PDE-Free Mass-Constrained Learning of Complex Systems with Hidden States: The Crowd Dynamics Case

Gianmaria Viola, Alessandro Della Pia, Lucia Russo , Ioannis Kevrekidis, Constantinos Siettos, Institute of Science, Technology for Energy, Sustainable Mobility (STEMS) Consiglio Nazionale delle Ricerche (CNR) , Naples 80125, Italy, Modeling, Engineering Risk, Complexity Scuola Superiore Meridionale · 2025

Read PDE-Free Mass-Constrained Learning of Complex Systems with Hidden States: The Crowd Dynamics Case by Gianmaria Viola, Alessandro Della Pia, Lucia Russo , Ioannis Kevrekidis, Constantinos Siettos, Institute of Science, Technology for Energy, Sustainable Mobility (STEMS) Consiglio Nazionale delle Ricerche (CNR) , Naples 80125, Italy, Modeling, Engineering Risk, Complexity Scuola Superiore Meridionale (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-17657.txt.

Introduction

Gianmaria Viola, with Alessandro Della Pia and three co-authors, learns hidden dynamics without writing equations in a 2025 arXiv paper. Many complex systems follow partial differential equations, meaning rules that link change across space and time, but the exact rules are unknown. Some system states also stay hidden from view. Their framework learns the dynamics in three tiers while respecting a fixed total mass.

First, a method called Diffusion Maps compresses the data into a low dimensional latent space. Next, simple models learn how the compressed state evolves. Finally, a neighbor based step lifts predictions back to full detail. Tests cover a crowd movement model and a fluid flow tracer problem. The latent models beat a standard compression baseline over long horizons.

Catalog record

CREATORGianmaria Viola, Alessandro Della Pia, Lucia Russo , Ioannis Kevrekidis, Constantinos Siettos, Institute of Science, Technology for Energy, Sustainable Mobility (STEMS) Consiglio Nazionale delle Ricerche (CNR) , Naples 80125, Italy, Modeling, Engineering Risk, Complexity Scuola Superiore Meridionale
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE72 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-17657

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

PDE-Free Mass-Constrained Learning of Complex Systems with Hidden States: The Crowd Dynamics Case Gianmaria Viola, Alessandro Della Pia, Lucia Russo , Ioannis Kevrekidis, Constantinos Siettos, Institute of Science, Technology for Energy, Sustainable Mobility (STEMS) Consiglio Nazionale delle Ricerche (CNR) , Naples 80125, Italy, Modeling, Engineering Risk, Complexity Scuola Superiore Meridionale, School for Advanced Studies , Naples 80138, Italy, Department of Chemical, Biomolecular Engineering, Department of Applied Mathematics and Statistics & Department of Urology, Johns Hopkins University , Baltimore, MD 21218, USA, Dipartimento di Matematica e Applicazioni “Renato Caccioppoli" Università degli Studi di Napoli “Federico II" , Naples 80126, Italy Abstract We propose a machine learning (ML) framework based on the next-generation Equation-Free algorithm for learning the spatio-temporal dynamics of mass-constrained complex systems with unobserved (hidden) states, whose dynamics can in principle be described by partial differential equations (PDEs), but lack explicit models. In these cases, the underlying variables, closures, and potentials governing the dynamics are generally unknown or not directly observable and therefore must be inferred from data. Here, we construct reduced-order models (ROMs) within a learned latent space – obtained via manifold learning – using delayed coordinates, thus exploiting the Takens’/Whitney’s embedding theorems. In the first stage, we employ both linear (Proper Orthogonal Decomposition, POD) and nonlinear manifold learning (Diffusion Maps, DMs) to extract low-dimensional latent representations of the complex spatio-temporal evolution. In the second step, we learn predictive ROMs using delayed manifold coordinates to approximate the solution operator on the latent space. In the final step, the latent dynamics are lifted back to the original high-dimensional space by solving a pre-image problem. We demonstrate that both POD and the particular k k -nearest neighbors ( k k -NN) lifting operators preserve the mass, a crucial property in the context of many problems, including computational fluid dynamics (CFD) and crowd dynamics. Actually, the proposed framework reconstructs the solution operator of the unavailable mass-constrained PDE, bypassing the need to discover an explicit form of the PDE per se, based on a manifold-informed objective map that bridges the scales . For our illustrations, we generated detailed spatio-temporal data using the Hughes model, which couples a continuity PDE with an Eikonal equation for the potential field, modeling how individuals choose optimal paths to minimize travel time while avoiding obstacles and high-density regions. We show that DMs provide nonlinear embeddings that outperform POD in reconstruction accuracy, thus resulting in better parsimonious manifold-informed ROMs that provide a stable and accurate approximation of the solution operator both in the latent space and, via reconstruction, in the high-dimensional space, and can therefore be integrated reliably over long time horizons. 1 Introduction Many real-world systems exhibit complex spatio-temporal dynamics that, in principle, can be described by partial differential equations (PDEs) but also depend on hidden variables [ 1 , 2 , 3 , 4 ] that are not directly observable. These hidden variables often represent unknown closures, as in the case of computational fluid dynamics (CFD) [ 5 , 6 , 7 , 8 , 9 , 10 ] , or internal states, as in biological systems [ 11 , 12 , 13 , 14 ] , or optimal transport strategies—such as those governing crowd dynamics [ 15 , 16 , 17 , 18 , 19 , 20 ] . For modeling such systems, one may consider a general data-driven setting in which the specific functional forms or entire operators of the macroscopic PDEs governing the emergent crowd dynamics are not known explicitly. The goal in this case is to use data, either from high-fidelity microscopic simulations or from real-world measurements, to approximate the right-hand sides or closures of the effective macroscopic equations within a finite-dimensional space. This can be achieved through black-box or gray-box surrogate models, typically implemented as (deep) neural networks (DNNs) [ 21 , 22 , 13 , 23 , 24 ] . Alternative machine learning schemes, such as convolutional neural networks (CNNs) [ 25 , 26 ] and Gaussian processes [ 27 , 28 ] , have also been explored in similar contexts. However, approximating the right-hand side of PDEs in finite spaces using DNNs poses various challenges and limitations. First, such an approach restricts the learned solutions to a particular discretization in time and space, thus limiting their ability to generalize across different resolutions. Another fundamental challenge is the accurate approximation of spatial and temporal derivatives. This task requires access to high-resolution samples, and additional pre-processing is needed in the case of noisy data [ 21 , 13 ] . Moreover, in real-world scenarios, data are typically available as discrete-time snapshots, limiting the accuracy of temporal derivative approximations. This can introduce significant modeling biases and numerical instabilities. Another approach is to approximate the integral evolution (solution) operator of PDEs, as defined by the corresponding Cauchy problem, using Neural Operators (NOs) [ 29 , 30 , 31 , 32 , 33 , 34 ] . NOs approximate continuous evolution operators of PDEs in infinite-dimensional spaces, theoretically enabling mesh-independent solutions and broad generalization. However, the NO paradigm requires learning a highly nonlinear mapping between function spaces, which can demand large training datasets and may suffer from the curse of dimensionality. Importantly, NOs face significant challenges in real-world observations of complex systems—such as in the case of crowd dynamics—when the training data are sparse or obtained from computationally expensive high-fidelity simulations. These issues can impede the applicability of NOs and their ability to accurately learn dynamics and may lead to poor generalization. To circumvent the “curse of dimensionality” inherent in the training of surrogate DNN- or NO-based PDE models, learning Reduced Order Models (ROMs) [ 35 , 36 ] in well-constructed latent spaces is crucial to significantly lowering computational cost while capturing the essential dynamics using surrogate DNNs, Gaussian Processes, or NOs [ 37 , 3 , 38 , 39 , 23 ] . These ROMs must also inherently preserve fundamental principles, such as stability for long-run simulations and “physical” accuracy when their dynamics are lifted back to the reconstructed space. Within this context, we propose a fully data-driven “PDE-free” machine learning framework based on manifold and machine learning, inspired by our previous work on the Equation-Free multiscale framework [ 40 ] and “Next-Generation Equation-Free” algorithms [ 41 , 3 , 37 , 42 , 9 , 10 , 43 , 44 , 4 ] . The proposed framework aims to learn the solution operator of the unavailable PDE for high-dimensional systems involving hidden or unobservable dynamics, without explicitly constructing a surrogate PDE per se, but by reconstructing the spatio-temporal dynamics based on ROMs in latent spaces. We illustrate the methodology through a crowd-dynamics paradigm. The objective is to learn appropriate low-dimensional manifold-informed surrogate models that can implicitly represent hidden dynamics in latent spaces by augmenting observables with delayed coordinates based on Takens’ and Whitney’s embedding theorems [ 45 ] . Spatio-temporal dynamics, usually modeled by PDEs, evolve on infinite-dimensional state spaces, but their dynamics often collapse to finite-dimensional attractors. Therefore, by applying time-delay embeddings to POD or DMs modes, Whitney’s/Takens’ theorem establishes a diffeomorphic equivalence between the delay-coordinate latent model and the original system dynamics (see results and discussion in [ 45 , 46 , 47 , 48 , 49 , 37 , 50

Read the complete text