Stable Differentiable Modal Synthesis for Learning Nonlinear Dynamics

Victor Zheleznov, Stefan Bilbao, Alec Wright, Simon King · 2026

Read Stable Differentiable Modal Synthesis for Learning Nonlinear Dynamics by Victor Zheleznov, Stefan Bilbao, Alec Wright, Simon King (2026) 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-2601-10453.txt.

Introduction

Victor Zheleznov, Stefan Bilbao, Alec Wright, and Simon King make modal sound synth that can learn. Modal synth builds sound from the tones of a thing. A string, a bar, or a plate each rings in such tones. Past modal tools are fixed once built. They can be slow to tune by ear.

High swings make the tones link in dense ways. The team give a form that stays firm as it runs. It can also be tuned by a learn step. Slopes flow through it, so a net can fit its terms to a target sound. The work joins old synth craft with new learn tools.

Catalog record

CREATORVictor Zheleznov, Stefan Bilbao, Alec Wright, Simon King
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-10453

Browse the full library · About the Tower of Babel

Full text

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

Stable Differentiable Modal Synthesis for Learning Nonlinear Dynamics Victor Zheleznov, Stefan Bilbao, Alec Wright, Simon King Abstract Modal methods are a long-standing approach to physical modelling synthesis. Extensions to nonlinear problems are possible, including the case of a high-amplitude vibration of a string. A modal decomposition leads to a densely coupled nonlinear system of ordinary differential equations. Recent work in scalar auxiliary variable techniques has enabled construction of explicit and stable numerical solvers for such classes of nonlinear systems. On the other hand, machine learning approaches (in particular neural ordinary differential equations) have been successful in modelling nonlinear systems automatically from data. In this work, we examine how scalar auxiliary variable techniques can be combined with neural ordinary differential equations to yield a stable differentiable model capable of learning nonlinear dynamics. The proposed approach leverages the analytical solution for linear vibration of system’s modes so that physical parameters of a system remain easily accessible after the training without the need for a parameter encoder in the model architecture. As a proof of concept, we generate synthetic data for the nonlinear transverse vibration of a string and show that the model can be trained to reproduce the nonlinear dynamics of the system. Sound examples are presented. 1 INTRODUCTION The aim of physical modelling synthesis is to generate sound by numerically solving ordinary or partial differential equations (ODEs/PDEs) that describe the dynamics of an acoustic system. Throughout its long history, various simulation techniques have been employed, including modal synthesis [ 23 ] , finite-difference time-domain methods [ 3 ] and port-Hamiltonian methods [ 19 ] . Numerical stability of such approaches has always been a critical topic in the literature, which has seen a resurgence due to the scalar auxiliary variable (SAV) technique [ 2 , 32 , 30 ] . Under the non-negativity condition on potential energy, the SAV technique provides a way to construct explicit and provably stable numerical solvers for general classes of nonlinear systems. More recently, machine learning approaches have seen an increasing interest in numerical simulation [ 6 , 28 , 18 , 20 ] , including for modelling of distributed musical systems [ 25 , 22 , 7 , 17 , 21 , 8 ] , and can yield differentiable models that are well-suited for learning from data while respecting physical priors [ 9 ] . In contrast with classical approaches, machine learning methods usually operate without stability guarantees. Extrapolation in time beyond intervals seen during training is a significant challenge and can lead to a rapid degradation in solution accuracy [ 25 , 22 , 7 ] . In addition to stability concerns, a common drawback of machine learning approaches is that sampling rate and physical parameters of a system, affecting pitch, timbre and other sonic characteristics, can not be changed after training [ 25 , 22 , 7 , 21 , 8 ] , or the model architecture relies on a parameter encoder to condition the solution [ 17 ] , leading to more trainable parameters and the requirement of a larger dataset containing ground truth data for all desired configurations of a system. In this work, we aim to more deeply integrate the physics of a distributed system into a machine learning framework. We use modal decomposition to construct a system of finite dimension and separate the linear and nonlinear parts of the problem. Then, we replace only a dimensionless memoryless nonlinearity that describes coupling between the modes with a neural network. In particular, we employ gradient networks (GradNets) [ 5 ] that allow an interpretation in terms of a closed-form and non-negative potential function. Finally, we use neural ordinary differential equations (NODEs) [ 6 ] in combination with the SAV technique to yield a stable differentiable model that can learn nonlinear dynamics from data. As a result, physical parameters remain easily accessible and the model generalises to physical parameters, sampling rates and time scales not seen during training. This paper is an expanded version of a paper presented at DAFx25 [ 36 ] and is organised as follows. A simple model of nonlinear transverse string vibration is described in Section 2 . Section 3 derives modal equations for a string, which are then discretised in time using an explicit and stable numerical solver in Section 4 . Section 5 outlines the proposed differentiable model, which is evaluated in Section 6 . Sound examples are available on the accompanying page 1 1 1 https://victorzheleznov.github.io/jaes-modal-node . 2 NONLINEAR TRANSVERSE STRING VIBRATION The general equation of motion describing the transverse nonlinear vibration of a string in a single polarisation is: \mathcal{L}u=\mathcal{F}+\mathcal{F}_{\mathrm{e}}. Here u=u(x,t)\!\!:[0,L]\times\mathbb{R}^{+}\rightarrow\mathbb{R} denotes the transverse displacement of a string of length L and depends on spatial coordinate x in \mathrm{m} and time t in \mathrm{s} . Initial conditions are assumed to be zero. The string is assumed to be simply supported at both ends, implying the following boundary conditions: u(0,t)=\partial_{x}^{2}u(0,t)=u(L,t)=\partial_{x}^{2}u(L,t)=0,\;\forall t\in\mathbb{R}^{+}. Output is assumed to be drawn directly from the string displacement at position x_{\mathrm{o}} as w(t)=u(x_{\mathrm{o}},t) . 2.1 Linear Vibration The linear part of the string vibration is encapsulated in the operator \mathcal{L} , defined as: \mathcal{L}=\rho A\partial_{t}^{2}-T_{0}\partial_{x}^{2}+EI\partial_{x}^{4}+2\sigma_{0}\rho A\partial_{t}-2\sigma_{1}\rho A\partial_{t}\partial_{x}^{2}, where \partial_{t} and \partial_{x} represent partial derivatives with respect to x and t , respectively. Physical parameters that appear in \mathcal{L} are: the material density \rho in \mathrm{kg{\cdot}m^{-3}} ; the string cross-sectional area A=\pi r^{2} in \mathrm{m^{2}} for a string of radius r ; the tension T_{0} in \mathrm{N} ; Young’s modulus E in \mathrm{N{\cdot}m^{-2}} ; and moment of inertia I=\frac{1}{4}\pi r^{4} in \mathrm{m^{4}} . Frequency-independent and dependent loss is characterised by parameters \sigma_{0}\geq 0 and \sigma_{1}\geq 0 , respectively. See [ 3 ] for more on these terms in the context of linear string vibration. 2.2 Nonlinearity Nonlinear dynamics of the string are described in a force density \mathcal{F} that can be expressed through the potential \mathcal{V} : \mathcal{F}(x,t)=\frac{EA-T_{0}}{2}\partial_{x}\big(\mathcal{V}^{\prime}(\xi)\big),\quad\xi\triangleq\partial_{x}u, where prime denotes a derivative of a scalar function. A general model for the potential \mathcal{V} is given by Morse and Ingard [ 24 ] and includes both longitudinal and transverse motion of a string in two polarisations. In this work we neglect the longitudinal motion and one of the two polarisations, leading to the following nonlinear function [ 2 ] : \mathcal{V}(\xi)=\Big(\sqrt{1+\xi^{2}}-1\Big)^{2}. The potential ( 2.2 ) can be approximated by Taylor series as \frac{1}{4}\xi^{4} which corresponds to the model used in our previous work [ 36 ] . Further simplification can lead to a Kirchhoff-Carrier model [ 15 , 4 ] , where the nonlinearity is averaged over the length of the string. Compared to the Kirchhoff-Carrier model, which adequately reproduces only the pitch glide effect, the potential ( 2.2 ) is capable of capturing other perceptually important effects such as phantom partials. 2.3 Plucking Excitation The string is excited by a pointwise external force \mathcal{F}_{\mathrm{e}} , which can be modelled as: \mathcal{F}_{\mathrm{e}}(x,t)=\delta(x-x_{\mathrm{e}})f_{\mathrm{e}}(t), where \delta(x-x_{\mathrm{e}}) is the Dirac delta function at the excitation location x_{\mathrm{e}} . The driving function f_{\mathrm{e}}(t) resembles a pluck of a string and is of the following form [ 1 ] : f_{\mathrm{e}}(t)=\begin{cases}\frac{1}{2}f_{\mathrm{amp}}\big[1-\cos\big(\frac{\pi t}{T_{\mathrm{e}}}\big)\big],\quad&t\in[0,T_{\mathrm{e}}]\\ 0,&\text{otherwise}\end{cases} Here f_{\mathrm{amp}} is the excitation amplitude in \mathrm{N} and T_{\mathrm{e}} is the excitation duration in \mathrm{s} . The excitation starting time is

Read the complete text