NEO Model Family Overview

Updated 5 February 2026

NEO Model Family is a collection of diverse models spanning visual navigation, asteroid surveys, quantum chemistry, material mechanics, psychometrics, and language processing.
It integrates advanced methodologies like variational Bayesian inference, kernel density estimation, symmetry-exploiting Hamiltonians, and spectral clustering to achieve state-of-the-art performance.
Its practical implications include enhanced data efficiency, cross-domain generalization, and improved interpretability for research, mission design, and AI applications.

The NEO model family comprises a diverse set of statistical, physical, and algorithmic models unified by the acronym “NEO” (which may variously denote Next Expected Observation, Near-Earth Object, Nuclear-Electronic Orbital, Non-Equilibrium Orbit, neo-Hookean, or the NEO personality model). The family encompasses prominent approaches in fields as diverse as planetary science, quantum chemistry, statistical inference, mechanics of materials, deep reinforcement learning, and personality psychometrics. Despite their disciplinary breadth, several NEO models serve as foundational or state-of-the-art tools within their respective domains.

Several recent visual navigation architectures implement “Next Expected Observation” (NEO) models, notably NeoNav (Wu et al., 2019). These models learn to predict the agent’s future sensory input given current and target observations, thereby facilitating model-based, target-driven control. The NeoNav formalism is a latent-variable variational Bayesian model, described by the joint: $p(x_{t+1},z_t|x_t,y,a_t) = p(x_{t+1}|x_t,y,z_t)\,p(z_t|x_t,a_t).$ Inference is amortized with an encoder network $q(z_t|x_t,x_{t+1},y)$ , and training maximizes the conditional likelihood lower bound (ELBO): $\log p(x_{t+1}|x_t,y,a_t) \geq \mathbb{E}_q[\log p(x_{t+1}|x_t,y,z_t)] - \mathrm{KL}[q(z_t|x_t,x_{t+1},y)||p(z_t|x_t,a_t)].$ A critical innovation is the use of a mixture-of-Gaussians prior on the latent code, conditioned on both current observation and action, with the mixture weights and parameters generated by neural networks. The decoder is a conditional generative network for next-view synthesis given current state, target, and latent.

These models support adaptation to multiple sensor modalities, alternative goal representations, and alternative (e.g., flow-based) priors. Their modular design yields significant gains in cross-target and cross-scene generalization, data efficiency, and close-to-goal stability relative to model-free and classical model-based approaches. Empirical results include ∼5%–10% improvements in navigation success rate and significant gains in Sampled Path Length (SPL) on real-world and synthetic benchmarks (Wu et al., 2019).

2. NEO Surveyor, NEOMOD 2, and Near-Earth Asteroid Population Models

In planetary science, the NEO model family encompasses formal asteroid population and survey-completeness tools for Near-Earth Objects (NEOs). Among these are:

Known Object Model (KOM) (Grav et al., 2023): A synthetic-but-empirically-calibrated model of the NEA population used by NASA’s NEO Surveyor mission, constructed by kernel density estimation over (a, e, i), calibrated size–frequency distributions, and albedo assignment. KOM “replays” historical survey performance, flagging as “discovered” those synthetic NEAs that would have been detectable on each epoch, yielding cumulative completeness

$C(D > D_0) = N_{\rm known}(D>D_0)/N_{\rm tot}(D>D_0),$

with $C(D>1\,\rm km)\approx88\%$ and $C(D>140\,\rm m)\approx38.3\%$ by the end of 2022. The projected post-Surveyor completeness for $D>140\,\rm m$ is $\sim76\%$ (Grav et al., 2023).

NEOMOD 2 (Nesvorny et al., 2023): An updated CSS/Pan-STARRS–calibrated NEO population model providing precise size, orbital, and detection-efficiency distributions. NEOMOD 2 consolidates a decade of Catalina Sky Survey data, resolves previous normalization inconsistencies for large NEOs, and introduces an explicit “tidal-disruption source” to model the observed low-eccentricity, low-inclination excess among small NEOs (H>25). NEOMOD 2 estimates $936\pm29$ NEOs with $H<17.75$ (D>1 km, $p_V=0.14$ ), a total population $N(H<28)=(1.20\pm0.04)\times10^7$ , and an Earth impact flux (for $H<28$ ) of $0.034\pm0.002$ yr⁻¹ (main-belt sources only) rising to $0.06\pm0.01$ yr⁻¹ with tidal disruption included.

These models constitute the current reference standards for NEO population, risk, and survey design studies, and their structure—synthetic population draw, completeness replay, detection-efficiency parametrization, and explicit source mixing—is increasingly incorporated into mission studies and International Asteroid Warning Network (IAWN) protocols.

3. Dynamically Cold Resonant NEO Families and the Problem of Genetic Families

The NEO model family also includes dynamical approaches for identifying resonant subpopulations and genetic families.

Dynamically cold 1:1–resonant family (Marcos et al., 2013): Identified through direct N-body integration as a cluster of low-e, low-i NEOs transiently trapped in the 1:1 mean–motion resonance with Earth (horseshoe and quasi-satellite librators), with well-defined orbital boundaries: $0.985\,\rm au < a < 1.013\,\rm au$ , $0 < e < 0.1$, $0^\circ < i < 8.56^\circ$ , $H\approx24$ –30. At least 11 confirmed members are catalogued, and these objects display uniquely low encounter velocities ( $U<0.16$ ), high repeated trapping rates, and exceptional accessibility for spacecraft rendezvous. Their presence impacts flux estimates, transient satellite capture predictions, and low- $\Delta v$ mission targeting.
Genetic NEO families (Schunova et al., 2012): Systematic clustering studies employing the Southworth–Hawkins metric and high-fidelity “fuzzy-real” surrogate catalogs have failed to identify any statistically significant (≥3 $\sigma$ ) genetic near-Earth families despite multiple attempts. A weak (∼2 $\sigma$ ) candidate cluster of four objects in the Amor population survives various physical consistency tests (taxonomy, SFD, tidal disruption plausibility) but is not confirmed in proper-element space. This null result underscores the utility of the “fuzzy-real” model approach and highlights the apparent rarity of post-catastrophic tightly clustered NEO fragments among the present survey catalog.

4. NEO Models in Quantum Chemistry, Mechanics, and Probabilistic Inference

The NEO model family is prominent in several computational and theoretical domains outside planetary science.

Nuclear-Electronic Orbital (NEO) methods (Kovyrshin et al., 2023): Treat select light nuclei quantum-mechanically on equal footing with electrons, moving beyond the Born–Oppenheimer approximation. The NEO Hamiltonian in second quantization, with tight symmetry exploitation (parity, point-group $\mathbb{Z}_2$ , spin projection), is mapped to reduced-qubit quantum circuits. Advanced initialization (parameter transfer from classical VQE solutions) and hardware-efficient ansätze (NEO-UCCSDT) attain chemical and entanglement accuracy (errors below $10^{-6}$ Ha and $10^{-4}$ respectively) for H $_2$ and malonaldehyde, while achieving polynomial scaling and robust tensor reduction for NISQ quantum hardware.
Compressible neo-Hookean (NEO) material models (Korobeynikov et al., 27 Jun 2025): The “mixed” and “vol-iso” neo-Hookean energy forms represent generalizations of the classical incompressible NEO hyperelastic model. Parameter choices (notably the Hartmann–Neff volumetric term $h^{(q)}(J)$ with $q\geq2$ ) ensure thermodynamic admissibility, monotonicity, and robust stress/extreme-states responses. Mixed models offer broader flexibility in volumetric law selection and algebraic simplicity for implementation. These variants form the basis for contemporary finite-element solid-mechanics codes dealing with slightly compressible elastomers, foams, and soft matter.
Non-Equilibrium Orbit (NEO) samplers (Thin et al., 2021): The NEO-IS and NEO-MCMC algorithms build weighted, unbiased estimators on the orbits generated by deterministic, invertible maps $T$ that fail to leave the target $\pi$ invariant. The NEO-IS estimator for the normalizing constant $Z$ or expectation $\pi(f)$

$\widehat Z_N = \frac{1}{N}\sum_{i=1}^N \sum_{k\in\mathbb{Z}} w_k(X^i) L(T^k(X^i))$

consistently outperforms standard IS and AIS on multimodal, high- $d$ distributions. The associated Markov kernel exhibits uniform geometric ergodicity with explicit mixing time bounds, and specialization to conformal-Hamiltonian symplectic flows yields state-of-the-art MCMC performance on challenging targets.

5. NEO Models in Personality Psychometrics

IPIP–NEO and Factor Models (Brocklebank et al., 2011): The International Personality Item Pool–NEO (IPIP–NEO) implements a 300-item textual instrument operationalizing the Five Factor Model of personality (Neuroticism, Extraversion, Openness, Agreeableness, Conscientiousness). Large-scale spectral clustering studies corroborate the five-domain structure but further reveal a naturally emerging sixth cluster interpretable as Honesty–Humility (HEXACO); this sixth dimension, found by network-theoretic clustering but missed by conventional factor analysis, challenges strict five-factor dogma and motivates cross-validation of personality structures using spectral and graph-based approaches. Cluster assignment for the five-cluster solution matches varimax-rotated FA at 95.3%, but the sixth cluster supports a more nuanced, multi-scale personality model.

6. NEO-Based LLM Families (GPT-Neo)

GPT-Neo family (Kashyap et al., 2022, Lai et al., 2023): Comprises open-source decoder-only transformers (125 M – 2.7 B parameters, EleutherAI) trained on “The Pile.” GPT-Neo models demonstrate predictable scaling in commonsense reasoning benchmarks, achieving reasonable but sub-leading accuracy compared to top-tier LLMs. Attention-head probe studies reveal layerwise algorithmic specialization: token copying, induction, and abstraction. Robustness tests show strength to lexical perturbations but sensitivity to negations and role swaps.
Heaps’ Law adherence (Lai et al., 2023): Across all model sizes, GPT-Neo–generated corpora follow Heaps’ law ( $V(N) = K N^\beta$ ). β decreases from 0.79 (125 M) to 0.72 (2.7 B), approaching the human-authored biomedical corpus value β ≈ 0.64 as model size increases. The smallest models overproduce singleton types, inflating β; larger models more accurately emulate human vocabulary growth rates. Scaling and architectural refinements are recommended to further minimize singleton proliferation and enhance linguistic plausibility.

7. Cross-Model Implications, Best Practices, and Limitations

Across the NEO model family, representation choices—be they action-conditional priors (NeoNav), composite source mixture models (NEOMOD 2), symmetry-exploiting Hamiltonians (NEO–quantum chemistry), or spectral network-based clustering (NEO–psychometrics)—substantially enhance model generalization, data efficiency, or interpretability.

A consistent theme is the rigorous integration of model-based inference with careful empirical calibration (as in asteroid population synthesis and detection-efficiency replay) or principled structure-extraction strategies (as in high-dimensional data clustering). Limiting factors in several cases include dependence on accurate detection/survey modeling (planetary NEOs), adequacy of underlying numerical approximations (NEO–quantum), and the need for continual rebenchmarking as new data or computational modalities arise.

The broad applicability and architectonic diversity of the NEO model family underscore the importance of model structure, prior flexibility, modular inference, and robust cross-validation in the design and deployment of contemporary scientific models.

References

“NeoNav: Improving the Generalization of Visual Navigation via Generating Next Expected Observations” (Wu et al., 2019)
“The NEO Surveyor Near Earth Asteroid Known Object Model” (Grav et al., 2023)
“NEOMOD 2: An Updated Model of Near-Earth Objects from a Decade of Catalina Sky Survey Observations” (Nesvorny et al., 2023)
“A resonant family of dynamically cold small bodies in the near-Earth asteroid belt” (Marcos et al., 2013)
“Searching for the first Near-Earth Object family” (Schunova et al., 2012)
“A Quantum Computing Implementation of Nuclear-Electronic Orbital (NEO) Theory” (Kovyrshin et al., 2023)
“Two types of compressible isotropic neo-Hookean material models” (Korobeynikov et al., 27 Jun 2025)
“NEO: Non Equilibrium Sampling on the Orbit of a Deterministic Transform” (Thin et al., 2021)
“Characteristic Characteristics” (Brocklebank et al., 2011)
“GPT-Neo for commonsense reasoning – a theoretical and practical lens” (Kashyap et al., 2022)
“Heaps' Law in GPT-Neo LLM Emulated Corpora” (Lai et al., 2023)