WorksComputer Science
A Fast Algorithm for the Finite Expression Method in Learning Dynamics on Complex Networks
Abstract
Complex network data is prevalent in various real-world domains, including physical, technological, and biological systems. Despite this prevalence, predicting trends and understanding behavioral patterns in complex systems remain challenging due to poorly understood underlying mechanisms. While data-driven methods have advanced in uncovering governing equations from time series data, efforts to extract physical laws from network data are limited and often struggle with incomplete or noisy data. Additionally, they suffer from computational costs on network data, making it difficult to scale to real-world networks. To address these challenges, we introduce a novel approach called the Finite Expression Method (FEX) and its fast algorithm for learning dynamics on complex networks. FEX represents dynamics on complex networks using binary trees composed of finite mathematical operators. The nodes within these trees are trained through a combinatorial optimization process guided by reinforcement learning techniques. This unique configuration allows FEX to capture complex dynamics with minimal prior knowledge of the system and a small dictionary of mathematical operators. We also integrate a fast, stochastic algorithm into FEX, reducing the computational complexity from to . Our extensive numerical experiments demonstrate that FEX excels in accurately identifying dynamics across diverse network topologies and dynamic behaviors.
Introduction
Complex network data is prevalent in various real-world domains, spanning from transportation [52, 76, 87, 89] to biology [16, 9, 51] and epidemiology [57, 20, 65]. For instance, proteins frequently interact with one another to execute a wide range of functions. As with many complex systems, comprehending the underlying dynamics within these networks is crucial for analyzing and predicting their behaviors. In practice, time-series data for individual nodes within a network are often accessible. Still, extracting the dynamic interactions among these nodes to gain insights into the system’s behavior proves to be a challenging endeavor. Additionally, the interaction dynamics among particles in the network significantly increase the computational expense of identifying physical laws. Consequently, the physical laws governing the majority of real-world complex networks remain elusive and underexplored.
In recent years, machine learning has emerged as a promising avenue for data-driven discovery of physical laws. Notably, deep learning has demonstrated its potential to identify governing partial differential equations (PDEs) from noisy data [14, 34, 58, 60, 56, 80, 45, 46, 22, 67, 3, 6, 34, 17, 18, 69, 28, 48, 84, 39, 7, 63, 71, 74, 12, 15, 68, 75, 13, 59]. The success of these methods is attributed to the expressive power of deep learning models in capturing relationships among different variables. However, deep learning methods suffer from the black-box nature of their solutions, making it difficult to interpret and explicitly identify the governing equation in mathematical form. To address this, recent works have explored symbolic regression techniques to infer governing equations, thereby obtaining an explicit mathematical representation of the system [83, 49, 79, 41, 35]. Nonetheless, there has been limited focus on the discovery of physical laws within complex networks. Discovering dynamics on network data involves various challenges:
Complexity and scale: Real-world network data can be extremely large and complex, making analysis computationally expensive. In the full interaction model with nodes in the network, each node interacts with every other particle, leading to interaction terms that need to be computed at each time, resulting a complexity.
Data quality and availability: The quality of network data can vary greatly. Issues like missing data, noise, and topological inaccuracies can significantly impact the analysis.
In Subsection 1.1, we summarize some of the recent work on learning dynamics from network data, and we also introduce some of the recent progress on fast algorithms. In Subsection 1.2, we introduce the overall idea of our proposed method and how it tackles these challenges.
Related work
In this Subsection, we introduce related work on identifying governing equations on network data and fast algorithms for scientific computing.
Discovering physical laws on network
SINDy: The Sparse Identification of Nonlinear Dynamics (SINDy) [10] algorithm aims to extract underlying governing equations from data in an interpretable and efficient way, with theoretical convergence analysis [90]. Taking time-series data as inputs, SINDy infers the underlying physical laws of the system by using Sequentially Thresholded Least Squares algorithm to choose a parsimonious model from a large library of candidate functions, usually including polynomials and trigonometric functions. SINDy has been used to identify models in fluid dynamics [29], control systems [44], and neuroscience [19], etc. However, it has been shown that the vanilla SINDy performs poorly in identifying dynamics on network data. In addition, SINDy assumes direct access to the time derivative of the dynamics in the candidate function library, which is often unavailable in most real-world applications.
Two-Phase SINDy: In [31], Gao and Yan proposed a two-phase procedure consisting of global regression and local fine-tuning for graph dynamics inference. This method works by building two comprehensive function libraries, and , for self and interaction dynamics. Then, in the first phase, this method performs a sparse regression on the normalized input data and the libraries and to identify potential terms in the network dynamics. In the second phase, it performs topological sampling to fine-tune the coefficients of functions from the first phase. This method can identify the explicit functional forms of the system dynamics, given sufficiently large function libraries. However, users often lack a priori knowledge of the candidate functions, so they assume a prior solid knowledge of the system. In addition, this method is sensitive to noisy and low-resolution data because topological samplings are used in the fine-tuning stage.
Graph Neural Networks: Graph neural networks (GNNs) are popular models to learn the dynamics on the networks and make predictions on the systems [64, 88, 27, 30]. These models typically use an encoder neural network to denote node states and create a latent representation. Following this, Neural Ordinary Differential Equations (Neural ODEs) are applied on the latent space to learn the dynamics. Finally, a decoder is used to project the latent embedding back to the original high-dimensional state. While this approach is flexible enough to be applied to a wide variety of network data, it essentially functions as a “black box”, making it difficult to understand the underlying dynamics. Additionally, it struggles to generalize well to out-of-distribution data.
ARNI: Algorithm for Revealing Network Interactions (ARNI) [11] is a model-free method originally proposed to identify direct interactions of particles in a network data. It works by decomposing the dynamics of each node as a linear combination of basis functions in pairwise and higher-order interactions with other nodes in the system, and recast the reconstruction problem into a mathematical regression problem with grouped variables. Authors of [31] modified ARNI appropriately to make it capable of identifying network dynamics instead of only inferring the graph topology.
Fast algorithms for scientific computation
In the realm of physical sciences, it is important to develop fast and efficient algorithms to handle complex systems while keeping a high accuracy. For example, Jin et al. [43] proposed the Random Batch Methods (RBM), which splits particles in a physical system into several batches at each time to reduce the computational cost from to , where is the number of particles. In machine learning, stochastic gradient descent (SGD) [2, 8, 38] takes a random sample of the entire dataset in each iteration to compute the gradient and update model parameters. Similar to SGD, stochastic coordinate descent [54, 55] randomly samples some of the coordinates to update at each time step. In addition, randomized algorithms have been widely applied to fast matrix factorizations [23, 37, 62] and Bayesian inference [86, 61], etc. In spirit, our work is similar to RBM; however, instead of taking a random batch of particles for the evolution of the dynamics, we use a stochastic approach to select a random batch of particles at each step for discovering the physical laws, i.e., solving an optimization problem.
Our contributions
To tackle the problems seen in previous methods, we propose the finite expression method (FEX) that represents dynamics on network data by binary trees, whose nodes are mathematical operators (e.g., sin, cos, exp, etc). FEX selects the operator of each node via combinatorial optimization (CO) by a novel reinforcement learning (RL) approach. In particular, our method has the following advantages: (1) Compared to the widely used SINDy approach and its variants, FEX does not require a large library of candidate functions. In contrast, thanks to the binary tree representation, FEX can produce complicated composition of functions with a few simple mathematical operators provided by the user. In addition, unlike SINDy, FEX directly estimates the time derivatives from time series data without direct access to the derivatives. This is a more challenging task due to the difficulty of estimating derivatives from noisy time series data. (2) Our model employs a RL approach to perform operator selection in a data-driven manner (discussed in § 3). This design has been demonstrated to promote the expressiveness of our method of identifying network dynamics. (3) To mitigate the intensive computational burden when calculating the pairwise interaction dynamics on network data, we utilize a stochastic algorithm to accelerate the computation while keeping a reasonable accuracy. Throughout this work, we demonstrate the advantage of our method by applying it to several synthetic complex network data to extract governing equations. An overview figure of FEX to infer physical laws on network data is provided in Figure 1.

Figure 1. (A) Based on the input data of the time series and adjacency matrix, FEX performs combinatorial optimization to infer the mathematical structures of the dynamics and then implements fine-tuning on the top-performing candidates in the pool to identify the optimal dynamics. (B) FEX implements combinatorial optimization by training a controller to output probability mass function for each node in the FEX binary tree.
Problem setup
Learning dynamics on complex networks
A common model [31, 53] that describes the evolution of the states of the network of size is given by
where represents node ’s -dimensional feature and . The dynamics of are driven by , denoting the self-dynamics, and , representing the interaction dynamics between any pair of nodes and . denotes the -th entry in the adjacency matrix , where if there is a connection between node and node , and otherwise. Therefore, the dynamics are determined by self-dynamics, interaction dynamics and the topology of the network. However, in many real-world complex network systems, we do not have the information of the explicit forms of and . Therefore, it is important to distill the functional forms of and from node activity data, which is the focus of our work.
However, inferring underlying dynamics from real-world time series data presents a formidable challenge, primarily due to the often unclear or only partially known governing laws. The endeavor to discover dynamics within a complex network system introduces an additional layer of complexity, attributable to the interactions among entities described by the graph topology. Moreover, numerous oscillator systems exhibit varied coupling behaviors contingent upon the coupling strength between entities and the inherent graph topology, a prime example being synchronization phenomena. In addition, the data in real-world are often sparse, noisy, and come from incomplete network topology. Hence, to demonstrate the capability of FEX to discover dynamics on complex networks, we subject FEX to a rigorous examination against three classical network dynamics: Hindmarsh-Rose (HR, ) [40, 78, 85], FitzHugh-Nagumo (FHN, ) neuronal systems [24], and coupled heterogeneous Rössler oscillators () [5], where denotes the dimension of feature for each node. Regarding to the graph topology, we demonstrate that FEX is capable of identifying governing equations on various synthetic networks, including Erdős-Rényi (ER) [33] and scale-free (SF) [1] networks. Through this extensive examination, we aim to demonstrate the effectiveness and robustness of FEX in decoding the intricate dynamics inherent in complex network systems. Details are elaborated more in § 4.
FEX for learning dynamics on networks
The motivation behind FEX is to develop a robust methodology for identifying physical laws from network data. Our model depicts network dynamics using two fixed-size binary trees: one representing the self-dynamics while another one representing the interaction dynamics . Each tree node denotes a simple mathematical operator—either unary (sin, cos, and exp ...) or binary (, , ...). The aim of the learning process is to identify the most suitable mathematical operator for each node of the tree, ensuring accurate representation of the dynamics. The input to FEX algorithm consists of time series data for each node in the network, and the adjacency matrix representing the network topology, as illustrated by Figure 2.
Figure 2. Representation of the components of our implementation with FHN dynamics. (A) In the offline stage, we generate time series data according to the governing law of the system, potentially with noise or low resolution. (B) In the online phase, based on the time series data, we use two FEX binary trees to infer the dynamics of the system.
Learning dynamics from sparse and noisy data
In real-world scenarios, data are inevitably contaminated by noise or by the incompleteness of network topology. Thus, evaluating the robustness of the proposed method across diverse scenarios—such as low resolution, observational noises, and spurious or missing links—is paramount. Moreover, a comparative analysis with baseline models is conducted to ascertain the relative efficacy of the proposed method.
Low resolution: Practical constraints or technological limitations often result in the collection of sparse raw data. To emulate such circumstances, we simulate nonlinear dynamics with a time step size of , followed by a uniform down-sampling of the time series data. The outcomes reveal that FEX retains a high degree of accuracy, i.e., relative errors of coefficients identified by FEX and the true coefficients are less than 1%, even with merely 5% or 10% of the original data.
Spurious and missing links: In real-world applications, the evolving network topology makes capturing all node interactions a challenging task. To assess the robustness of FEX against incomplete topology, we randomly either add or remove a proportion of links from the authentic network topology. The results indicate that FEX can accommodate approximately 15% missing or spurious links, showcasing its robustness in handling topological inaccuracies.
Observational noise: Observational noise is another common issue in data measurement. We replicate such noise by introducing Gaussian noises to the time series data for every node within the network. The noise intensity is modulated via the signal-to-noise ratio (SNR). The findings underscore that FEX sustains a high accuracy level, even amidst a 40-dB observational noise environment. In contrast, other baseline methods such as Two-Phase SINDy [31] and ARNI [11] identified erroneous and redundant terms with low-accuracy coefficients.
Through this thorough evaluation across varied scenarios and a comparative analysis with baseline models, we aim to demonstrate the robustness and stability of FEX in extracting dynamics from real-world, noisy data with potential topological inaccuracies. Further details are provided in Supplementary Information.
Methods
Overview of FEX
FEX [50, 77] introduces a novel strategy for solving PDEs and identifying physical laws from data [42]. FEX represents the system’s dynamics using a learnable binary tree to identify dynamics on network data. This tree comprises unary operators (e.g., sin, cos, etc.) and binary operators (+, −, ×, etc.). Additionally, each unary operator incorporates learnable weights and biases, augmenting the expressiveness of the expression. Given a specific set of operators in the tree, input variables are passed through the FEX tree via the leaves. The complete FEX expression is then derived through a preorder traversal of the tree. Unlike many symbolic regression (SR) methods that generate equations autoregressively, FEX allows users to predetermine the tree’s depth and size. This distinctive feature has been demonstrated to enhance FEX’s accuracy in solving PDEs and discovering governing equations. Subsequently, with a suitable loss function, FEX conducts CO using a RL strategy to perform operator selection, and continuous optimization to refine the parameters associated with unary operators.
To address this mixed optimization problem, FEX employs a CO technique to determine the best selection of operators while leveraging a continuous optimization method to finetune the corresponding parameters of tree nodes. Regarding the CO, FEX utilizes a RL strategy to transform CO into a continuous optimization over probability distributions. To achieve this transformation, a controller network (modeled as a fully connected neural network) is integrated into FEX, which produces the probability mass function for each operator’s selection at every tree node. Consequently, selecting the most suitable operators for tree nodes is recast as the problem of training the optimal controller network capable of sampling top-performing operators. The refinement process of identifying the best controller is fundamentally a continuous optimization task. The controller network’s parameters are refined through policy gradient techniques to maximize the anticipated reward, consistent with RL conventions. As a result, the best operators can be efficiently identified by sampling from the probability distribution of the learned controller network.
In particular, analyzing dynamics in complex networks brings forth several challenges, primarily the computational complexity involved. Computing particle interactions on a graph requires complexity, where is the number of nodes in a network. This is prohibitive for large-scale systems. We utilize a stochastic algorithm when calculating the loss function to mitigate this. Simply put, nodes in the network are grouped into random batches and particle interactions are limited to their respective batches. With this strategy, we achieve faster computational speeds, i.e., computational complexity, without sacrificing the precision required to identify the accurate structure of the FEX tree.
Functional space of finite expressions
FEX models the dynamics of network data within a functional space characterized by a finite set of operators. As such, it is crucial to clearly define this functional space where the solution resides.
Definition 3.1 (Mathematical expression [50]). A mathematical expression is a combination of symbols, which is well-formed by syntax and rules and forms a valid function. The symbols include operands (variables and numbers), operators (e.g., “+”, “sin”, integral, derivative), brackets, and punctuation.
Definition 3.2 (-finite expression [50]). A mathematical expression is called a -finite expression if the number of operators in this expression is .
Definition 3.3 (Finite expression method). The finite expression method aims to numerically capture dynamics by using a limited size expression, so that the produced function closely matches the actual dynamics.
We denote the functional space that consists of functions formed by finite expressions with the number of operators less than or equal to .
The combinatorial optimization problem in FEX
In FEX, the loss functional varies based on the problem. For this study, we adopt the least squares loss functional on each dimension of the dynamics, as presented in [31], which is described as follows:
where is the adjacency matrix of the network, denotes the self-dynamics modeled by a FEX tree parameterized by , denotes the interaction dynamics by another FEX tree parameterized by . In FEX, the solution is found by solving the combinatorial optimization problem
where the solution space will be elaborated in § 3.4.1.
Implementation of FEX
As mentioned previously, FEX’s computational process begins with the formation of a mathematical binary tree, where each node represents either a unary or a binary operator. The candidate solution is derived from evaluating the function represented by this tree. Subsequently, the CO denoted by (3) is employed to dynamically choose the best operators for all tree nodes. This optimization aims to find operators that recover the structure of the genuine network dynamics.
Finite expressions with binary trees
FEX uses a binary tree structure to represent finite expressions. The user defines the sets of potential unary and binary operators, represented by and . Common choices of unary and binary operators are respectively,
Every unary operator in the leaf level operates on the inputs in an element-wise manner, followed by scaling parameters , where , and a bias parameter . For instance:
All parameters associated with unary operators are collectively represented by . The complete expression is derived through a preorder traversal of the operator sequence, , in the binary tree . Consequently, this finite expression is represented as , as a function of input . Given a fixed , the maximum number of operators remains constrained and is upper-bounded by a constant, labeled . In FEX, the functional space wherein the dynamics of the system is solved is defined as:
The computation within the binary tree proceeds recursively, beginning at the leaf nodes. Unary operators at these leaf nodes act on the input in an element-wise manner, with the scaling factor , transforming the dimension from to . Subsequently, computations proceed in a bottom-up recursive pattern until reaching the tree’s root in the top level. A representation of the computation flow of such tree is presented in Figure 3.
Figure 3. Computational rule of a FEX binary tree. Each unary operator is equipped with a weight and bias to enhance expressiveness. For tree depth greater than one (i.e., ), the computation is implemented by recursion.
Workflow of FEX
The FEX solution represented as is derived by tackling the CO problem as:
To achieve this, FEX adopts a two-phase approach:
Operator sequence optimization: The primary goal here is to refine the operator sequence such that it reflects the structure of the authentic dynamics.
Parameter optimization: Following the sequence refinement, FEX then optimizes the parameter set aiming to minimize the functional described in (4). The FEX framework is organized into four components:
Score Computation: To discover the solution’s structure, FEX deploys a mix-order optimization algorithm. This step assesses the score of the operator sequence . A stochastic algorithm is utilized to accelerate the score’s calculation over the network data.
Operator Sequence Generation: Here, a neural network (NN) is leveraged to conceptualize a controller. This controller subsequently outputs a probability mass function, which is used for sampling optimal operator sequences.
Controller Update: This step refines the parameters of the controller based on the reward outcomes of the generated operator sequences, guiding it to yield more effective sequences over training iterations.
Candidate Optimization: FEX dynamically keeps a pool of the operator sequences with the highest rewards during the search process. Upon completion of training, each sequence candidate undergoes a fine-tuning phase. The ultimate objective is to identify the best operator sequence as the approximation of system dynamics.
Subsequent sections will elaborate on the details of each of these components.
Score computation
The evaluation of an operator sequence through its score is pivotal during training. This score guides the controller’s parameter updates, aiming to produce optimal probability mass functions for the sampling of efficient operators to approximate the network dynamics. The score of , denoted as , is defined as follows:
For an efficient computation of the score , we employ a hybrid mix-order optimization strategy to update the parameter . We first take iterations using a first-order optimization algorithm, such as stochastic gradient descent [72] or Adam [47], and this yields . will be used as a warm start, followed by a iterations of a second-order optimization method, such as Newton’s method [4] or BFGS [25], resulting in .
Stochastic-FEX To address the computational burden when performing the total optimization steps, we propose to approximate the interaction terms using only a subset of the particles at each step. In particular, we divide the particles into batches and only consider the interactions within each batch, then the number of interactions that need to be computed is significantly reduced. Assuming that each batch contains approximately nodes (), then the number of interactions to be computed for each batch is on the order of . Since there are such batches, the total number of interactions at each step is on the order of , which is since is a constant independent of . This significantly reduces the computational cost when evaluating the loss function during the steps of optimization.
Upon completion of these steps, the score associated with the operator sequence is denoted as:
We have the following theorem on the convergence of Stochastic-FEX within this phase. The proof is provided in the Supplementary Information.
Theorem 1. Fix some tolerance level , and integer . Let be a regular minimizer of defined in (2), and suppose that Stochastic-FEX is run with a step-size schedule of the form for some and large enough . Furthermore, assume that the following bounds hold:
, where denotes the -th component of the vector.
Then:
There exist neighborhoods and of such that, if , the event
occurs with probability at least .
Conditioned on , we have
Operator sequence generation
The primary objective of the controller is to generate operator sequences that achieve high scores throughout the training process. We denote the controller as , which is parameterized by a fully connected neural network with parameters . Given an operator sequence comprising nodes, the controller produces probability mass functions for . The operator is then sampled based on the probability mass function . To encourage exploration within the set of operators, the -greedy strategy [81] is implemented. To be more specific, the operator is selected from a uniform distribution over the entire operator set with a probability of .
Controller update
As previously discussed, we introduce a controller represented by the neural network , to generate a probability mass function for each node in the binary tree. This transformation changes the problem from a CO problem into a continuous optimization one. The primary goal now becomes the optimization of the controller parameters so that the controller is capable of producing optimal probability mass function for each node. To achieve this, we consider the expected value of scores of operator sequences sampled from the controller , i.e.,
The derivative of (7) with respect to is
where is the probability of the sampled . To efficiently approximate the expectation (8), we sample a batch of operator sequences each time and compute
where is the batch size. Subsequently, the model parameter is updated using gradient ascent, denoted as . Nonetheless, the practical aim is to identify the operator sequence with the highest score, rather than optimizing the average scores of all generated operator sequences. Therefore, we follow the approach in [66] and consider
where represents the -quantile of the score distribution generated by . In a discrete form, the gradient computation becomes
where denotes an indicator function which evaluates to 1 when the condition is met, and 0 otherwise. Meanwhile, represents the -quantile of the scores within the set .
Candidate optimization
During the training phase of optimizing the controller parameters , efficient evaluation of expressions sampled from the controller is crucial. As introduced in § 3.4.3, this is achieved by computing the score through “coarse-tune” iterations. However, this coarse score might not reflect the performance of operator sequence owing to the optimization of a nonconvex function. Therefore, we maintain a pool with a fixed size of during the training of controller, which dynamically holds the top-performing candidate expressions. When the controller’s training is completed, we further refine the coefficients of each candidate expression within the pool to attain high accuracy. To enhance the robustness of our method against link perturbations, we employ a random sampling strategy. In particular, for each candidate function, we perform random sampling of times, where we randomly sample nodes from the graph each time. For each batch, we perform iterations of optimizations based on the objective function with a first-order algorithm with a small learning rate, and the final coefficients of the expression are obtained by averaging the coefficients over times.
Dynamics learning results
In this section, we present the results of identification of HR, FHN, and coupled Rössler dynamics with FEX on the Scale-Free (SF) network with clean data. The set up of SF network follows the description in Supplementary Information. We choose SF network since it demonstrates more complex phenomena than the ER model, and many real-world networks have scale-free property. During the offline stage, we generate the time series data with the true dynamics (ODEs). During the inference stage, we use FEX only with the time series data to identify the governing equations of the system, without accessing to the genuine ODEs. We emphasize that we only have access to the time series data and approximate the state time derivatives using numerical derivatives. This contrasts with the SINDy method, which requires direct access to the time derivatives of the states. Throughout all numerical experiments, we set the batch size in stochastic-FEX as 32. In addition, we provide the depths of FEX trees used in each of the considered dynamics in the numerical examples in Table 1. Robustness analyses of FEX and other baseline methods, aimed at extracting dynamics from low-resolution data, networks with spurious or missing links, and noisy data, are discussed in the Supplementary Information.
| Dynamics | Depth of Tree |
| HR | 4 |
| FHN | 3 |
| Coupled Rössler | 3 |
Table 1. Tree depth of FEX tree for each dynamics.
Numerical derivatives and error metric
To approximate the time-varying derivative of each node, we apply the five-point approximation [73]
where is the time step.
To evaluate the accuracy of inferred dynamics, we use the metric of symmetric mean absolute percentage error (sMAPE) [26]:
where is the total number of terms of both true and inferred functions, is the inferred coefficient and is the true coefficient. sMAPE is an appropriate metric in evaluating the performance of baselines since it not only captures the coefficient discrepancies but also penalizes incorrect inferred terms. sMAPE ranges from 0 to 1, and the lower the value the higher the accuracy, and vice versa.
Hindmarsh-Rose dynamics
The HR model [40, 78, 85] was introduced by Hindmarsh and Rose in the early 1980s to describe neuronal activity. The HR model is a simplified model of Hodgkin-Huxley model [36] while still reproducing many of the dynamic behaviors observed in real neurons. The governing equations of HR model is given as
where the coupling term is
Here is the membrane potential of neuron , is the transport rate of ions across the membrane through the ion channels, and is the adaptation current. Parameters are set as: , , , , , , , , , and is external current. Notice that the dynamic behavior of the HR neuron can be adjusted by varying these parameters.
By doing so, one can observe various neuronal activities, such as tonic spiking, phasic spiking and chaotic bursting. We generate the time series data with Equation (14), with the adjacency matrix of SF network, terminal time and time step .
Applying FEX to the neuronal activities data generated by HR dynamics on a directed SF network, we obtain the inferred equations as
From Eq. (15), we observe that FEX not only identifies all the correct terms in the dynamics in each dimension of the node, but also learns the coefficient of each term with relative errors lower than 1%. The neuronal activities generated by FEX and true governing equations are shown in Figure 4.

Figure 4. True HR dynamics and the dynamics identified by FEX.
| True Dynamics | Two-Phase | ARNI | FEX | |
| None | ||||
| None | None | |||
Table 3. Numerical results for Hindmarsh-Rose dynamics (SNR = 40) in self-dynamics and interaction dynamics term by term.
Observational noises
In this subsection, we demonstrate the robustness of our method against observational noises. We focus on the HR dynamics on a directed SF network. We generate noisy data through
where is observational noise, which follows a Gaussian distribution with zero mean and standard deviation one, so the parameter is used to tune the level of noise.
From Table 2, 3 and 4, FEX demonstrates its capability to accurately identify all terms within the HR dynamics with their respective coefficients with SNR = 45 and 40. The Two-Phase method can indeed infer the correct terms in the dynamics in the case of SNR = 45. However, the accuracy of inferred coefficients is low, particularly in the interaction dynamics. In addition, in the case of SNR = 40, Two-Phase method fails to infer the term in the interaction dynamics. ARNI can infer most of the terms correctly when SNR = 45, with an extra term in the self-dynamics. Furthermore, the coefficients in the self-dynamics are poorly estimated, but it is interesting to notice that the coefficients of interaction-dynamics are more accurate than those of Two-Phase, owing again to ARNI’s capability to infer the network structure. As for SNR = 40, ARNI infers more redundant terms in self-dynamics, and it fails to identify both true terms in the original dynamics, and incorrectly identifies a term. This shows that ARNI can reasonably well infer most of the dynamics terms in low-noise case, while it fails to work as noise level increases. Therefore, FEX has demonstrated its superior performance compared to Two-Phase and ARNI to handle noisy data.
| True Dynamics | Two-Phase | ARNI | FEX | |
Table 2. Numerical results for Hindmarsh-Rose dynamics (SNR = 45) in self-dynamics and interaction dynamics term by term.
| SNR | Two-Phase | ARNI | FEX |
| 45 | 0.0924 | 0.2720 | 0.0158 |
| 40 | 0.2190 | 0.6252 | 0.0376 |
Table 4. sMAPE of HR dynamics with different levels of noise.
FitzHugh-Nagumo dynamics
We further test our method to the neuronal activities data generated by FHN dynamics [24]. The FHN model is another mathematical representation used to describe the spiking behavior of neurons. It is defined by a system of two ordinary differential equations, capturing the main features of excitability in nerve membrane dynamics. The equations governing the FHN neuronal network dynamics are
where the first component represents the membrane potential containing self and interaction dynamics, is the in-degree of neuron (representing the number of incoming connections to node ), and . The second component represents a recovery variable where , and . We generate the time series data with Equation (16), with the adjacency matrix of SF network, terminal time and time step .
The equations inferred by our approach from the neuronal activities data generated on a directed SF network are shown below:
The trajectories generated by physical laws inferred by FEX and the true governing equations are shown in Figure 5. FEX has shown to accurately infer both dimensions of FHN system and coincide closely with the true dynamics.
Figure 5. True FHN dynamics and the dynamics identified by FEX.
We compare the robustness of FEX against spurious and missing links using FHN dynamics, and evaluate it alongside Two-Phase and ARNI methods. FEX demonstrates superior accuracy compared to the baselines, even with random addition or deletion of 15% of the links. Detailed results are provided in the Supplementary Information.
Coupled Rössler dynamics
We also consider the coupled Rössler oscillators [70, 82], which is a classical model often used to study chaotic dynamics and synchronization phenomena in complex networks. We generate chaotic activities data according to the true equations governing heterogeneous Rössler dynamics [5, 75],
where coupling strength , , , and . The natural frequencies of the oscillators, denoted by , follow a normal distribution with a mean value of 1 and a standard deviation of 0.1. We generate the time series data with Equation (17), with the adjacency matrix of SF network, terminal time and time step .
The equations inferred by FEX from the data generated on a directed SF network with are
Note that the coefficient -1.0093 of in the first equation of Eq. (18) and 0.9909 of in the second equation of Eq. (18) are estimated average value of nodes’ natural frequencies, otherwise inferring a distinctive frequency for each node would lead to an -fold increase in the dimensionality of model space. The trajectories generated by physical laws inferred by FEX and the true governing equations are shown in Figure 6.
Figure 6. True coupled Rössler dynamics and the dynamics identified by FEX.
Utilizing Coupled Rössler dynamics, we further assess the robustness of FEX against low-resolution observational data. We compare the performance of FEX with Two-Phase and ARNI methods when observing only 10% and 5% of the time series data. FEX demonstrates superior accuracy in identifying the dynamics within this low data regime compared to the other baselines. Detailed results are provided in the Supplementary Information.
Details of performance of stochastic-FEX
As introduced before, computation on a large complex network is demanding or even infeasible. Therefore, we incorporate a stochastic algorithm into the FEX algorithm to accelerate the identification of the optimal mathematical structure of network dynamics while maintaining a high accuracy.
To demonstrate the efficiency, we create a SF model of network size 64, 128, 256, and 512, respectively. Then, we use the time series data of FHN model on the four SF networks, and implement FEX to infer the dynamics. We record the clock time of computation of coarse-tune in the FEX training algorithm by full interaction method and stochastic-FEX (with a batch size of 32), respectively. The log-log plot of time vs. number of particles is shown in Figure 7. We observe that the fitting line clock time of the full interaction model on different network sizes has a slope of approximately 2. The stochastic-FEX’s slope is roughly 1, indicating that the stochastic-FEX demonstrates a linear scaling performance.
Figure 7. Clock time comparisons of full interaction model and stochastic-FEX. The -axis shows 64, 128, 256, and 512 particles of the SF network, respectively. The -axis shows the clock time of one iteration computation for each network size.
Furthermore, we comment that we only use this stochastic algorithm in the CO part of FEX to identify the optimal structure of the dynamics. The CO stage is the computational bottleneck of FEX, and we have demonstrated its capability to determine the correct structure of dynamics while accelerating the computation significantly. After the structures of top-performing candidates are learned, we use the full interaction model in the iterations of coefficients fine-tuning stage to achieve high accuracy of dynamics inferred by FEX. The computational burden at this stage is manageable since we only have to fine-tune the coefficients of the top candidates in the pool .
Conclusion
In this work, we proposed FEX and its fast algorithm as a novel methodology to learn the dynamics on complex networks. FEX has shown its capability to accurately discover physical laws on various synthetic networks, even with noise and low resolution of data. Furthermore, we incorporate a stochastic algorithm into our proposed FEX framework, which scales up FEX to handle large data when the network size increases significantly. Therefore, we have demonstrated that FEX has the potential to identify real-world dynamics in an interpretable and reliable way.
There are also questions that remained to be addressed. First, throughout our work, we consider only the pairwise interaction dynamics in the physical laws. However, it is common for complex systems to exhibit three-point or higher-order couplings among nodes in the network. Yet, it is straightforward to adapt such a setting into the FEX methodology, with the extra computational cost of including more binary trees to represent higher-order dynamics. Second, dynamics often demonstrate implicit stochasticity in real-world data, which can be described by stochastic differential equations (SDEs) [21, 32]. Lastly, this work focuses on classical dynamics with an explicit form of governing equations. However, dynamics are largely unknown or underexplored in most situations. Hence, it is interesting to observe how accurately FEX can approximate empirical dynamics and the insights it provides from inferring complex network dynamics.
Acknowledgments C. W. was partially supported by National Science Foundation under award DMS-2206332. H. Y. was partially supported by the US National Science Foundation under awards DMS-2244988, DMS-2206333, the Office of Naval Research Award N00014-23-1-2007, and DARPA D24AP00325-00.
Competing interests The authors declare no competing interests.
References
- [1]Réka Albert and Albert-László Barabási. Statistical mechanics of complex networks. Reviews of modern physics, 74(1):47, 2002.
- [2]Shun-ichi Amari. Backpropagation and stochastic gradient descent method. Neurocomputing, 5(4-5):185–196, 1993.
- [3]Hassan Arbabi, Judith E Bunder, Giovanni Samaey, Anthony J Roberts, and Ioannis G Kevrekidis. Linking machine learning with multiscale numerics: Data-driven discovery of homogenized equations. Jom, 72:4444–4457, 2020.DOI
- [4]Mordecai Avriel. Nonlinear programming: analysis and methods. Courier Corporation, 2003.
- [5]Baruch Barzel, Yang-Yu Liu, and Albert-László Barabási. Constructing minimal models for complex system dynamics. Nature communications, 6(1):7186, 2015.
- [6]Tom Bertalan, Felix Dietrich, Igor Mezić, and Ioannis G Kevrekidis. On learning hamiltonian systems from data. Chaos: An Interdisciplinary Journal of Nonlinear Science, 29(12), 2019.
- [7]Josh Bongard and Hod Lipson. Automated reverse engineering of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 104(24):9943–9948, 2007.
- [8]Léon Bottou. Large-scale machine learning with stochastic gradient descent. In Proceedings of COMPSTAT’2010: 19th International Conference on Computational StatisticsParis France, August 22-27, 2010 Keynote, Invited and Contributed Papers, pages 177–186. Springer, 2010.
- [9]Cameron P Bracken, Hamish S Scott, and Gregory J Goodall. A network-biology perspective of microrna function and dysfunction in cancer. Nature Reviews Genetics, 17(12):719–732, 2016.DOI
- [10]Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the national academy of sciences, 113(15):3932–3937, 2016.DOI
- [11]Jose Casadiego, Mor Nitzan, Sarah Hallenberg, and Marc Timme. Model-free inference of direct network interactions from nonlinear collective dynamics. Nature communications, 8(1):2192, 2017.DOI
- [12]Kathleen Champion, Bethany Lusch, J Nathan Kutz, and Steven L Brunton. Data-driven discovery of coordinates and governing equations. Proceedings of the National Academy of Sciences, 116(45):22445–22451, 2019.
- [13]Haibin Chang and Dongxiao Zhang. Identification of physical processes via combined data-driven and data-assimilation methods. Journal of Computational Physics, 393:337–350, 2019.DOI
- [14]Sheng Chen, Stephen A Billings, and PM Grant. Non-linear system identification using neural networks. International journal of control, 51(6):1191–1214, 1990.
- [15]Yuan Chen and Dongbin Xiu. Learning stochastic dynamical system via flow map operator. Journal of Computational Physics, 508:112984, 2024.
- [16]Dong-Yeon Cho, Yoo-Ah Kim, and Teresa M Przytycka. Chapter 5: Network biology approach to complex diseases. PLoS computational biology, 8(12):e1002820, 2012.DOI
- [17]Victor Churchill and Dongbin Xiu. Learning fine scale dynamics from coarse observations via inner recurrence. Journal of Machine Learning for Modeling and Computing, 3(3), 2022.arxiv.org/abs/2206.01807
- [18]Victor Churchill and Dongbin Xiu. Flow map learning for unknown dynamical systems: overview, implementation, and benchmarks. Journal of Machine Learning for Modeling and Computing, 4(2), 2023.arxiv.org/abs/2307.11013
- [19]Bryan C Daniels and Ilya Nemenman. Automated adaptive inference of phenomenological dynamical models. Nature communications, 6(1):8133, 2015.DOI
- [20]Leon Danon, Ashley P Ford, Thomas House, Chris P Jewell, Matt J Keeling, Gareth O Roberts, Joshua V Ross, Matthew C Vernon, et al. Networks and the epidemiology of infectious disease. Interdisciplinary perspectives on infectious diseases, 2011, 2011.
- [21]Gustavo Deco, Edmund T Rolls, and Raulfo Romo. Stochastic dynamics as a principle of brain function. Progress in neurobiology, 88(1):1–16, 2009.DOI
- [22]Felix Dietrich, Alexei Makeev, George Kevrekidis, Nikolaos Evangelou, Tom Bertalan, Sebastian Reich, and Ioannis G Kevrekidis. Learning effective stochastic differential equations from microscopic simulations: Linking stochastic numerics to deep learning. Chaos: An Interdisciplinary Journal of Nonlinear Science, 33(2), 2023.
- [23]Petros Drineas, Ravi Kannan, and Michael W Mahoney. Fast monte carlo algorithms for matrices ii: Computing a low-rank approximation to a matrix. SIAM Journal on computing, 36(1):158–183, 2006.DOI
- [24]Richard FitzHugh. Impulses and physiological states in theoretical models of nerve membrane. Biophysical journal, 1(6):445–466, 1961.
- [25]Roger Fletcher. Practical methods of optimization. John Wiley & Sons, 2013.
- [26]Benito E Flores. A pragmatic view of accuracy measurement in forecasting. Omega, 14(2):93–98, 1986.DOI
- [27]Cornelius Fritz, Emilio Dorigatti, and David Rügamer. Combining graph neural networks and spatio-temporal disease models to improve the prediction of weekly covid-19 cases in germany. Scientific Reports, 12(1):3930, 2022.DOI
- [28]Xiaohan Fu, Lo-Bin Chang, and Dongbin Xiu. Learning reduced systems via deep neural networks with memory. Journal of Machine Learning for Modeling and Computing, 1(2), 2020.
- [29]Kai Fukami, Takaaki Murata, Kai Zhang, and Koji Fukagata. Sparse identification of nonlinear dynamics with low-dimensionalized flow representations. Journal of Fluid Mechanics, 926:A10, 2021.DOI
- [30]Junyi Gao, Rakshith Sharma, Cheng Qian, Lucas M Glass, Jeffrey Spaeder, Justin Romberg, Jimeng Sun, and Cao Xiao. Stan: spatio-temporal attention network for pandemic prediction using real-world evidence. Journal of the American Medical Informatics Association, 28(4):733–743, 2021.
- [31]Ting-Ting Gao and Gang Yan. Autonomous inference of complex network dynamics from incomplete and noisy data. Nature Computational Science, 2(3):160–168, 2022.
- [32]Mikhail Genkin, Owen Hughes, and Tatiana A Engel. Learning non-stationary langevin dynamics from stochastic observations of latent trajectories. Nature communications, 12(1):5986, 2021.DOI
- [33]Jesús Gómez-Gardeñes and Yamir Moreno. From scale-free to erdos-rényi networks. Physical Review E, 73(5):056124, 2006.
- [34]Raul González-García, Ramiro Rico-Martìnez, and Ioannis G Kevrekidis. Identification of distributed parameter systems: A neural net based approach. Computers & chemical engineering, 22:S965–S968, 1998.DOI
- [35]Gary J Gray, David J Murray-Smith, Yun Li, Ken C Sharman, and Thomas Weinbrenner. Nonlinear model structure identification using genetic programming. Control Engineering Practice, 6(11):1341–1352, 1998.DOI
- [36]John Guckenheimer and Ricardo A Oliva. Chaos in the hodgkin–huxley model. SIAM Journal on Applied Dynamical Systems, 1(1):105–114, 2002.
- [37]Nathan Halko, Per-Gunnar Martinsson, and Joel A Tropp. Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions. SIAM review, 53(2):217–288, 2011.
- [38]Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International conference on machine learning, pages 1225–1234. PMLR, 2016.
- [39]John Harlim, Shixiao W Jiang, Senwei Liang, and Haizhao Yang. Machine learning for prediction with missing dynamics. Journal of Computational Physics, 428:109922, 2021.
- [40]James L Hindmarsh and RM Rose. A model of the nerve impulse using two first-order differential equations. Nature, 296(5853):162–164, 1982.
- [41]Hitoshi Iba, Hugo deGaris, and Taisuke Sato. A numerical approach to genetic programming for system identification. Evolutionary computation, 3(4):417–452, 1995.DOI
- [42]Zhongyi Jiang, Chunmei Wang, and Haizhao Yang. Finite expression methods for discovering physical laws from data. arXiv preprint arXiv:2305.08342, 2023.arxiv.org/abs/2305.08342
- [43]Shi Jin, Lei Li, and Jian-Guo Liu. Random batch methods (rbm) for interacting particle systems. Journal of Computational Physics, 400:108877, 2020.
- [44]Eurika Kaiser, J Nathan Kutz, and Steven L Brunton. Sparse identification of nonlinear dynamics for model predictive control in the low-data limit. Proceedings of the Royal Society A, 474(2219):20180335, 2018.DOI
- [45]George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics, 3(6):422–440, 2021.
- [46]Felix P Kemeth, Sergio Alonso, Blas Echebarria, Ted Moldenhauer, Carsten Beta, and Ioannis G Kevrekidis. Black and gray box learning of amplitude equations: Application to phase field systems. Physical Review E, 107(2):025305, 2023.DOI
- [47]Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.arxiv.org/abs/1412.6980
- [48]John H Lagergren, John T Nardini, Ruth E Baker, Matthew J Simpson, and Kevin B Flores. Biologically-informed neural networks guide mechanistic modeling from sparse experimental data. PLoS computational biology, 16(12):e1008462, 2020.DOI
- [49]Mikel Landajuela, Chak Shing Lee, Jiachen Yang, Ruben Glatt, Claudio P Santiago, Ignacio Aravena, Terrell Mundhenk, Garrett Mulcahy, and Brenden K Petersen. A unified framework for deep symbolic regression. Advances in Neural Information Processing Systems, 35:33985–33998, 2022.DOI
- [50]Senwei Liang and Haizhao Yang. Finite expression method for solving high-dimensional partial differential equations. arXiv preprint arXiv:2206.10121, 2022.arxiv.org/abs/2206.10121
- [51]Gipsi Lima-Mendez and Jacques Van Helden. The powerful law of the power law and other myths in network biology. Molecular BioSystems, 5(12):1482–1493, 2009.
- [52]Jingyi Lin and Yifang Ban. Complex network topology of transportation systems. Transport reviews, 33(6):658–685, 2013.
- [53]Bing Liu, Wei Luo, Gang Li, Jing Huang, and Bo Yang. Do we need an encoder-decoder to model dynamical systems on networks? arXiv preprint arXiv:2305.12185, 2023.arxiv.org/abs/2305.12185
- [54]Ji Liu and Stephen J Wright. Asynchronous stochastic coordinate descent: Parallelism and convergence properties. SIAM Journal on Optimization, 25(1):351–376, 2015.
- [55]Ji Liu, Steve Wright, Christopher Ré, Victor Bittorf, and Srikrishna Sridhar. An asynchronous parallel stochastic coordinate descent algorithm. In International Conference on Machine Learning, pages 469–477. PMLR, 2014.
- [56]Yuxuan Liu, Zecheng Zhang, and Hayden Schaeffer. Prose: Predicting operators and symbolic expressions using multimodal transformers. arXiv preprint arXiv:2309.16816, 2023.arxiv.org/abs/2309.16816
- [57]Alun L Lloyd and Steve Valeika. Network models in epidemiology: an overview. Complex population dynamics: nonlinear modeling in ecology, epidemiology and genetics, pages 189–214, 2007.
- [58]Zichao Long, Yiping Lu, and Bin Dong. Pde-net 2.0: Learning pdes from data with a numeric-symbolic hybrid deep network. Journal of Computational Physics, 399:108925, 2019.DOI
- [59]Zichao Long, Yiping Lu, Xianzhong Ma, and Bin Dong. Pde-net: Learning pdes from data. In International conference on machine learning, pages 3208–3216. PMLR, 2018.
- [60]Bethany Lusch, J Nathan Kutz, and Steven L Brunton. Deep learning for universal linear embeddings of nonlinear dynamics. Nature communications, 9(1):4950, 2018.DOI
- [61]Yi-An Ma, Tianqi Chen, and Emily Fox. A complete recipe for stochastic gradient mcmc. Advances in neural information processing systems, 28, 2015.arxiv.org/abs/1506.04696
- [62]Per-Gunnar Martinsson, Vladimir Rokhlin, and Mark Tygert. A randomized algorithm for the decomposition of matrices. Applied and Computational Harmonic Analysis, 30(1):47–68, 2011.
- [63]Daniel A Messenger and David M Bortz. Weak sindy for partial differential equations. Journal of Computational Physics, 443:110525, 2021.
- [64]Charles Murphy, Edward Laurence, and Antoine Allard. Deep learning of contagion dynamics on complex networks. Nature Communications, 12(1):4720, 2021.
- [65]Romualdo Pastor-Satorras, Claudio Castellano, Piet Van Mieghem, and Alessandro Vespignani. Epidemic processes in complex networks. Reviews of modern physics, 87(3):925, 2015.
- [66]Brenden K Petersen, Mikel Landajuela Larma, Terrell N. Mundhenk, Claudio Prata Santiago, Soo Kyung Kim, and Joanne Taery Kim. Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients. In International Conference on Learning Representations, 2021.
- [67]Yorgos M Psarellis, Seungjoon Lee, Tapomoy Bhattacharjee, Sujit S Datta, Juan M Bello-Rivas, and Ioannis G Keverekidis. Data-driven discovery of chemotactic migration of bacteria via machine learning. arXiv preprint arXiv:2208.11853, 2022.arxiv.org/abs/2208.11853
- [68]Tong Qin, Zhen Chen, John D Jakeman, and Dongbin Xiu. Data-driven learning of nonautonomous systems. SIAM Journal on Scientific Computing, 43(3):A1607–A1624, 2021.
- [69]Tong Qin, Kailiang Wu, and Dongbin Xiu. Data driven governing equations approximation using deep neural networks. Journal of Computational Physics, 395:620–635, 2019.
- [70]Michael G Rosenblum, Arkady S Pikowsky, and Jürgen Kurths. Phase synchronization of chaotic oscillators. Physical review letters, 76(11):1804, 1996.
- [71]Samuel H Rudy, Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Data-driven discovery of partial differential equations. Science advances, 3(4):e1602614, 2017.
- [72]David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. nature, 323(6088):533–536, 1986.
- [73]Timothy Sauer. Numerical solution of stochastic differential equations in finance. In Handbook of computational finance, pages 529–550. Springer, 2011.
- [74]Hayden Schaeffer. Learning partial differential equations via data discovery and sparse optimization. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 473(2197):20160446, 2017.DOI
- [75]Michael Schmidt and Hod Lipson. Distilling free-form natural laws from experimental data. science, 324(5923):81–85, 2009.DOI
- [76]Harold Soh, Sonja Lim, Tianyou Zhang, Xiuyu Fu, Gary Kee Khoon Lee, Terence Gih Guang Hung, Pan Di, Silvester Prakasam, and Limsoon Wong. Weighted complex network analysis of travel routes on the singapore public transportation system. Physica A: Statistical Mechanics and its Applications, 389(24):5852–5863, 2010.DOI
- [77]Zezheng Song, Maria K Cameron, and Haizhao Yang. A finite expression method for solving high-dimensional committor problems. SIAM Journal on Scientific Computing, 47(1):C1–C21, 2025.
- [78]Marco Storace, Daniele Linaro, and Enno de Lange. The hindmarsh–rose neuron model: bifurcation analysis and piecewise-linear approximations. Chaos: An Interdisciplinary Journal of Nonlinear Science, 18(3), 2008.DOI
- [79]Fangzheng Sun, Yang Liu, Jian-Xun Wang, and Hao Sun. Symbolic physics learner: Discovering governing equations via monte carlo tree search. arXiv preprint arXiv:2205.13134, 2022.arxiv.org/abs/2205.13134
- [80]Yifan Sun, Linan Zhang, and Hayden Schaeffer. Neupde: Neural network based ordinary and partial differential equations for modeling time-dependent data. In Mathematical and Scientific Machine Learning, pages 352–372. PMLR, 2020.arxiv.org/abs/1908.03190
- [81]Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018.
- [82]Longkun Tang, Xiaoqin Wu, Jinhu Lü, Jun-an Lu, and Raissa M D’Souza. Master stability functions for complete, intralayer, and interlayer synchronization in multiplex networks of coupled rössler oscillators. Physical Review E, 99(1):012304, 2019.DOI
- [83]Silviu-Marian Udrescu and Max Tegmark. Ai feynman: A physics-inspired method for symbolic regression. Science Advances, 6(16):eaay2631, 2020.
- [84]Pantelis R Vlachas, Wonmin Byeon, Zhong Y Wan, Themistoklis P Sapsis, and Petros Koumoutsakos. Data-driven forecasting of high-dimensional chaotic systems with long short-term memory networks. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 474(2213):20170844, 2018.DOI
- [85]X-J Wang. Genesis of bursting oscillations in the hindmarsh-rose model and homoclinicity to a chaotic saddle. Physica D: Nonlinear Phenomena, 62(1-4):263–274, 1993.DOI
- [86]Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688, 2011.
- [87]Zengwang Xu and Robert Harriss. Exploring the structure of the us intercity passenger air transportation network: a weighted complex network approach. GeoJournal, 73:87–102, 2008.
- [88]Chengxi Zang and Fei Wang. Neural dynamics on complex networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 892–902, 2020.
- [89]Massimiliano Zanin, Xiaoqian Sun, and Sebastian Wandelt. Studying the topology of transportation systems through complex networks: handle with care. Journal of Advanced Transportation, 2018, 2018.DOI
- [90]Linan Zhang and Hayden Schaeffer. On the convergence of the sindy algorithm. Multiscale Modeling & Simulation, 17(3):948–972, 2019.




