A modeling study asks whether a present-day matter map can carry information about the earlier and final averaged state of an uneven universe. A three-dimensional neural network made predictions that tracked the true simulated values, with reported R² scores from 0.942 to 0.985. The finding concerns cosmic backreaction, the paper’s term for the averaged effects represented by its curvature and kinematical-backreaction parameters. It is a proof of principle within simulations, not a new measurement of the real Universe.
The network used the final density map as its input and predicted density parameters and Hubble parameters at both the initial and final epochs. The authors interpret the result as evidence that a present-time density field can contain information about averaged cosmological history, including backreaction. The test therefore concerns predictive correspondence inside a simulated model family.
Inside a controlled model
To create training examples, the researchers generated 92,610 independent relativistic, simplified cosmological simulations spanning initial conditions and average cosmological parameters. The simulations supplied both the final matter distribution and the averaged quantities the network was asked to recover. Because the exercise was built from simplified simulations, its result is bounded by that model setup.
The data were divided into 60 percent for training, 20 percent for validation and 20 percent for the held-out test. The researchers equalized the shares of different parameter combinations, then normalized both the density-map inputs and the cosmological targets. This separation gave the model a test set that was not used during training.
The model used four convolutional blocks, followed by global average pooling and a dense output layer. Dropout was set at 0.3. Training could run for up to 250 epochs and stopped after 20 successive epochs without an improvement in validation loss, using the AdamW optimization method. The researchers assessed the outputs with R² and with histograms of relative prediction errors.
Close tracking, with a weaker point
On the held-out simulations, the reported R² values ranged from 0.942 to 0.985. In practical terms, the study says the predictions generally followed the known simulated values closely. Relative-error histograms were roughly centered around zero, and only a small fraction of predictions reached percent-level errors or above. For most parameters, the error tails extended to about minus 0.1 to plus 0.2, while the initial matter-density parameter, Ωm,i, had a much narrower tail of about 0.0001.
The weakest reported output was the cosmological-constant parameter, ΩΛ. Its initial-time R² was 0.942, compared with 0.969 at the present-time endpoint. The analysis showed vertical prediction patterns associated with discrete input values. Initial outputs were generally more accurate than present outputs, except for ΩΛ, and the model slightly under-predicted the largest initial kinematical-backreaction values.
Within the simulated averages, the backreaction signal appeared mainly through a curvature parameter that was systematically larger than its FLRW counterpart, a nonzero kinematical-backreaction term and a larger averaged dimensionless Hubble parameter. These are patterns in synthetic results, not measurements of those quantities in the real Universe.
The boundary is the data
The central limitation is realism. The authors explicitly describe the simplified simulations as insufficiently realistic for sensible application of the trained network to real data. The paper points to realistic N-body simulations and weak-lensing inputs as future extensions, rather than treating the current result as an observational method.
A further warning came from the LTB check. Evolution was reproduced closely when exact initial conditions were used. With simplified initial conditions, however, the results differed for shear and Weyl curvature, and present-time density and expansion showed noticeable deviations, especially near the density peak.
The authors identify dataset compilation as the main obstacle to extending the approach. Their proposed next tests include inputs from N-body simulations, weak-lensing maps and other observational probes. Until work reaches those settings, the study remains a demonstration that present-time density maps can predict averaged quantities within the simplified simulations used here.
A simulated result, not a cosmic measurement
The manuscript is an arXiv preprint, identified as arXiv:2608.27962v1, version August 31, 2026. The project was funded by Villum Fonden through grant VIL53032, with SMK as principal investigator. Its evidence is limited to synthetic simulations, a held-out test split and the reported model checks, so it does not yet establish how the method would perform on the real Universe’s data.
Paper data and sources
Original title: Learning the averaged history of an inhomogeneous universe from its present day density field
Authors: Jonas Broe Bendtsen, Sofie Marie Koksbang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text