Abstract:Recovering a physics-driven model, i.e. a governing set of equations of the underlying dynamical systems, from the real-world data has been of recent interest. Most existing methods either operate on simulation data with unrealistically high sampling rates or require explicit measurements of all system variables, which is not amenable in real-world deployments. Moreover, they assume the timestamps of external perturbations to the physical system are known a priori, without uncertainty, implicitly discounting any sensor time-synchronization or human reporting errors. In this paper, we propose a novel liquid time constant neural network (LTC-NN) based architecture to recover underlying model of physical dynamics from real-world data. The automatic differentiation property of LTC-NN nodes overcomes problems associated with low sampling rates, the input dependent time constant in the forward pass of the hidden layer of LTC-NN nodes creates a massive search space of implicit physical dynamics, the physics model solver based data reconstruction loss guides the search for the correct set of implicit dynamics, and the use of the dropout regularization in the dense layer ensures extraction of the sparsest model. Further, to account for the perturbation timing error, we utilize dense layer nodes to search through input shifts that results in the lowest reconstruction loss. Experiments on four benchmark dynamical systems, three with simulation data and one with the real-world data show that the LTC-NN architecture is more accurate in recovering implicit physics model coefficients than the state-of-the-art sparse model recovery approaches. We also introduce four additional case studies (total eight) on real-life medical examples in simulation and with real-world clinical data to show effectiveness of our approach in recovering underlying model in practice.

What problem does this paper attempt to address?

This paper attempts to solve the problem of recovering physically - driven models from data under real - world constraints. Specifically, the paper focuses on how to accurately reconstruct the dynamic models of physical systems from unperturbed and perturbed system data, while overcoming the following five major challenges: 1. **Low Sampling Rate (C1)**: Storage and energy limitations in actual deployments often lead to low sensing frequencies, which makes it difficult to recover models from data. 2. **Perturbed Systems (C2)**: Since system operations cannot be interrupted, all observed data comes from systems affected by external inputs (such as human user inputs), and there is no data from unperturbed systems. 3. **Sparse Structures in High - Dimensional Non - Linear Function Spaces (C3)**: Physically - driven models are usually sparse, but in real life, this sparse information is often unknown. 4. **Implicit Dynamics (C4)**: Due to privacy, cost, and resource limitations, only a subset of system variables can be measured in actual deployments, resulting in some dynamics being implicit. 5. **Input Uncertainty (C5)**: In the real world, the timestamps of input perturbations are uncertain, which may be caused by sensor time synchronization or manual reporting errors. To solve these problems, the authors propose a new architecture (LTC - NN - MR) based on Liquid - Time - Constant Neural Networks (LTC - NN) and propose corresponding solutions for each challenge: - **To Deal with C1**: By using the automatic differentiation feature of LTC - NN, the system dynamics can be represented with arbitrary precision at a low sampling rate. - **To Deal with C2**: The forward - pass form of LTC - NN is the same as the control - affine dynamics, and the perturbed and unperturbed dynamics can be learned separately. - **To Deal with C3 and C4**: Through bilinear approximation, LTC - NN can search the implicit dynamics space and convert the implicit dynamics into an over - determined system of equations regarding the model coefficients. - **To Deal with C5**: By searching for the optimal input shift through dense - layer nodes to minimize the reconstruction loss, thus handling time synchronization or reporting errors. In addition, the paper also verifies the superiority of LTC - NN - MR in recovering implicit physical model coefficients through four benchmark dynamic systems (three simulated data and one real - world data) and four additional real - world medical case studies. In summary, the main contribution of this paper is to propose a new neural network architecture, LTC - NN - MR, which can more accurately recover the dynamic models of physical systems under the complex constraints of the real world.

Recovering implicit physics model under real-world constraints

Physics Constrained Learning for Data-driven Inverse Modeling from Sparse Observations

Physics-Informed Regularization for Domain-Agnostic Dynamical System Modeling

Physics-informed learning of governing equations from scarce data

DeepPhysics: a physics aware deep learning framework for real-time simulation

Extrapolation of Physics-Inspired Deep Networks in Learning Robot Inverse Dynamics

Estimating model inadequacy in ordinary differential equations with physics-informed neural networks

Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning

Recovering the state and dynamics of autonomous system with partial states solution using neural networks

Deep Physics Corrector: A physics enhanced deep learning architecture for solving stochastic differential equations

Combining physics and deep learning to learn continuous-time dynamics models

Multi-fidelity physics constrained neural networks for dynamical systems

Validation and parameterization of a novel physics-constrained neural dynamics model applied to turbulent fluid flow

Bidirectional Dynamic Neural Networks with Physical Analyzability

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

Learning to Represent Mechanics via Long-term Extrapolation and Interpolation

Physics-guided machine learning from simulated data with different physical parameters

Learning from Integral Losses in Physics Informed Neural Networks

Capturing dynamical correlations using implicit neural representations

Recovering missing CFD data for high-order discretizations using deep neural networks and dynamics learning