面向一维量子阱问题的物理信息神经网络
Physics-Informed Neural Networks for One-Dimensional Quantum Well Problems
AI总结:
本文用物理信息神经网络求解一维无限深势阱、有限深势阱和有限势垒的定态薛定谔方程,通过边界条件嵌入、归一化约束和可训练本征能量,验证了PINN学习基态波函数与本征值的可行性。
摘要:
我们实现了物理信息神经网络(PINNs)来求解三个典型一维量子势的定态薛定谔方程:无限深方势阱、有限深方势阱和有限势垒。PINN模型引入了在边界处精确满足边界条件(域边界上的狄利克雷零点)的试探波函数,并优化一个将偏微分方程残差与归一化约束相结合的损失泛函。对于无限深势阱,基态能量已知(无量纲单位下E = pi^2),并在训练中保持固定;而对于有限深势阱和势垒,本征能量被作为可训练参数处理。我们使用具有光滑激活函数的全连接神经网络来表示波函数,并证明PINNs能够学习这些量子系统的基态本征函数和本征值。结果表明,PINN预测的波函数与解析解或预期行为高度吻合,学习到的本征能量收敛到已知值。我们展示了训练日志和能量参数的收敛过程,以及将PINN解与精确结果进行比较的图形。讨论部分涉及PINNs相对于传统数值方法的性能,重点指出了一些挑战,例如收敛到正确本征值的问题、对初始化的敏感性,以及建模不连续势的困难。我们还讨论了归一化项对于解决波函数尺度模糊性的重要性。最后,我们得出结论:PINNs是求解量子本征值问题的一种可行方法,并概述了未来方向,包括扩展到更高维和含时薛定谔方程。
英文摘要:
We implement physics-informed neural networks (PINNs) to solve the time-independent Schrödinger equation for three canonical one-dimensional quantum potentials: an infinite square well, a finite square well, and a finite barrier. The PINN models incorporate trial wavefunctions that exactly satisfy boundary conditions (Dirichlet zeros at domain boundaries), and they optimize a loss functional combining the PDE residual with a normalization constraint. For the infinite well, the ground-state energy is known (E = pi^2 in dimensionless units) and held fixed in training, whereas for the finite well and barrier, the eigenenergy is treated as a trainable parameter. We use fully-connected neural networks with smooth activation functions to represent the wavefunction and demonstrate that PINNs can learn the ground-state eigenfunctions and eigenvalues for these quantum systems. The results show that the PINN-predicted wavefunctions closely match analytical solutions or expected behaviors, and the learned eigenenergies converge to known values. We present training logs and convergence of the energy parameter, as well as figures comparing the PINN solutions to exact results. The discussion addresses the performance of PINNs relative to traditional numerical methods, highlighting challenges such as convergence to the correct eigenvalue, sensitivity to initialization, and the difficulty of modeling discontinuous potentials. We also discuss the importance of the normalization term to resolve the scaling ambiguity of the wavefunction. Finally, we conclude that PINNs are a viable approach for quantum eigenvalue problems, and we outline future directions including extensions to higher-dimensional and time-dependent Schrödinger equations.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。