Artificial neural networks (NNs) bridge input data into output results by approximately encoding the function that relates them. This is achieved after training the network with a collection of known inputs and results leading to an adjustment of the neuron connections and biases. In the context of quantum detection schemes, NNs find a natural playground. In particular, in the presence of a target (e.g. an electromagnetic field), a quantum sensor delivers a response, i.e. the input data, which can be subsequently processed by a NN that outputs the target features. In this work we demonstrate that adequately trained NNs enable to characterize a target with (i) minimal knowledge of the underlying physical model (ii) in regimes where the quantum sensor presents complex responses and (iii) under a significant shot noise due to a reduced number of measurements. We exemplify the method with a development for ¹⁷¹Yb⁺ atomic sensors. However, our protocol is general, thus applicable to arbitrary quantum detection scenarios.

人工神经网络 (NN) 通过对与输入数据相关的函数进行近似编码，将输入数据桥接到输出结果。这是在使用一组已知输入和结果对网络进行训练后实现的，从而调整神经元连接和偏差。在量子检测方案的背景下，神经网络找到了一个天然的游乐场。特别地，在存在目标（例如电磁场）的情况下，量子传感器传递响应，即输入数据，其随后可由输出目标特征的NN处理。在这项工作中，我们证明了经过充分训练的 NN 能够以 (i) 对基础物理模型的最少了解 (ii) 在量子传感器呈现复杂响应的情况下以及 (iii) 由于减少的散粒噪声而导致的显着散粒噪声下表征目标测量次数。我们通过开发来举例说明该方法¹⁷¹ Yb ⁺原子传感器。然而，我们的协议是通用的，因此适用于任意量子检测场景。