Ten-Dimensional Neural Network Emulator for the Nonlinear Matter Power Spectrum

Area of Science:

• Cosmology
• Astrophysics
• Computational Science

Background:

• The nonlinear matter power spectrum is crucial for cosmological parameter inference.
• Next-generation surveys require efficient and accurate theoretical predictions.

Purpose of the Study:

• Introduce gokunemu, a 10D neural network emulator for the nonlinear matter power spectrum.
• Support next-generation cosmological analyses with rapid and accurate predictions.

Main Methods:

• Utilized the Goku N-body simulation suite and t2n-muse emulation framework.
• Developed a 10-dimensional emulator covering parameters beyond Lambda-CDM (ΛCDM).
• Trained the emulator to predict the matter power spectrum for 0≤z≤3 and 0.006≤k/(h Mpc⁻¹)≤10.

Main Results:

• Achieved ~0.5% average accuracy in matter power spectrum predictions.
• Emulator covers extended parameters: dynamical dark energy (w₀, wₐ), massive neutrinos (∑mν), N<0xE2><0x82><0x91><0xE2><0x82><0x91>, and spectral index running (α<0xE2><0x82><0x9B>).
• Prediction time is ~2 milliseconds per cosmology, orders of magnitude faster than existing emulators.

Conclusions:

• Gokunemu is the only emulator covering dynamical dark energy models favored by DESI constraints.
• Its speed and broad parameter coverage make it ideal for upcoming surveys (LSST, Euclid, Roman, CSST).
• Enables efficient interpretation of complex cosmological data.