neuroGravity: Neural Inference Meets Gravity

• neuroGravity is a multidisciplinary term denoting three distinct neural frameworks that integrate gravity-based physics for urban mobility reconstruction, 3D geophysical inversion, and quantum sensing of entanglement.
• In urban mobility, it employs a physics-informed meta-gravity kernel combined with graph neural networks to reconstruct sparse origin–destination flows and achieve up to 38% improvement in R² over traditional baselines.
• The geophysical and quantum variants leverage implicit neural representations and quantum reservoir computing respectively to enhance gravity data inversion and detect gravity-induced entanglement with superior resolution.

In the cited arXiv literature, neuroGravity denotes three distinct constructs rather than a single unified framework: a physics-informed deep learning model for reconstructing and transferring urban human mobility networks (Yang et al., 26 Apr 2026), an implicit neural representation approach to three-dimensional gravity inversion (Mishra et al., 17 Oct 2025), and a quantum neuromorphic platform applied to sensing gravity-induced entanglement (Krisnanda et al., 2022). The shared nomenclature reflects a recurring combination of neural parameterization with gravity-based physical structure or gravity-related inference targets, but the underlying domains, data modalities, objectives, and mathematical formalisms are different.

1. Terminological scope and principal usages

In the available literature, the term is attached to three research programs with non-overlapping technical meanings.

Usage	Domain	Core definition
neuroGravity	Urban mobility	Physics-informed deep learning for OD reconstruction and transfer
NeuroGravity	Geophysical inversion	INR-based three-dimensional gravity inversion
neuroGravity	Quantum sensing	Quantum neuromorphic sensing of gravity-induced entanglement

The urban-mobility usage defines neuroGravity as a framework that reconstructs human mobility flows from sparse observations and transfers to cities with no observed flows, using population and open built-environment data (Yang et al., 26 Apr 2026). The geophysical usage defines NeuroGravity as an implicit neural representation that maps spatial coordinates to a continuous subsurface density field and is trained directly through a physics-based forward-model loss (Mishra et al., 17 Oct 2025). The quantum usage refers to a quantum reservoir processor that learns to estimate entanglement, including gravity-induced entanglement between two masses, from reservoir observables (Krisnanda et al., 2022).

A common misconception is that the name refers to a single method portable across urban science, geophysics, and quantum information. The literature instead supports a disambiguated reading: the same label has been reused for separate architectures in separate fields.

2. neuroGravity in urban human mobility reconstruction

In urban mobility research, neuroGravity is introduced to reconstruct an origin–destination directed weighted graph with flows $F_{ij}$ denoting the average daily number of trips from region $i$ to region $j$, under the realistic regime where only a small subset of flows is observed (Yang et al., 26 Apr 2026). A city is partitioned into administrative or functional zones, and the available inputs for each region are population $P_i$, 52 OpenStreetMap-derived built-environment features, and pairwise distances $D_{ij}$. The target output is the full OD matrix $\hat F = [\hat F_{ij}]$, with absolute volumes rather than only normalized trip probabilities.

The framework is explicitly physics-informed. It uses classical mobility-law structure as a prior, especially the gravity family of models, but avoids the requirement—common in doubly constrained gravity formulations and some deep baselines—of knowing total outflows $O_i$ or inflows $D_j$. Its learnable gravity kernel, termed meta-Gravity, is

$$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$

where $h^0_i$ and $i$0 are the initial regional feature vectors, and $i$1 and $i$2 are outputs of small MLPs. This provides a physically grounded base estimate that is then refined by a graph model.

The architecture has four stages. First, node features are constructed from population and the 52 OSM variables; edge features begin from distances. Second, a LightGBM classifier predicts which OD pairs likely carry stable flows, producing a candidate edge set $i$3 and exploiting the sparsity of urban mobility. Third, meta-Gravity yields base flows and initial edge features $i$4. Fourth, an edge-enhanced Graph-BERT updates node and edge embeddings with edge-aware attention and predicts log-flows through a head of the form

$i$5

Operating in log space is used to linearize multiplicative gravity relationships and mitigate long-tailed flow distributions.

Training combines meta-Gravity pretraining with joint optimization of the gravity kernel and the graph model. Two Huber-based losses are used over observed edges with positive flows, with $i$6. In few-shot reconstruction, larger flows are emphasized through

$i$7

Early stopping uses a validation set of internal flows from 20% of the observed zones and stops if neither $i$8 nor CPC improves for 100 epochs.

The sparse-observation regime is deliberately stringent. In Boston, with 250 ZCTAs and 51,786 OD pairs, a 10% observation ratio of zones covers roughly 1% of all flow links; the model is also reported to work under even scarcer observations, including 1% of links observed. Under the 10% zones-observed setting in Boston, the reported comparison is: Gravity, $i$9, CPC $j$0; neuroGravity, $j$1, CPC $j$2. Similar advantages are reported in LA and SF Bay, and averaged over cities under the 10% observation scenario, neuroGravity improves $j$3 by 38% over the strongest baseline, a plain GNN (Yang et al., 26 Apr 2026).

3. Transferability, embeddings, and global mobility proxies

A defining claim of the mobility formulation is that it transfers to cities with no observed flows. In zero-shot mode, a model trained on a source city is applied to an unobserved target city using only population and OSM features; multiple source models trained on independently sub-sampled source networks can be ensembled for robustness (Yang et al., 26 Apr 2026). With Boston as source, the reported zero-shot transfer performance is $j$4 for LA, $j$5 for SF Bay, $j$6 for Bogotá, and $j$7 for Rio de Janeiro. For LA and SF Bay, these values are stated to be comparable to training with 10% local observation, namely $j$8 and $j$9.

The model also yields regional embeddings $P_i$0 that correlate with socioeconomic and livability indicators not observed during mobility training, including income, education, NO$P_i$1, household carbon footprint, and radius of gyration. Using a gradient boosting machine with inputs equal to OSM features plus embeddings improves prediction accuracy over OSM alone in multiple cities, and SHAP analyses rank embedding components among the most influential predictors. In Boston, GBMs using OSM plus neuroGravity embeddings achieve high $P_i$2 for NO$P_i$3, carbon footprint, and radius of gyration, reported as $P_i$4, $P_i$5, and $P_i$6, with moderate accuracy for income and education, reported as $P_i$7 and $P_i$8.

A central empirical result is that spatial income segregation governs transferability. The paper defines a spatial income segregation index

$P_i$9

where $D_{ij}$0 captures between-cluster segregation and $D_{ij}$1 captures within-cluster segregation. The transfer $D_{ij}$2 from source $D_{ij}$3 to target $D_{ij}$4 declines as $D_{ij}$5 grows, and a linear regression using SI together with other covariates such as region area distribution and OSM feature density is reported to achieve a fit $D_{ij}$6 of 0.97. The most influential factor is $D_{ij}$7, followed by $D_{ij}$8 and $D_{ij}$9.

The framework has been applied to more than 1,200 cities, and neuroGravity-estimated OD networks for over 1,200 cities/regions worldwide are publicly released. The released flows are described as usable in downstream models; one stated example is that SEIR simulations in LA and SF Bay using neuroGravity flows closely match those using observed flows.

Its limitations are explicit. OSM completeness is heterogeneous globally, and missingness above approximately 30% may degrade performance. The architecture does not impose explicit mass constraints and therefore trades strict conservation for the ability to infer absolute flows from open data. Privacy constraints prevent release of raw mobile traces, and socioeconomic inference is cautioned to inherit data biases (Yang et al., 26 Apr 2026).

4. NeuroGravity as an implicit neural representation for gravity inversion

In geophysics, NeuroGravity is an entirely different construct: a scientific machine-learning approach to three-dimensional inversion of gravity data using an implicit neural representation (Mishra et al., 17 Oct 2025). The subsurface density is parameterized as a continuous field

$\hat F = [\hat F_{ij}]$0

where $\hat F = [\hat F_{ij}]$1, $\hat F = [\hat F_{ij}]$2 is an MLP, $\hat F = [\hat F_{ij}]$3 is a positional encoding, and $\hat F = [\hat F_{ij}]$4 is a bounded activation. The model is described as mesh-free from a modeling standpoint because its trainable parameters are shared network weights rather than per-cell densities, even though the forward operator is evaluated numerically on a chosen discretization.

The underlying physics is standard gravity forward modeling. The vertical component at observation location $\hat F = [\hat F_{ij}]$5 is

$\hat F = [\hat F_{ij}]$6

In implementation, the volume integrals are computed with the rectangular-prism kernel, assembling a sensitivity matrix once so that

$\hat F = [\hat F_{ij}]$7

with $\hat F = [\hat F_{ij}]$8 populated by evaluating $\hat F = [\hat F_{ij}]$9 at cell centers. Observed data are whitened as $O_i$0, and the loss is the whitened data misfit without explicit smoothness, smallness, or depth weighting terms.

The default INR is a fully connected MLP with three hidden layers of widths $O_i$1, LeakyReLU activations with slope $O_i$2, and bounded output. Positional encoding uses inclusive sinusoidal encoding with dyadic frequencies; by default $O_i$3 and $O_i$4, giving 63 input features. The paper states that positional encoding mitigates spectral bias and, in ablations, reduces density error and RMS gravity residual by approximately an order of magnitude relative to a plain coordinate MLP.

Two synthetic benchmarks are emphasized. In a Gaussian random field test on a $O_i$5 grid with 40 $O_i$6 40 surface stations, the encoded MLP recovers sharper contrasts and textures consistent with the GRF morphology, converges faster, and reaches substantially lower loss. In a dipping block model on a $O_i$7 grid, both the INR and a deterministic regularized inversion fit data near the noise level, but the INR recovers sharp lateral boundaries and the dipping geometry with only mild rounding at corners and slight amplitude underestimation, whereas the deterministic inversion spreads contrast, rounds edges, and broadens the body vertically.

The method is presented as parameter-efficient and scalable, but not as a resolution of the classical non-uniqueness of gravity inversion. The paper explicitly states that non-uniqueness remains, that solutions depend on encoding bandwidth, capacity, and forward discretization, and that very large surveys require careful management of the sensitivity operator in memory and batching.

5. neuroGravity in quantum neuromorphic sensing of gravity-induced entanglement

In quantum information and optomechanics, neuroGravity refers to the application of a quantum reservoir processor to the sensing of gravity-induced entanglement between two masses (Krisnanda et al., 2022). The platform is also called an “uncontrolled quantum network”: a quantum network of randomly interacting nodes with minimal control, where only a single classical output layer is trained. The internal Hamiltonian parameters need not be precisely tuned, and randomness and dissipation are allowed.

The inputs are the objects whose entanglement is to be sensed; the reservoir comprises $O_i$8 quantum nodes. In the continuous-variable setting, the dynamics are governed by Langevin equations for the quadrature vector and by covariance-matrix evolution

$O_i$9

After an evolution time $D_j$0, local reservoir observables such as quadrature means, variances, and mean excitations are recorded as a feature vector $D_j$1. The trainable output layer is linear:

$D_j$2

so the training problem is ridge regression from reservoir observables to the unique elements of the input covariance matrix or density matrix.

A notable feature is that training can use only separable input states without compromising performance on entangled test inputs, because the network learns a state estimator rather than a direct entanglement-only map. Entanglement is then computed analytically from the reconstructed state, using negativity or logarithmic negativity,

$D_j$3

Measurement errors are modeled as Gaussian perturbations of the reservoir observables, and the paper reports precision scaling beyond the standard quantum limit. With $D_j$4 independent observables, the SQL scales as $D_j$5; numerics show beyond-SQL scaling with increasing node number and with time-multiplexing, and in the gravity application the reported standard deviation obeys $D_j$6 over the explored range.

The gravity application considers two identical spherical masses in one-dimensional harmonic traps, separated by equilibrium distance $D_j$7, with gravitational interaction

$D_j$8

At time $D_j$9, cavity probes are turned on, and the joint cavity readout is used to reconstruct the two-mode Gaussian covariance matrix of the masses. For the reported parameter set—$$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$0 kg, $$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$1 Hz, initial squeezing $$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$2, $$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$3s, cavity lengths $$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$4 mm, laser wavelength $$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$5 nm, and $$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$6 mW—the platform achieves $$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$7 across test instances with $$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$8, $$\hat{F}^{g}_{ij} = \frac{G(h^0_i \oplus h^0_j)\, P_i P_j}{D_{ij}^{\alpha(h^0_i \oplus h^0_j)}},$$9, and $h^0_i$0. This is described as a two-order-of-magnitude improvement over a state-of-the-art value of approximately $h^0_i$1, while direct measurements on the masses would require three orders of magnitude smaller measurement error to match $h^0_i$2.

6. Conceptual relations, distinctions, and recurring themes

The three usages share the combination of neural inference with gravity-related physics, but their technical commonalities are limited. In the mobility framework, gravity is a prior over human trips and is blended with graph representation learning; in the geophysical framework, gravity is the forward operator linking density to observations; in the quantum framework, gravity is the mediator generating entanglement to be sensed. This suggests that the shared name is primarily nominative rather than indicative of a common architecture.

Several distinctions are essential for accurate interpretation. In the mobility literature, neuroGravity is not a doubly constrained gravity model and does not require known $h^0_i$3 or $h^0_i$4; if strict marginals are needed, post-hoc rebalancing via iterative proportional fitting is suggested (Yang et al., 26 Apr 2026). In the geophysical literature, “mesh-free” refers to the model representation rather than the forward computation, because the forward gravity operator is still evaluated on a discretization of rectangular prisms (Mishra et al., 17 Oct 2025). In the quantum literature, the platform does not employ an entanglement witness $h^0_i$5 and does not perform full tomography of the inputs; instead, it reconstructs state parameters from local reservoir observables and computes entanglement analytically (Krisnanda et al., 2022).

Taken together, these works place neuroGravity at three different interfaces: urban science and transfer learning, geophysical inversion and implicit neural fields, and quantum sensing and reservoir computing. Their coexistence under the same name is best understood as a case of terminological convergence across disciplines rather than as a single research lineage.