Neural Conditional Simulation (NCS)

• Neural Conditional Simulation (NCS) is a family of generative methods that map conditioning information to samples from complex probability distributions.
• It employs various architectures—including encoder-decoder GANs, score-based diffusion models, and optimal transport maps—to surrogate expensive numerical simulators and solve inverse problems.
• NCS is applied in fields such as spatial statistics, biomechanical simulation, and crowd dynamics, offering efficiency gains by amortizing simulation over repeated queries.

Searching arXiv for papers on Neural Conditional Simulation and related conditional generative simulation frameworks. Neural Conditional Simulation (NCS) denotes a family of conditional generative approaches in which a neural model is trained to produce samples from a target conditional distribution under specified conditions. In one formulation, given physical or physiological parameters $c$, a learned model outputs samples from an approximation of the conditional distribution $p(x \mid c)$ implicitly defined by an expensive numerical simulator (Ma et al., 2022). In another formulation, NCS is a general method for spatial conditional simulation based on score-based diffusion models, aimed at sampling from predictive distributions at unobserved locations given observations and spatial process parameters (Walchessen et al., 27 Aug 2025). Across the cited literature, the term functions as an umbrella for neural surrogates of simulators, conditional posterior samplers, and data-driven simulators used for virtual experiments.

1. Scope and evolution of the concept

The literature does not present NCS as a single architecture. Rather, it appears as a unifying description for neural systems that map conditioning information to samples from complex conditional laws. The 2019 Neural Conditioner learns an exponential family of conditional distributions $P(X_r \mid X_a = x_a)$ of a random vector $X$ through masked inputs and adversarial training (Belghazi et al., 2019). The 2022 BioMime work applies a conditional generative neural network to emulate a finite element method (FEM) EMG simulator under varying physiological conditions (Ma et al., 2022). In the same year, conditional simulation using diffusion Schrödinger bridges extends diffusion-based generative modeling to posterior sampling in inverse problems (Shi et al., 2022). Later work broadens the paradigm to data-conditional simulation for stochastic differential equations in ABC-SMC (Jovanovski et al., 2023), conditional optimal transport maps for Bayesian inference (Wang et al., 2023), generative neural simulators for pedestrian crowds (Minartz et al., 2024), and a formalized NCS method for spatial predictive distributions based on neural diffusion models (Walchessen et al., 27 Aug 2025).

A common structural feature is that the neural system does not merely approximate a deterministic regression map. It represents, either explicitly or implicitly, a conditional law over high-dimensional outputs. In the surrogate-simulation setting, the output may be a waveform tensor such as a motor unit activation potential. In spatial statistics, it is a predictive distribution over unobserved field values. In inverse problems, it is a posterior $p(x \mid y)$. In crowd modeling, it is a conditional distribution over accelerations and hence trajectories (Ma et al., 2022, Walchessen et al., 27 Aug 2025, Shi et al., 2022, Minartz et al., 2024).

This suggests that NCS is best understood as a methodological stance: conditional simulation is amortized into a neural generator so that repeated downstream queries no longer require repeated execution of the original simulator, repeated MCMC, or repeated solution of an inverse problem.

2. Mathematical formulations of conditional simulation

The most compact formulation of NCS appears in the simulator-surrogate setting. A teacher simulator defines a conditional relation between system conditions and outputs, and the neural model learns

$$\boldsymbol{x} \sim p_\theta(\boldsymbol{x} \mid \boldsymbol{c_s}),$$

where $\boldsymbol{c_s}$ contains specified generative factors and $\boldsymbol{x}$ is the simulated object of interest (Ma et al., 2022). In BioMime, $\boldsymbol{x}$ is a MUAP tensor

$$\boldsymbol{x} \in \mathbb{R}^{T \times H \times W},$$

with $p(x \mid c)$0, $p(x \mid c)$1, and $p(x \mid c)$2, while $p(x \mid c)$3 encodes fiber density, conduction velocity, innervation zone position, fiber length, motor unit depth, and motor unit medial-lateral position (Ma et al., 2022).

In spatial statistics, the conditional target is a predictive distribution over unobserved locations. Let $p(x \mid c)$4 be a set of spatial locations, $p(x \mid c)$5 the observed subset, and $p(x \mid c)$6 the prediction subset. If $p(x \mid c)$7 and $p(x \mid c)$8, then the conditional simulation target is

$p(x \mid c)$9

with $P(X_r \mid X_a = x_a)$0 denoting spatial process parameters (Walchessen et al., 27 Aug 2025). The central claim of that framework is that a neural diffusion model can be trained using only unconditional simulations from the spatial process.

The Neural Conditioner gives a broader mask-based formulation. Given a random vector $P(X_r \mid X_a = x_a)$1, an available mask $P(X_r \mid X_a = x_a)$2, a requested mask $P(X_r \mid X_a = x_a)$3, and noise $P(X_r \mid X_a = x_a)$4, the model

$P(X_r \mid X_a = x_a)$5

is trained so that

$P(X_r \mid X_a = x_a)$6

for arbitrary mask pairs $P(X_r \mid X_a = x_a)$7 (Belghazi et al., 2019). This formulation is especially general because $P(X_r \mid X_a = x_a)$8 and $P(X_r \mid X_a = x_a)$9 need not be complementary.

A trajectory-level conditional law appears in neural crowd simulation. There, the learned stochastic dynamics are written at the level of acceleration,

$X$0

so that full trajectory simulation is obtained by integrating conditionally sampled accelerations over time (Minartz et al., 2024). A similar posterior-sampling interpretation is used in inverse problems, where the objective is to sample from

$X$1

by learning a neural stochastic process conditioned on $X$2 (Shi et al., 2022).

3. Generative mechanisms and training objectives

Several neural mechanisms instantiate NCS. BioMime uses an encoder-decoder conditional GAN with a hybrid VAE-GAN style loss. Its generator has a morphing mode, which transforms an existing MUAP $X$3 to new conditions $X$4, and an ab initio mode, which generates a MUAP from a latent draw $X$5 and desired conditions $X$6. The generator loss is

$X$7

combining an adversarial term, a KL regularizer toward $X$8, and a cycle-consistency penalty. The discriminator is explicitly conditional: it must reject not only fake waveform-condition pairs but also real waveforms paired with incorrect conditions (Ma et al., 2022).

The spatial-process formulation of NCS uses score-based diffusion models. Its forward process is a diffusion on the unobserved part of the field, while the observed values remain fixed. A neural score network approximates the conditional score needed in the reverse-time SDE, and spatial masks determine which coordinates are diffused. The model is trained with a score-matching objective derived from the diffusion transition kernel, and the paper adopts a variance-preserving SDE equivalent to a DDPM forward process (Walchessen et al., 27 Aug 2025).

Conditional Schrödinger bridges provide a different route. In that framework, the goal is to construct, for each $X$9, a bridge whose initial marginal is $p(x \mid y)$0 and whose terminal marginal is a reference Gaussian. The bridge is obtained by minimizing a KL divergence over path measures, and the formulation is amortized over the marginal law of $p(x \mid y)$1 so that training does not require direct sampling from the posterior at training time (Shi et al., 2022). This places conditional simulation inside a finite-horizon diffusion-bridge formalism rather than a long-horizon denoising process.

Conditional optimal transport gives yet another construction. In the static approach, a conditional Brenier-style map is approximated as the gradient of a partially input-convex neural network, while the dynamic approach represents conditional transport through the flow map of a regularized neural ODE. Training minimizes an expected negative log-likelihood,

$p(x \mid y)$2

subject to monotonicity or dynamic regularization, so that later conditional sampling reduces to pushing a tractable reference law through the learned map (Wang et al., 2023).

The Neural Conditioner is adversarial rather than diffusion- or transport-based. Its discriminator sees the requested block, the available block, and the masks, and training matches the conditional distribution of requested coordinates given available ones and the mask pattern. Because the conditioning operator is shared across mask pairs, a single model is trained to represent a combinatorially large family of conditional distributions (Belghazi et al., 2019).

4. Surrogate simulation of physical and physiological systems

BioMime is a mature instance of NCS as a teacher-student surrogate simulator. The teacher is Neurodec, an FEM-based EMG simulator operating in anatomically realistic forearm geometries. The original EMG forward model decomposes into motor neuron spike trains, which are cheap to simulate, and motor unit action potential templates, which are computationally expensive because they require solving a Poisson equation in a deforming volume conductor and, after FEM discretization, a large sparse linear system $p(x \mid y)$3 (Ma et al., 2022).

The dataset used for the neural surrogate consists of MUAP templates from 1500 distinct motor units under 256 combinations of fiber density, conduction velocity, innervation zone position, and fiber length, for a total of

$p(x \mid y)$4

MUAP templates. The six specified conditions are linearly normalized to $p(x \mid y)$5 and used as the conditioning vector for each sample (Ma et al., 2022). The paper argues that this lower-dimensional conditioning structure allows the neural model to interpolate between simulated conditions and thereby emulate FEM output at new settings without rerunning FEM.

The application-level importance of this surrogate construction is tied to dynamic movement. Under isometric conditions, surface EMG is modeled as

$p(x \mid y)$6

but during naturalistic motion the volume conductor deforms and the expensive part of the computation must be repeated at many time steps. BioMime replaces direct FEM generation of $p(x \mid y)$7 with a conditional neural simulator, after which full EMG can be synthesized by convolving the generated MUAP templates with motor neuron spike trains and adding Gaussian noise (Ma et al., 2022).

A related but distinct use of neural conditional simulation appears in crowd dynamics. The Neural Crowd Simulator, or NeCS, is trained on large-scale real-world trajectory data and used for surrogate crowd dynamics experiments without training on specific scenarios. It combines message-passing graph neural networks, node-wise LSTMs, and a conditional VAE, with a latent variable $p(x \mid y)$8 per pedestrian and an edge-correction module inspired by E(n)-equivariant flows (Minartz et al., 2024). Here the neural simulator is not a surrogate for a closed-form PDE or FEM solver, but it serves an analogous role: it provides laboratory-like controllability while preserving the statistical resolution of large-scale real-world data.

5. Spatial, Bayesian, and data-conditional inference settings

The 2025 spatial-process paper gives the most explicit formalization of NCS as a named method. It uses spatial masks to build a conditional score-based diffusion model that evolves Gaussian noise into samples from a predictive distribution when given a partially observed spatial field and spatial process parameters as inputs. A notable property is amortization: once trained, the same neural network can be reused for different observation patterns, different numbers and locations of observations, and different process parameters without retraining (Walchessen et al., 27 Aug 2025). The paper assesses NCS against exact Gaussian-process conditional simulation and against MCMC for a Brown-Resnick process model, and concludes that it enables efficient and accurate conditional simulation from predictive distributions that are difficult for traditional methods.

Conditional simulation in Bayesian inference is also addressed through Schrödinger bridges and optimal transport. Diffusion Schrödinger bridges extend unconditional bridge-based generative modeling to posterior sampling in inverse problems such as image super-resolution, optimal filtering for state-space models, and refinement of pre-trained networks. The stated motivation is that standard conditional score models require long diffusion horizons, whereas the bridge formulation can shorten generation time (Shi et al., 2022). Conditional optimal transport pursues the same broad objective through learned maps $p(x \mid y)$9 from a reference distribution to approximate posterior samples $$\boldsymbol{x} \sim p_\theta(\boldsymbol{x} \mid \boldsymbol{c_s}),$$0, emphasizing conditional sampling and conditional density estimation in simulation-based Bayesian inference (Wang et al., 2023).

A further variant is data-conditional simulation for discretely observed SDEs in ABC-SMC. That framework distinguishes forward simulation,

$$\boldsymbol{x} \sim p_\theta(\boldsymbol{x} \mid \boldsymbol{c_s}),$$1

from data-conditional simulation,

$$\boldsymbol{x} \sim p_\theta(\boldsymbol{x} \mid \boldsymbol{c_s}),$$2

and implements the latter using lookahead sequential importance sampling and backward simulation particle smoothing. The goal is to produce trajectories consistent with the observed trajectory so as to increase the ABC acceptance rate. The method additionally uses an invariant neural network to learn the summary statistics function, incrementally retrained using new ABC-SMC samples at each round (Jovanovski et al., 2023). The reported result is accurate inference with a runtime about three times faster than standard forward-only ABC-SMC.

Taken together, these works show that NCS can operate at several inferential levels: predictive distributions in spatial statistics, posteriors in inverse problems, smoothing distributions in partially observed dynamical systems, and simulator surrogates in mechanistic modeling.

6. Conceptual boundaries, logical interpretation, and acronym ambiguity

One conceptual boundary is that NCS is not identical to conditional reasoning in classical causal models. “On the Conditional Logic of Simulation Models” develops a formal semantics in which a conditional $$\boldsymbol{x} \sim p_\theta(\boldsymbol{x} \mid \boldsymbol{c_s}),$$3 is evaluated by intervening on a simulator and running it under clamped conditions. Although the simulator in that paper is a Turing machine rather than a neural network, the framework is explicitly presented as conceptually transferable to neural generative models (Ibeling et al., 2018). A central result is that the logic of simulation-based conditionals is weaker than the logics associated with structural equation models and normality-ordering models: principles such as Cautious Monotonicity can fail because clamping additional variables can disrupt the very procedural pathway that produced the consequent.

A plausible implication is that neural conditional simulators should not be assumed to inherit the algebra of Pearl-style intervention semantics simply because they support counterfactual or conditional generation. Their procedural and stateful nature matters. This point is especially relevant when NCS is used for “if-then” reasoning by rerunning a learned simulator under hypothetical conditions (Ibeling et al., 2018).

Another boundary concerns terminology. The acronym “NCS” is not unique to Neural Conditional Simulation. In “Enhancing Adversarial Transferability Through Neighborhood Conditional Sampling,” NCS refers strictly to Neighborhood Conditional Sampling, an adversarial attack constructed as a max-min bi-level optimization problem for transfer-based attacks (Qiu et al., 2024). That work is not a generic conditional generative simulation framework. The acronym collision is therefore substantive, not merely stylistic.

For this reason, the term Neural Conditional Simulation is most precise when reserved for methods whose core operation is to learn a neural mechanism that maps conditioning information to samples from a target conditional distribution, whether the target arises from a numerical simulator, a spatial predictive law, a posterior, or a stochastic dynamical system (Ma et al., 2022, Walchessen et al., 27 Aug 2025).