This paper addresses the problem of estimating rate coefficients for complex chemical reactions. The strong nonlinearity (stiffness) of real-world atmospheric chemical systems leads to training instability and poor convergence in existing learning-based approaches. To address this, we propose SPIN-ODE, a physics-based neural ODE framework that incorporates strong nonlinearity. SPIN-ODE involves a three-step optimization process: 1) fitting concentration trajectories using a black-box neural ODE; 2) pretraining a chemical reaction neural network (CRNN) to learn the mapping between concentration and time derivatives; and 3) integrating the pretrained CRNN with the ODE to fine-tune the rate coefficients. Extensive experiments on synthetic data and a newly proposed real-world dataset demonstrate the effectiveness and robustness of the proposed method. This is the first study to apply a neural ODE with strong nonlinearity to chemical rate coefficient discovery, suggesting promising directions for integrating neural networks with detailed chemistry.
