Modeling continuous processes from data

Discovering time-varying aerodynamics of a prototype bridge by sparse identification of nonlinear dynamical systems

Vortex-induced vibrations (VIVs) have been observed on a long-span suspension bridge. The nonstationary wind in the field characterized by the time-varying mean wind speed is likely to lead to time-varying aerodynamics of the wind-bridge system during VIVs, which is different from VIVs induced by stationary or even steady wind in wind tunnels. In this paper, data-driven methods are proposed to reveal the time-varying aerodynamics of the wind-bridge system during VIV events based on field measurements on a long-span suspension bridge. First, a variant of the sparse identification of nonlinear dynamics algorithm is proposed to identify parsimonious, time-varying aerodynamical systems that capture VIV events of the bridge. Thus we are able to posit new, data-driven, and interpretable models highlighting the aeroelastic interactions between the wind and bridge. Second, a density-based clustering algorithm is applied to discovering the potential modes of dynamics during VIV events. As a result, the time-dependent model is obtained to reveal the evolution of the aerodynamics of the wind-bridge system over time during an entire VIV event. It is found that the level of self-excited effects of the wind-bridge system is significantly time varying with the real-time wind speed and bridge motion state. The simulations of VIVs by the obtained time-dependent models show high accuracies of the models with an averaged normalized mean-square error of 0.0023. The clustering of obtained models shows underlying distinct dynamical regimes of the wind-bridge system, which are distinguished by the level of self-excited effects.

Characterizing chaotic dynamics from simulations of large strain behavior of a granular material under biaxial compression

For a given observed time series, it is still a rather difficult problem to provide a useful and compelling description of the underlying dynamics. The approach we take here, and the general philosophy adopted elsewhere, is to reconstruct the (assumed) attractor from the observed time series. From this attractor, we then use a black-box modelling algorithm to estimate the underlying evolution operator. We assume that what cannot be modeled by this algorithm is best treated as a combination of dynamic and observational noise. As a final step, we apply an ensemble of techniques to quantify the dynamics described in each model and show that certain types of dynamics provide a better match to the original data. Using this approach, we not only build a model but also verify the performance of that model. The methodology is applied to simulations of a granular assembly under compression. In particular, we choose a single time series recording of bulk measurements of the stress ratio in a biaxial compression test of a densely packed granular assembly-observed during the large strain or so-called critical state regime in the presence of a fully developed shear band. We show that the observed behavior may best be modeled by structures capable of exhibiting (hyper-) chaotic dynamics.

Uncovering bifurcation patterns in cortical synapses

Individual cortical synapses are known to exhibit a very complex short-time dynamic behaviour in response to simple "naturalistic" stimulation. We describe a computational study of the experimentally obtained excitatory post-synaptic potential trains of individual cortical synapses. By adopting a new nonlinear modelling scheme we construct robust and repeatable models of the underlying dynamics. These models suggest that cortical synapses exhibit a wide range of either periodic or chaotic dynamics. For stimulus at a fixed rate our models predict that the response of the individual synapse will vary from a fixed point to periodic and chaotic, depending on the frequency of stimulus. Dynamics for individual synapses vary widely, suggesting that the individual behaviour of synapses is highly tuned and that the dynamic behaviour of even a small network of synapse-coupled neurons could be extremely varied.

Modeling global vector fields of chaotic systems from noisy time series with the aid of structure-selection techniques

We address the problem of reconstructing a set of nonlinear differential equations from chaotic time series. A method that combines the implicit Adams integration and the structure-selection technique of an error reduction ratio is proposed for system identification and corresponding parameter estimation of the model. The structure-selection technique identifies the significant terms from a pool of candidates of functional basis and determines the optimal model through orthogonal characteristics on data. The technique with the Adams integration algorithm makes the reconstruction available to data sampled with large time intervals. Numerical experiment on Lorenz and Rossler systems shows that the proposed strategy is effective in global vector field reconstruction from noisy time series.

Constraining the topology of neural networks to ensure dynamics with symmetry properties

This paper addresses the training of network models from data produced by systems with symmetry properties. It is argued that although general networks are global approximators, in practice some properties such as symmetry are very hard to learn from data. In order to guarantee that the final network will be symmetrical, constraints are developed for two types of models, namely, the multilayer perceptron (MLP) network and the radial basis function (RBF) network. In global modeling problems it becomes crucial to impose conditions for symmetry in order to stand a chance of reproducing symmetry-related phenomena. Sufficient conditions are given for MLP and RBF networks to have a set of fixed points that are symmetrical with respect to the origin of the phase space. In the case of MLP networks such conditions reduce to the absence of bias parameters and the requirement of odd activation functions. This turns out to be important from a dynamical point of view since some phenomena are only observed in the context of symmetry, which is not a structurally stable property. The results are illustrated using bench systems that display symmetry, such as the Duffing-Ueda oscillator and the Lorenz system.

Minimum description length neural networks for time series prediction

Artificial neural networks (ANN) are typically composed of a large number of nonlinear functions (neurons) each with several linear and nonlinear parameters that are fitted to data through a computationally intensive training process. Longer training results in a closer fit to the data, but excessive training will lead to overfitting. We propose an alternative scheme that has previously been described for radial basis functions (RBF). We show that fundamental differences between ANN and RBF make application of this scheme to ANN nontrivial. Under this scheme, the training process is replaced by an optimal fitting routine, and overfitting is avoided by controlling the number of neurons in the network. We show that for time series modeling and prediction, this procedure leads to small models (few neurons) that mimic the underlying dynamics of the system well and do not overfit the data. We apply this algorithm to several computational and real systems including chaotic differential equations, the annual sunspot count, and experimental data obtained from a chaotic laser. Our experiments indicate that the structural differences between ANN and RBF make ANN particularly well suited to modeling chaotic time series data.