EVENT DOSSIER
The researchers proposed a fast differentiable agent model based on parameter-conditioning Fourier neural operators (FNOs) for simulating the excitability and oscillation dynamics of the FitzHugh-Nagumo system. This study focused on the spatial domain containing five physiological parameters, aiming to address the high computational cost of traditional finite difference solvers during parameter scanning. The Fourier layer was conditioned using feature linear modulation (FiLM), and dedicated operators were trained separately for the oscillatory mode and the excitatory mode. Experimental results showed that in the oscillatory mode, the agent model had a relative L² error of less than 0.1%, operated three orders of magnitude faster than the finite difference baseline, and possessed uniform generalization and extrapolation capabilities. In the excitatory mode, the model accurately reproduced the relationship between the discharge threshold and the square root of the conduction velocity and diffusion coefficient, fully capturing the full propagation pulses and the excitatory bifurcation structure.
研究人员训练参数条件化傅里叶神经算子（FNOs）作为 FitzHugh-Nagumo 系统的快速可微代理模型。该研究针对激活 - 抑制结构，在包含五个生理参数的空间内探索神经元电压动力学，以解决经典有限差分求解器参数扫描成本高的问题。通过特征线性调制（FiLM）对傅里叶层进行条件化，并在振荡（持续放电）和兴奋性（动作电位传播）两个区域分别训练一个算子。结果显示，在振荡模式下，代理模型在两个场变量上相对 $L^2$ 误差低于 0.1%，运行速度比有限差分基线快三个数量级，且能均匀泛化并外推至极低百分比误差；在兴奋性模式下，该算子准确复现了放电阈值和传导速度与 $\sqrt{D_u}$ 成正比的规律，完整捕捉了全传播脉冲及兴奋性分岔结构。