Optical Thermodynamic Computing光学热力学计算

Optical thermodynamics seeks to describe the collective behavior of highly multimoded nonlinear photonic systems using macroscopic concepts such as entropy, temperature, chemical potential, and equilibrium. Rather than tracking every optical mode individually, it asks whether large ensembles of interacting modes can be understood through a small set of thermodynamic principles. This perspective has provided a unified language for phenomena such as thermalization, mode redistribution, and the emergence of equilibrium states, while helping address long-standing questions that were difficult to resolve through conventional wave-based descriptions alone.现代光学正进入一个高度多模的时代。在许多光子系统中，成百上千个光学模式可以同时存在；当这些模式通过非线性效应彼此作用时，系统便会呈现出极其复杂的多体运动问题，难以再通过传统理论逐一求解。光热力学为理解这类复杂系统提供了一种新的视角。它不再试图追踪每一个光学模式的具体演化，而是借助熵、温度、化学势等宏观参量，描述高度多模、强非线性光学系统的整体行为，研究能量如何在不同模式之间流动与重新分配，系统如何趋于平衡，以及新的宏观状态如何从复杂的微观相互作用中涌现。这就如同我们会用温度、压强等宏观参量描述屋子里的气体，而从不追踪每一个分子的位置和速度一样，我们放弃追踪每一个光子，并且使用熵来描述系统的演化。通过这一框架，光的热化、模态竞争与能量输运等现象得以被纳入统一的描述，一些长期困扰多模非线性光学领域的问题也因此获得了新的理解与答案。

Beyond explaining complex optical systems, optical thermodynamics may also offer a new way to compute. In thermodynamic matrix inversion, the solution to a linear system is encoded in the correlations and fluctuations of an optical system as it relaxes toward equilibrium, allowing the underlying physics to perform the computation directly. This is particularly promising because modern accelerators are highly efficient at matrix–vector multiplication but far less efficient at operations such as matrix inversion and linear-system solving. As a result, many algorithms that are mathematically attractive remain computationally impractical at scale. By removing this bottleneck, optical thermodynamic computing could make a broader class of algorithms viable and, in doing so, reshape some of the computational foundations of modern artificial intelligence.与此同时，光热力学也为计算提供了一条不同于传统数字处理的路径。以热力学矩阵求逆为例，线性方程组的解可以通过光学系统在平衡态下的涨落与关联直接获得，从而利用系统自身的物理演化完成计算。当前的 GPU 围绕矩阵乘法进行了高度优化，却难以同样高效地处理矩阵求逆和线性方程组求解；因此，一些在算法上很有吸引力的方法，常常由于与现有硬件不匹配而难以大规模应用。若光热力学计算能够显著降低这类运算的时间与能耗，它带来的就不只是更快的计算硬件，也可能让一批过去受制于计算成本的算法重新成为现实，并进一步改变现代人工智能所依赖的算法与计算基础。