PI: Alessio Spurio Mancini

The Euclid space telescope, launched in July 2023, is carrying out the most ambitious cosmological survey to date, mapping billions of galaxies across one-third of the sky. Extracting precise cosmological constraints from Euclid Data Release 1 (DR1) requires running the theoretical prediction pipeline millions of times within Bayesian inference frameworks — a task that is computationally prohibitive with traditional Boltzmann solvers such as CAMB.

During the 2025 reporting period (April–December 2025), we developed and validated COSMOPOWER neural network emulators specifically designed for the Euclid DR1 analysis pipeline. COSMOPOWER replaces the computationally expensive Boltzmann solver step with a neural network that delivers predictions for CMB and matter power spectra in milliseconds, achieving speed-ups of 3–4 orders of magnitude over CAMB while maintaining sub-percent accuracy.

Using DiRAC CSD3 CPU resources at Cambridge, we generated large training datasets of power spectra spanning the cosmological parameter space for both the baseline ΛCDM model and minimal extensions including the w₀wₐ dark energy parametrisation. The training data generation required approximately 500,000 CAMB evaluations per model, each producing power spectra at multiple redshifts — a task made feasible by the CSD3 Data Intensive system.

Key milestones achieved during 2025:

• June 2025: Completed emulators for ΛCDM and minimal extensions (w₀wₐCDM), validated against independent CAMB runs to sub-percent accuracy.
• December 2025: Completed emulators for extended cosmological models, including massive neutrinos and modified gravity scenarios.
• Integration of the emulators into the Euclid Likelihood code (CLOE), coupled with the COBAYA/PolyChord sampling framework, preparing the full inference pipeline for DR1 analysis.

These emulators will enable the Euclid Consortium to perform the large-scale parameter estimation runs required for DR1 cosmological results, expected in 2026. The COSMOPOWER framework developed with DiRAC resources is a critical component of the Euclid analysis pipeline.

Figure 1. CAMB (dashed) vs COSMOPOWER (solid) linear matter power spectra for four cosmologies. Residuals (bottom) are within 0.2% across all scales.

Publications (January–December 2025): None. The project started in April 2025. Publications based on the emulator development and Euclid DR1 analysis are expected from mid-2026 onwards.