MIT Center for Theoretical Physics – a Leinweber Institute

The measurement of the sum of neutrino masses is among the primary promises of precision cosmology, achievable by combining complementary early- and late-Universe probes. However, these datasets currently exhibit mild-to-strong disagreements within $Λ$CDM and its simplest extensions, giving rise to multiple tensions, including the Hubble tension, the preference for “negative” neutrino mass, and indications of evolving dark energy. It has recently been shown that these tensions can be alleviated by adopting a higher value of the optical depth to reionization parameter, $τ$, when large-scale cosmic microwave background (CMB) polarization data are ignored. We extend this proposal and show that an especially high prior on $τ= 0.11 \pm 0.006$ simultaneously addresses all three of these tensions, significantly reducing the need for new physics beyond $Λ$CDM. We determine the “concordance” value of $τ$ by requiring physical neutrino mass and consistency of the Hubble constant, $H_0$, inferred from the CMB and that preferred by the Dark Energy Spectroscopic Instrument (DESI) baryon acoustic oscillation (BAO) and full-shape measurements. Within this high-$τ$ Universe, we obtain the first $2σ$ detection of a positive neutrino mass, $Σm_ν = 0.10^{+0.04}_{-0.05}$~eV at 68% C.L., while restoring cosmological concordance between datasets within $Λ$CDM. In particular, low-redshift distance predictions are consistent with DESI BAO observations and the inferred dark-energy equation-of-state parameters are consistent with a cosmological constant, both with and without supernovae data. The concordance power of our $τ$ prior further motivates new measurements of $τ$, e.g., through large angular scale CMB polarization observations with the \textit{LiteBIRD}, CLASS, or proposed PICO experiments. (Abridged)

We compute the galaxy power spectrum at two-loop order in cosmological perturbation theory (effective field theory, EFT). We derive galaxy bias operators through the fifth order and obtain two-loop renormalization conditions for the their bias coefficients. We compute the two-loop integrals using a renormalization scheme consistent with the CLASS-PT code, allowing for an easy interface of our new computations with standard tools used in the one-loop galaxy power spectrum and bispectrum analyses. We also derive the relevant higher-derivative and stochastic contributions, and implement IR resummation using time-sliced perturbation theory. Having identified the redundant operators, we find that the two-loop galaxy power spectrum requires 21 additional EFT parameters per galaxy sample. We compare our computation with the galaxy-galaxy and galaxy-matter power spectra from the PT Challenge N-body simulation at $z=0.61$ and find a per mille-level agreement up to $k=0.85~h$Mpc$^{-1}$. We show that even with conservative priors on all EFT parameters, the two-loop model produces an unbiased measurement of the mass fluctuation amplitude $σ_8$ with three times narrower error-bars than the linear theory model. The improvement over the one-loop model is $\simeq 40\%$. This suggests significant gains in the two-loop EFT analyses of galaxy clustering and galaxy–lensing two-point functions ($2\times2$ pt) from CMB lensing maps and imaging surveys like Euclid, LSST, and Roman. In addition, our two-loop computation offers a probe of new physics scenarios that modify the shape of the matter power spectrum at wavenumbers $(0.4-0.8)~h$Mpc$^{-1}$ such as the presence of ultra-light axion dark matter sub-components with masses $m_a\sim 10^{-24}$ eV.

We present a field-level perturbative forward model for the Lyman-alpha (Lya) forest flux decrement. We validate it on two simulation suites: large-volume AbacusSummit N-body simulations with the Lya forest painted onto the dark matter field, and the Sherwood hydrodynamic simulations. Across the redshift range of the simulations (z=2.0-3.2), the 3D and 1D power spectra of the model match the simulated Lya fields at the 1% (5%) level up to k <= 0.3 (1.0) h/Mpc, with similar performance for the cross-correlation with massive dark matter halos. The counts-in-cells statistic shows excellent agreement down to cell radii of 2 Mpc/h. Leveraging cosmic variance cancellation, the model enables precision measurements of Lya bias parameters and robustly detects the full set of quadratic line-of-sight bias operators, consistent with the notion of naturalness in effective field theory (EFT). We quantify the stochasticity of the Lya forest (the analog to the one-halo term), and find it to be white (scale- and orientation-independent) on large scales, matching EFT predictions. We further find that phenomenological flux power spectrum models, based on modulations of the linear-theory power spectrum, fail at the field level even on quasi-linear scales. For the currently observing Dark Energy Spectroscopic Instrument (DESI), we generate large-scale clustering mocks of the Lya forest to validate cosmological parameter inference pipelines. Looking ahead to its successor, DESI-II, we produce large-volume mocks of representative samples of Lyman-break galaxies (LBGs) and Lya emitters (LAEs), calibrated on Astrid hydrodynamic simulations and matched to observations at z=3, enabling joint analyses of Lya forest and high-redshift galaxy data.

We review the concepts of interpretability and explainability as they apply to machine learning in physics. We define interpretability as concerning the structural transparency of a model (the ability to understand or approximate its inner workings) and explainability as concerning the scientific content of a model (the ability to map it onto domain knowledge). We discuss the trade-offs each entails (interpretability vs. expressivity; explainability vs. adaptability), the contexts in which each is needed, and the intrinsic and post-hoc tools available for achieving them. Throughout, we emphasize that machine-learned models are subject to the same scientific questions as classical models, differing only in scale, and that interpretability and explainability are best understood as deliberate modeling choices rather than inherent properties. We also emphasize the importance of task specification and intervention plans as a core aspect of model design.