Neural Kinetics: LEARNING KINETICS AS TRANSPORT IN COMPOSITIONAL SPACE
Year: 2026
Type: preprint
Abstract: Complex bio-derived and fossil mixtures contain thousands of molecules whose individual structures and reaction pathways cannot be fully resolved, making conventional mechanism-based kinetic modelling impractical. A reaction network detailed enough to describe every formula-level transformation would contain far more reactions and parameters than the available data can constrain. We propose to overcome this limitation by representing mixtures directly as abundance fields on a compositional grid indexed by molecular descriptors, specifically carbon number (#C) and double-bond equivalent (DBE), and by learning a time-conditioned, chemistry-informed operator that maps the initial composition to predict a later state. The evolution rates of the chemical space are learned by a two-dimensional Fourier neural operator (FNO) acting jointly on a set of heteroatom classes, with normalized time and compositional coordinates appended as input channels. Because experimental data are limited, we enlarge the training set with synthetic trajectories generated from the measured samples, pre-train on this ensemble, and then calibrate a small subset of parameters on the measured distributions. We tested the framework on a case of oxidation of a complex hydrocarbon mixture that is tracked by FTICR-MS over sixty seconds of reaction. On an independent sample the framework, that we named NeuralKinetics, reproduced the measured endpoint class composition within a mean absolute error of approximately one percent and arranged the heteroatom classes correctly on the compositional grid. The second test was an entirely different system: litter decomposition in a peatland warming experiment, also followed by FTICR-MS but over two years. The developed operator accurately predicted the endpoint class composition of the held-out sample proving that the framework is flexible enough to adapt to diverse reactive systems.
Source: ChemRxiv
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7216489601