Neural network predicts gas pressure in galaxy clusters from dark matter alone

The work

Technical · from the record

The authors trained a rotationally and permutation equivariant set-based neural network (a DeepSets architecture) on the IllustrisTNG-300 simulation to predict the continuous electron pressure field inside galaxy clusters directly from the dark matter particles of a gravity-only run, as a cheaper stand-in for a full hydrodynamic simulation. The architecture is split into separate modules: one that corrects mis-centering of an analytic GNFW profile, one that reads the local particle environment, an aggregator, and a conditional-VAE module that models stochasticity in the mapping. On the held-out clusters the paper reports its loss metric improving by 70% over analytic profiles fitted to the same simulation data, with a further 7% from the conditional-VAE extension, which the authors describe as limited by their small training set. Module-level tests indicated that cluster triaxiality had negligible impact and that local regions already carried enough information to infer global cluster properties.

How AI was used

Clusters above a mass cut were identified in the present-day gravity-only IllustrisTNG 300-1 snapshot with the Rockstar halo finder, and target electron pressure fields were voxelised from the matching full-physics run with Voxelize at a voxel size fixed in units of R200. Cluster-scale scalars (mass, offsets, angular momentum, inertia-tensor eigenvalues) and unit vectors (angular momentum, offset, inertia-tensor eigenvectors) were constructed as SO(3) scalars and vectors, self-similarly normalised, and perturbed with noise during training. A composed network was then fitted to the voxelised targets with a loss normalised by the characteristic pressure scale: a vector DeepSet (Origin) shifted the centre of an analytic GNFW profile, a scalar DeepSet (Local) produced features from particles within a radius of the target position, an Aggregator MLP with dropout combined these with target-position and cluster-scale information, and a conditional-VAE encoder (Stochastic) mapped shell-averaged residuals to a one-dimensional latent code. Particle sets were sparsely sampled during training as regularisation and sampled more densely at test time. Training used Adam with a one-cycle schedule, per-module learning rates, weight decay and gradient clipping, with annealed KL weighting for the stochastic variants; hyperparameters were searched with Optuna, using multi-objective optimisation and Pareto-frontier selection where the Stochastic module was present. The trained network was then run over all voxels of the held-out clusters, and individual modules were added or removed to study their contributions.