AI 中文总结
本文提出一种无增益、无非因果元件的原位时域伴随动力学实现方案，通过RLC网络实验验证其可优化时变目标，为复杂系统硬件原生原位时域优化奠定基础。
AI 中文摘要
通过物理系统的内在演化直接优化复杂波动力学，因缺乏可直接获取的梯度而存在根本局限。伴随方法提供了计算梯度的精确途径，已在数值求解器中实现优化，且近年在频域物理平台中也得到应用。然而，其向时域的扩展仍未实现，因为重现时间反向传播似乎需要非因果操作或补偿增益。本文开发了一种方案并通过实验证明，时域伴随动力学可在无增益、无非因果元件或辅助反向网络的线性物理系统中原位实现。通过将时间重映射与系统变量变换相结合，我们获得了可物理实现的伴随演化，其能从可测量信号中构建梯度。我们在复杂RLC网络中对该方法进行了实验验证，实现了针对时变目标的原位优化，包括时间窗口响应和宽带响应。该框架通过允许在同一平台内同时优化系统参数和源激励，统一了物理优化。我们的结果缩小了伴随理论与物理实现之间的差距，为广泛的复杂动力学系统的硬件原生原位时域优化奠定了通用基础。
英文摘要
Direct optimization of complex wave dynamics through the intrinsic evolution of physical systems is fundamentally limited by the lack of directly accessible gradients. Adjoint methods provide an exact route to gradient computation and have enabled optimization in numerical solvers and, more recently, in frequency-domain physical platforms. Yet their extension to the time domain has remained out of reach, as reproducing time-reversed propagation appears to require non-causal operations or compensating gain. Here we develop a protocol and experimentally demonstrate that time-domain adjoint dynamics can be realized in-situ in linear physical systems without gain, non-causal elements, or auxiliary backward networks. By combining time remapping with a transformation of system variables, we obtain a physically realizable adjoint evolution that constructs gradients from measurable signals. We experimentally demonstrate the approach in a complex RLC network, realizing in-situ optimization for time-dependent objectives, including time-windowed and broadband responses. This framework unifies physical optimization by enabling both system parameters and source excitations to be optimized within the same platform. Our results close the gap between adjoint theory and physical implementation, establishing a general foundation for hardware-native, in-situ temporal optimization across a broad class of complex dynamical systems.
Comments10 pages, 4 figures
