al. 2009).
• The prior. Slow background cooling plus a few exhumation pulses (2–50 Myr long) and reheating episodes (5–150 °C); temperatures of 0–200 °C; cooling rates up to ~90 °C/Myr; present-day temperature range set by the user (0–25 °C).
• Constraints. Boxes and cooling only are applied by rejection: paths are drawn without them and only those that pass are kept.
• Training uncertainties. Each helium age has a 12.3% error (1σ); fission-track age error is based on Ns.
• Training range. Up to 12 helium grains; radius 30–110 µm; eU 1–507 ppm; Dpar 1.2–3.5 µm; 20–150 track lengths; ages within the time span. Outside the training range, the model extrapolates, and the answer is less credible. Each such case is flagged in the result.

• Flowers et al. (2009). Geochimica et Cosmochimica Acta, 73(8), 2347–2365. doi:10.1016/j.gca.2009.01.015
• Ketcham et al. (2007). American Mineralogist, 92(5–6), 799–810. doi:10.2138/am.2007.2281
• Radev et al. (2022). IEEE Transactions on Neural Networks and Learning Systems, 33(4), 1452–1466. doi:10.1109/TNNLS.2020.3042395

Link to PyPI will be added here.

Jiao, R. Rapid Bayesian estimation of thermal histories from apatite fission-track and (U–Th)/He data by amortized inference. In preparation.