Tensor-Brain Model: Latent Dynamics

• The Tensor-Brain Model is a framework combining learned temporal evolution, spatial decoding, and low-dimensional latent representations to simulate high-dimensional systems.
• It employs a latent ODE operating in a compact space to capture temporal dynamics and reconstruct continuous spatial fields meshlessly.
• Key benefits include parameter efficiency, robust generalization, and scalability for applications in model reduction, control, and surrogate modeling.

A Temporal-Spatial-Latent-Dynamics Network (TSLDN) is a modeling paradigm that jointly learns temporal evolution, structural spatial dependencies, and low-dimensional latent representations, enabling accurate data-driven prediction and efficient simulation of high-dimensional, spatiotemporally varying systems. This framework underpins multiple recent advances in model reduction, system identification, and surrogate modeling for scientific computing, dynamical systems, and network science. Central to the TSLDN approach is the coupling of a compact latent state (that evolves in time under a learned dynamics) with an architecture for meshless or structure-aware spatial decoding, often driven by high-dimensional, possibly irregular data.

1. Architectural Principles and Manifold Discovery

A TSLDN typically comprises two coupled components: (1) a temporal dynamics module operating purely in a low-dimensional latent space, and (2) a spatial decoder capable of reconstructing observable fields at arbitrary locations. Unlike conventional autoencoder frameworks, which require explicit encoding of the high-dimensional system state, TSLDNs such as Latent Dynamics Networks (LDNets) directly learn the low-dimensional solution manifold and associated vector field end-to-end, bypassing explicit high-dimensional encodings altogether (Regazzoni et al., 2023).

The dynamical module parameterizes an ordinary differential equation (ODE) in latent coordinates:

$\dot{s}(t) = f_\theta(s(t), u(t)), \quad s(0) = 0,$

where $s(t) \in \mathbb{R}^{n_s}$ ($n_s \ll N_h$), $u(t)$ is an exogenous input or control, and $f_\theta$ is a compact fully connected neural network. All time evolution is performed in latent space; no full state vector is ever assembled.

The spatial field $y(\cdot, t)$ is then reconstructed meshlessly via a spatial decoder:

$\hat{y}(x, t) = g_\phi(s(t), u(t), x),$

where $x\in\Omega$ is an arbitrary spatial query, and $g_\phi$ is a neural network whose weights are shared for all $x$. This enables continuous, mesh-agnostic spatial outputs and strong parameter sharing, which underpins superior generalization and facilitates spatial arbitrarily-fine prediction.

Key to the approach is fully end-to-end training from input-output data, so that the latent manifold (and its associated dynamics) are automatically discovered without recourse to encoders operating on the full discretized state.

2. Mathematical Formulation of Latent Dynamics

The central innovation lies in formulating evolution entirely in a learned latent space. For deterministic systems, LDNets and related architectures utilize the ODE formalism above. For stochastic, discrete-event, or networked systems, alternative probabilistic evolution equations (e.g., Gaussian autoregressive priors, Markov chains, or Kalman state-space models) may govern the latent state (Turnbull et al., 2021, Artico et al., 2022).

In LDNets, the latent ODE is discretized via schemes such as Forward Euler:

$s(t) \in \mathbb{R}^{n_s}$0

for training and inference. All time evolution is thus performed without reference to high-dimensional physical variables.

In the case of latent space interacting network models, each node $s(t) \in \mathbb{R}^{n_s}$1 has a coordinate $s(t) \in \mathbb{R}^{n_s}$2, evolving as:

$s(t) \in \mathbb{R}^{n_s}$3

with the network topology at each $s(t) \in \mathbb{R}^{n_s}$4 governed by the distances among latent $s(t) \in \mathbb{R}^{n_s}$5 (Turnbull et al., 2021). Edge probabilities are typically modeled via distance-based link functions (e.g., logistic or Poisson variants), enabling embedding of evolving topologies as latent point clouds.

3. Meshless and Structure-Aware Spatial Decoding

A defining feature is the use of meshless, continuous spatial decoding. The spatial decoder $s(t) \in \mathbb{R}^{n_s}$6 receives the latent state, possible exogenous inputs $s(t) \in \mathbb{R}^{n_s}$7, and the query location $s(t) \in \mathbb{R}^{n_s}$8 to reconstruct the physical field value $s(t) \in \mathbb{R}^{n_s}$9.

In LDNets (Regazzoni et al., 2023), this meshless neural architecture enables output at any arbitrary spatial query point, not restricted to a fixed discretization. All spatial predictions share the same decoder weights, enforcing a strong spatial inductive bias and efficient weight-sharing.

Alternative instantiations use graph neural networks, convolutional autoencoders, or spectral graph methods to encode geometric priors or manifold constraints, but the unifying principle is that reconstruction is performed pointwise or per-location, with latent-to-physical mappings decoupled from any fixed grid (Wang et al., 5 Jul 2025, Xu et al., 2019).

4. Training, Loss Functions, and Recipe

Training in a TSLDN is fully supervised on input–output time-series pairs, with loss functions typically combining:

• A normalized mean-squared error (MSE) or negative log-likelihood over all observed space-time points:

$n_s \ll N_h$0

where $n_s \ll N_h$1 normalizes the field scale.

• $n_s \ll N_h$2-regularization terms for network weights:

$n_s \ll N_h$3

Optimization proceeds in two stages: initial training with Adam (e.g., a few hundred epochs at learning rate $n_s \ll N_h$4), followed by BFGS fine-tuning. Gradients are propagated through time (backpropagation-through-time on the ODE steps) and via ordinary backpropagation in the spatial decoder.

A canonical forward pass involves stepping the latent state via the ODE, reconstructing at all observed points, accumulating loss, and applying weight decay. The process is summarized in concise pseudocode in (Regazzoni et al., 2023).

5. Benchmark Accuracy and Generalization

TSLDNs deliver state-of-the-art performance across a range of scientific and engineering benchmarks (Regazzoni et al., 2023). Notable observations include:

• For linear advection–diffusion–reaction (ADR) systems with known intrinsic dimension 2, an LDNet with $n_s \ll N_h$5 recovers the underlying manifold precisely, yielding normalized RMSE $n_s \ll N_h$6 and capturing physically-relevant Fourier modes.
• In unsteady 2D Navier–Stokes (lid-driven cavity, $n_s \ll N_h$7), as few as 10 latents yield test NRMSE $n_s \ll N_h$8, with extrapolation accuracy maintained outside the training time horizon.
• For nonlinear cardiac (Aliev–Panfilov) dynamics, LDNets outperform all autoencoder+ODE, AE+LSTM, or POD-DEIM surrogates, achieving up to 5× lower normalized error with a parameter count reduced by an order of magnitude ($n_s \ll N_h$91.7k vs 18k–23k).

This pattern—order-of-magnitude parameter savings, tighter reconstruction error, strong extrapolation, and meshless flexibility—generalizes to diverse scientific and engineering domains.

A summary table of comparative performance on the Aliev–Panfilov system (see (Regazzoni et al., 2023)):

Model	Latent dim	Test NRMSE	Param count
POD-DEIM	60	$u(t)$0	–
AE+ODE, AE+LSTM	–	$u(t)$1	$u(t)$2–$u(t)$3k
End-to-end AE/ODE	–	$u(t)$4	$u(t)$518k
LDNet	12	$u(t)$6	$u(t)$7k

6. Interpretability, Generalization, and Design Implications

TSLDNs realize a modeling paradigm characterized by:

• Temporal dynamics confined to a learned low-dimensional manifold, permitting both compression and interpretability of system dynamics.
• Spatial structure enforced via meshless decoders, enabling weight sharing and evaluation at arbitrary spatial locations.
• End-to-end discovery of the intrinsic solution manifold and dynamics, without reliance on expensive encoders or fixed-grid operations.
• Capacity for time extrapolation and interpolation, attributed to the explicit, global nature of the latent ODE or Markovian core and the inductive bias from meshless decoding.
• Substantial parameter efficiency versus classical POD, DEIM, or autoencoder surrogates.

A plausible implication is that TSLDN models may serve as scalable surrogates for scientific simulation, real-time forecasting, or control applications in high-dimensional, stiff, or multi-scale systems.

7. Connections to Related Approaches and Extensions

While LDNet (Regazzoni et al., 2023) is the archetype, the TSLDN paradigm subsumes a much broader family of models:

• Temporal-Spatial-Latent Dynamics with Sequential Monte Carlo for dynamic networks (Turnbull et al., 2021), which parameterize the time-evolving topology of networks via latent positions that evolve temporally, allowing for flexible, online inference via high-dimensional particle filtering.
• Transformer-based latent ODE and spatio-temporal models, which replace explicit integration with continuous spatio-temporal attention mechanisms to encode latent dynamics without ODE solutions (Lagemann et al., 2023).
• Dynamic latent space models for networked relational events (Artico et al., 2022), which combine nonlinear distance-based intensity models with Kalman filtering for efficient tracking of evolving node latent positions.
• Graph convolutional and tensor-based variants, which further encode spatial structure into the latent space and exploit tensor decompositions or graph neural ops to capture geometry, cross-layer dynamics, or structured variable interactions (Wang et al., 5 Jul 2025, Lan et al., 3 Jun 2025, Xu et al., 2019).

In all variants, the unifying mechanism is decoupling the evolution of a compact, low-dimensional latent core from high-dimensional observations, reconstructing spatial structure meshlessly, and discovering both manifold and dynamics in a data-driven regime.

---

References

• Latent Dynamics Networks (LDNets): learning the intrinsic dynamics of spatio-temporal processes (Regazzoni et al., 2023)
• Sequential Estimation of Temporally Evolving Latent Space Network Models (Turnbull et al., 2021)
• Learning Latent Dynamics via Invariant Decomposition and (Spatio-)Temporal Transformers (Lagemann et al., 2023)
• Transformer with Koopman-Enhanced Graph Convolutional Network for Spatiotemporal Dynamics Forecasting (Wang et al., 5 Jul 2025)
• Multi-level Convolutional Autoencoder Networks for Parametric Prediction of Spatio-temporal Dynamics (Xu et al., 2019)