Photonic Extreme Learning Machine Using Ultrafast Transient Absorption in a Single Quasi-1D ZrSe₃ Nanoribbon

Artificial neural networks underpin modern artificial intelligence but face challenges of scalability, energy consumption, and hardware efficiency as model sizes grow. Photonic approaches offer an attractive alternative by exploiting the parallelism and low thermal footprint of light, yet many implementations still require complex device fabrication or engineered nonlinearities. Extreme learning machines (ELMs) simplify this paradigm by fixing the input-to-hidden mapping and training only a linear output layer, making them highly compatible with physical realizations. Here, a photonic ELM (PELM) framework is introduced based on ultrafast transient absorption (TA) spectroscopy, a widely adopted pump–probe technique operating intrinsically on the femtosecond–picosecond timescale. In this system, inputs are encoded through multiple probe-polarization channels, each parameterized by pump–probe delay, and the resulting TA spectral responses provide high-dimensional nonlinear features without pixelated modulators or nanofabrication. Using a quasi-1D ZrSe₃ nanoribbon, task versatility is demonstrated across nonlinear regression, spiral classification, and image recognition. The approach achieves near-perfect accuracy on the Iris dataset and robust performance on MNIST digits, underscoring the potential of TA-based encoding for physical computation. These results establish ultrafast TA spectroscopy as an experimentally accessible platform and lay the groundwork for future ultrafast, energy-efficient photonic learning systems.