calypso v1.0.0 — Trained PCA runtime artifact and training/evaluation dataset for circumbinary accretion light curves

Trained runtime artifact and training/evaluation dataset for calypso (Circumbinary Accretion Lightcurves Yielded via Predictive Sequence Outputs), a parameter-conditioned stochastic surrogate model for circumbinary accretion time-series.

Given a binary eccentricity eb and mass ratio qb, calypso returns synthetic accretion-rate light curves for the total binary (Mb) and each component (M1, M2). The model is trained on a suite of 100 2D hydrodynamic simulations of circumbinary accretion disks (Arepo, Navier-Stokes) spanning eb ∈ [0.0, 0.8] and qb ∈ [0.1, 1.0]. From each simulation, 500 detrended 10-orbit windows of the concatenated (Mb, M1, M2) time series form the training matrix.

The model has three layers:  1. A single global PCA basis fit by SVD over all training windows, retaining k=142 components (≳90% of the variance).  2. For each training (eb, qb), an empirical multivariate Gaussian over the PCA coefficients across its 500 windows. This captures the aleatoric uncertainty of the accretion process — including precession-driven long-term modulation — directly in the latent space.  3. To predict at unseen (eb, qb), per-binary mean vectors and Cholesky factors of the covariances are linearly interpolated across the parameter grid and recombined into a positive semi-definite covariance. Sampling from the resulting Gaussian and projecting back through the PCA basis produces synthetic time series.

Files: calypso_pca_emulator.pkl is the runtime artifact (PCA basis + per-binary Cholesky factors), loaded by the calypso Python package via calypso.load_emulator(). TS_{train,test}_{Mb,M1,M2}_*.pkl are the detrended time-series windows used to fit the model (train/test split: 87/13 simulations).

Code: https://github.com/mssiwek/calypsoReference: Siwek et al. (2026), in prep.