AInstein: Numerical Einstein Metrics via Machine Learning

TL;DR

AInstein introduces a semi-supervised, patch-based machine learning framework to approximate Einstein metrics on general manifolds without symmetry. The architecture splits the manifold into two atlas patches, each predicted by a subnetwork, and enforces the Einstein condition $R_{ij}=\lambda g_{ij}$ alongside patch-gluing via a Jacobian $J$ with overlap handling. The loss combines an Einstein term, an overlap term, and a finiteness term, all modulated by radial filters; the model is tested on spheres $S^n$ for $n=2$–$5$, with $\lambda\in\{+1,0,-1\}$, using both supervised baselines and semi-supervised learning. Results show strong performance for $\lambda=+1$ (round metrics) across dimensions, while $\lambda=0,-1$ become increasingly challenging, providing numerical insight into the existence/non-existence questions for Ricci-flat metrics on higher-dimensional spheres and highlighting the method's potential for exploring Einstein geometry and related physics problems.

Abstract

A new semi-supervised machine learning package is introduced which successfully solves the Euclidean vacuum Einstein equations with a cosmological constant, without any symmetry assumptions. The model architecture contains subnetworks for each patch in the manifold-defining atlas. Each subnetwork predicts the components of a metric in its associated patch, with the relevant Einstein conditions of the form $R_{μν} - λg_{μν} = 0$ being used as independent loss components (here $μ,ν= 1, 2, \cdots, n$, where $n$ is the dimension of the Riemannian manifold, and the Einstein constant $λ\in \{+1, 0, -1\}$). To ensure the consistency of the global structure of the manifold, another loss component is introduced across the patch subnetworks which enforces the coordinate transformation between the patches, $g' = J^T g J$, for an appropriate analytically known Jacobian $J$. We test our method for the case of spheres represented by a pair of patches in dimensions 2, 3, 4, and 5. In dimensions 2 and 3, the geometries have been fully classified. However, it is unknown whether a Ricci-flat metric can exist on spheres in dimensions 4 and 5. This work hints against the existence of such a metric.

Figures (11)

• Figure 1: Overview sketch of the AInstein architecture. Here, $T$ is a patch transition function layer which converts the points in patch 1 to their equivalents in patch 2, $\{H^{(l)}_\text{Patch $p$}\}$ a set of hidden layers with non-linear activations for each patch, and 'Concat' a concatenation layer, then followed by a Cholesky transform on the output of the hidden states in the pipeline, producing the metrics on both patches.
• Figure 2: Visualisations of the $(0,0)$ components of the learnt metrics and their respective Ricci tensors, in 2d on a single patch. These metrics solve the Einstein equation with Einstein constants of $\lambda \in \{+1, 0, -1\}$ respectively. We emphasise the $R_{00}$$(\lambda = 0)$ scale is $\sim 10^{-5}$, indicating Ricci-flat.
• Figure 3: Visualisations of the analytic round metric, $g_{ij}$, in 2d on a ball patch. This metric solves the Einstein metric equation with positive Einstein constant ($R_{ij} = g_{ij}$), such that each metric component $g_{ij}$ equals its equivalent Ricci component $R_{ij}$.
• Figure 4: Visualisations of the learnt metrics, $g_{ij}$, in 2d, on the 2 patches, trained with positive Einstein constant (such that $R_{ij} = g_{ij}$).
• Figure 5: Visualisations of the Ricci tensors, $R_{ij}$, of the learnt metrics in 2d, on the 2 patches, trained with positive Einstein constant (such that $R_{ij} = g_{ij}$).