Neural Network-Based Parameter Estimation for Non-Autonomous Differential Equations with Discontinuous Signals

Hyeontae Jo 1, 2, Krešimir Josić 3, 4, † \dagger, Jae Kyoung Kim 2, 5, † \dagger · 2025

Read Neural Network-Based Parameter Estimation for Non-Autonomous Differential Equations with Discontinuous Signals by Hyeontae Jo 1, 2, Krešimir Josić 3, 4, † \dagger, Jae Kyoung Kim 2, 5, † \dagger (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-2507-06267.txt.

Introduction

Fitting a model to data breaks when the outside signal driving it jumps suddenly, and a 2025 arXiv paper by Hyeontae Jo, Krešimir Josić, and Jae Kyoung Kim fixes the jump first. Their method, HADES-NN, alternates two steps: a neural network smooths the jagged signal, then the smooth version feeds standard parameter fitting.

The estimates came out accurate and tight across test cases. Two came from real biology: the body clock driven by light recorded on wearables, and yeast answering a mating signal. The authors say the trick widens the range of real systems that can be fitted at all.

Catalog record

CREATORHyeontae Jo 1, 2, Krešimir Josić 3, 4, † \dagger, Jae Kyoung Kim 2, 5, † \dagger
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-06267

Browse the full library · About the Tower of Babel

Full text

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

Neural Network-Based Parameter Estimation for Non-Autonomous Differential Equations with Discontinuous Signals Thanks: Submitted to the editors, August 8, 2026. Hyeontae Jo 1, 2, Krešimir Josić 3, 4, † \dagger, Jae Kyoung Kim 2, 5, † \dagger Abstract Non-autonomous differential equations are crucial for modeling systems influenced by external signals, yet fitting these models to data becomes particularly challenging when the signals change abruptly. To address this problem, we propose a novel parameter estimation method utilizing functional approximations with artificial neural networks. Our approach, termed Harmonic Approximation of Discontinuous External Signals using Neural Networks (HADES-NN), operates in two iterated stages. In the first stage, the algorithm employs a neural network to approximate the discontinuous signal with a smooth function. In the second stage, it uses this smooth approximate signal to estimate model parameters. HADES-NN gives highly accurate and precise parameter estimates across various applications, including circadian clock systems regulated by external light inputs measured via wearable devices and the mating response of yeast to external pheromone signals. HADES-NN greatly extends the range of model systems that can be fit to real-world measurements. 1 1 footnotetext: Division of Applied Mathematical Sciences, Korea University, Sejong City, Republic of Korea 2 2 footnotetext: Pioneer Research Center for Mathematical and Computational Sciences, Institute for Basic Science, Daejeon, Republic of Korea 3 3 footnotetext: Departments of Mathematics, and Biology and Biochemistry, University of Houston, Houston, TX, USA 4 4 footnotetext: Department of BioSciences, Rice University, Houston, TX, USA 5 5 footnotetext: Department of Mathematical Sciences, KAIST, Daejeon, Republic of Korea 2 2 footnotetext: Corresponding authors: K. Josić ( [email protected] ), J. K. Kim ( [email protected] ) keywords Non-autonomous differential equations, Discontinuous external signals, Neural networks, Parameter estimation, Non-smooth optimization † † runningheads: Parameter Estimation of Systems with Abrupt Signal / H. Jo, K. Josić, and J. K. Kim AMS 34C60, 92B05, 68T07, 93C15, 65K10 1 Introduction Non-autonomous differential equations are used widely to model the evolution of natural and engineered systems subject to varying inputs and perturbations. Such models can provide a mechanistic understanding of medical and biological systems and predict their behavior (e.g., vaccine strategies in epidemic dynamics [ 1 , 2 ] and cancer treatment approaches [ 3 ] ). Often, systems in nature are subject to abrupt changes in their inputs, such as the light intensity driving circadian pacemakers [ 4 , 5 ] , pheromone signals in the mating response of yeast [ 6 ] , and on-off switches in electrical circuits [ 7 ] . However, these abrupt changes make non-autonomous models difficult to apply in practice. In particular, fitting models with discontinuous inputs to data presents unresolved mathematical and computational challenges. Specifically, the presence of discontinuous inputs results in non-smooth loss functions, causing the failure of popular gradient descent-based optimization algorithms [ 8 ] (e.g., Sequential Least Squares Quadratic Programming (SLSQP [ 9 ] ), Limited-memory Broyden–Fletcher–Goldfarb–Shanno algorithm with bound support (L-BFGS [ 10 ] ), Levenberg-Marquardt method (LM [ 11 ] ), and Neural Ordinary Differential Equations (NeuralODE [ 12 ] )). While direct search methods (e.g., Nelder-Mead (NM [ 13 ] ) and differential evolution (DE [ 14 ] )) do not use the gradient of the loss function, they often fail to converge in finite time because loss functions are non-smooth [ 15 ] . Artina et al. proposed iterative smoothing and adaptive parametrization techniques to address non-smooth optimization problems [ 16 ] . However, this method is tailored for problems with linear constraints and is not directly applicable to optimization problems constrained by dynamic systems. Recently, artificial neural networks have been used to fit a model of an electric circuit subject to discontinuous forcing to data [ 7 ] . However, this approach requires extensive observations whenever the input to the circuit changes. The application of physics-informed neural networks has also been proposed to infer the underlying forcing term [ 17 ] , but only when the governing models are explicitly provided. Thus, a general, computationally efficient approach to fitting models with discontinuous inputs remains elusive. Here, we present a novel method, the Harmonic Approximation of Discontinuous External Signals using Neural Networks (HADES-NN), to address this challenge. HADES-NN employs neural networks to approximate discontinuous inputs with a sequence of smooth functions [ 18 , 19 ] . These approximate smooth functions are then used to replace the discontinuous input in the non-autonomous system, and the resulting sequence of systems is fitted to the data. Hence, HADES-NN minimizes a sequence of smooth loss functions instead of a single, non-smooth loss function. We prove that the sequence of minima produced by HADES-NN converges to the minimum of the original, non-smooth loss function, yielding accurate parameter estimates even with data is sparse. We illustrate the accuracy of HADES-NN using various examples: a Lotka-Volterra model driven by discontinuous seasonal input, a circadian pacemaker model driven by a discontinuous light signal measured using wearable devices, and a gene regulatory network model in mating yeast subject to abruptly changing pheromone levels. Our results indicate that HADES-NN is a flexible and computationally efficient method for fitting models with discontinuous inputs, greatly extending the range of model systems for which parameters can be estimated. By providing a robust and efficient solution to a long-standing problem, HADES-NN paves the way for more precise and reliable modeling of complex, dynamic systems. This advance has the potential to impact various fields, including biology, engineering, and environmental science, where accurate modeling of systems with discontinuous inputs is crucial. 2 Main results 2.1 Leading optimization algorithms fail to infer parameters of a simple system with a discontinuous input Figure 1: Parameter inference for the Lotka-Volterra system in a changing environment. (A) The Lotka-Volterra model with input S(t) describes the interaction between prey ( y_{1} ) and predator ( y_{2} ) in a changing environment ( S(t) ). (B) When S(t) is smooth (i), so is the corresponding numerical solution of the non-autonomous model (ii). In this case, the Levenberg-Marquardt (LM) and differential evolution (DE) algorithms successfully recover the parameters of the model from as few as 10 observations (iii). Here, the gray box represents the range of parameters from which initial values were randomly selected for optimization. Estimates were normalized by dividing with the true parameter values, located at the intersection of the dashed lines. The green and purple points indicate the final parameter estimates obtained using the LM and DE algorithms, respectively, each from independent random initializations (gray boxes). (C) In contrast, if S(t) changes discontinuously (i), the resulting solution is not differentiable (ii). In this case, both LM and DE produce inaccurate parameter estimates (iii). The discontinuous signal S(t) was generated using a continuous Markov process with a transition rate equal to 0.05 between the states S(t)=0 and S(t)=1 (See Fig. 5 for details). (D) The failure of these two standard optimization algorithms can be explained by the roughness of the loss landscape defined as the L_{2} norm of the difference between the measurements and the numerical solution of the model system at different parameter values: With a smooth S(t) , the contours of the loss landscape are also smooth. However, with discontinuous input, S(t), the contour lines are no longer smooth, resulting in numerous local minima.

Read the complete text