Abstract:Learning dynamics governing physical and spatiotemporal processes is a challenging problem, especially in scenarios where states are partially measured. In this work, we tackle the problem of learning dynamics governing these systems when parts of the system's states are not measured, specifically when the dynamics generating the non-measured states are unknown. Inspired by state estimation theory and Physics Informed Neural ODEs, we present a sequential optimization framework in which dynamics governing unmeasured processes can be learned. We demonstrate the performance of the proposed approach leveraging numerical simulations and a real dataset extracted from an electro-mechanical positioning system. We show how the underlying equations fit into our formalism and demonstrate the improved performance of the proposed method when compared with baselines.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to address the challenge of learning the dynamic equations of physical systems under partial measurements. Specifically, when some states of the system are not measured or the generated dynamics are unknown, how to accurately learn the dynamic models of these systems from data. #### Main problem description 1. **Partial measurement problem**: - In many practical applications, some states of the system may not be directly measurable, and these unmeasured states may have an indirect impact on the measured states. For example, in a complex physical system, some internal states may not have sensors to monitor. - In this case, the traditional Neural Ordinary Differential Equations (NODEs) methods will encounter difficulties during the optimization process because they usually assume that all states are known. 2. **Dynamics of unmeasured states are unknown**: - When the dynamics of unmeasured states are unknown, traditional methods are difficult to accurately estimate these states, resulting in the problem of multiple solutions in the joint estimation of model parameters and unmeasured states. - To solve this problem, the paper proposes a sequential optimization framework, which uses known physical knowledge and partial measurement data to learn the dynamics of unmeasured states. 3. **Optimization and recursive learning**: - The paper introduces a recursive optimization strategy based on Newton's method. By alternately optimizing states and model parameters, it gradually improves the estimation of unmeasured states. - This method not only improves the estimation accuracy of unmeasured states but also can effectively learn the global model of the system under partial measurements. #### Solution overview The paper proposes a sequential optimization framework that combines Physics Informed Neural ODEs and recursive state estimation methods. The specific steps are as follows: 1. **Define cost function**: - Construct a cost function \( L_N(\Theta_N, X_N) \) that contains a physical - information loss term, a data - driven loss term, and a regularization term, which is used to optimize the hidden state \( x(t) \) and model parameters \( \theta(t) \). 2. **Recursive optimization**: - Use recursive formulas to gradually update the hidden state and model parameters. By decomposing the optimization process into alternating optimization steps, first optimize the state \( x(t_i) \), and then optimize the parameter \( \theta(t_i) \), thereby avoiding the vanishing gradient problem. 3. **Experimental verification**: - Verify the effectiveness of the proposed method through numerical simulation and real - data sets (such as electromechanical positioning systems), and show the performance improvement in different dynamic systems. #### Mathematical formula representation - **Cost function**: \[ L_N(\Theta_N, X_N) = \frac{1}{2} \sum_{i = 1}^N \left( \|x(t_i)-f_o(x(t_{i - 1}),\theta(t_{i - 1}))\|^2_{Q_x^{-1}}+\|y(t_i)-h(x(t_i))\|^2_{R_y^{-1}}+\|\theta(t_i)-\theta(t_{i - 1})\|^2_{Q_\theta^{-1}} \right) \] - **Recursive update formula**: \[ \hat{x}(t_i)=f_o(\hat{x}(t_{i - 1}),\hat{\theta}(t_{i - 1}))-P_x^{-}H_i^T(H_iP_x^{-}H_i^T + R_y)^{-1}(h(f_o(\hat{x}(t_{i - 1}),\hat{\theta}(t_{i - 1})))-y(t_i)) \] \[ \hat{\theta}(t_i)=\hat{\theta}(t_{i - 1})-P_\theta^{-}F_{\theta i - 1}^

Learning Physics Informed Neural ODEs With Partial Measurements

Learning dynamics on invariant measures using PDE-constrained optimization

Physics-Informed Neural State Space Models via Learning and Evolution

Learning Governing Equations of Unobserved States in Dynamical Systems

Learning Stable Deep Dynamics Models for Partially Observed or Delayed Dynamical Systems

Physics-informed learning of governing equations from scarce data

Stochastic physics-informed neural ordinary differential equations

Recognition Models to Learn Dynamics from Partial Observations with Neural ODEs

Learning Dynamics from Noisy Measurements using Deep Learning with a Runge-Kutta Constraint

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

How to Learn and Generalize From Three Minutes of Data: Physics-Constrained and Uncertainty-Aware Neural Stochastic Differential Equations

Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations

Some of the variables, some of the parameters, some of the times, with some physics known: Identification with partial information

Physics-informed neural ODE (PINODE): embedding physics into models using collocation points

Physics guided neural networks for modelling of non-linear dynamics

Capturing Actionable Dynamics with Structured Latent Ordinary Differential Equations

Learning a Neural Solver for Parametric PDE to Enhance Physics-Informed Methods

Encoding physics to learn reaction–diffusion processes

An Analysis of Physics-Informed Neural Networks

Deep Hidden Physics Models: Deep Learning of Nonlinear Partial Differential Equations