Upcoming measurements of the high-redshift 21 cm signal from the Epoch of Reionization (EoR) are a promising probe of the astrophysics of the first galaxies and of cosmological parameters. In particular, the optical depth τ to the last scattering surface of the cosmic microwave background (CMB) should be tightly constrained by direct measurements of the neutral hydrogen state at high redshift. A robust measurement of τ from 21 cm data would help eliminate it as a nuisance parameter from CMB estimates of cosmological parameters. Previous proposals for extracting τ from future 21 cm data sets have typically used the 21 cm power spectra generated by semi-numerical models to reconstruct the reionization history. We present here a different approach which uses convolution neural networks (CNNs) trained on mock images of the 21 cm EoR signal to extract τ. We construct a CNN that improves upon on previously proposed architectures, and perform an automated hyperparameter optimization. We show that well-trained CNNs are able to accurately predict τ, even when removing Fourier modes that are expected to be corrupted by bright foreground contamination of the 21 cm signal. Typical random errors for an optimized network are less than 3.06%, with biases factors of several smaller. While preliminary, this approach could yield constraints on τ that improve upon sample-variance limited measurements of the low-ℓ EE observations of the CMB, making this approach a valuable complement to more traditional methods of inferring τ.

即将对来自再电离时代 (EoR) 的高红移 21 厘米信号进行的测量是对第一批星系的天体物理学和宇宙学参数的有希望的探测。特别是，宇宙微波背景 (CMB) 最后一个散射面的光学深度τ应该受到高红移下中性氢态的直接测量的严格限制。从 21 厘米数据中对τ 进行稳健的测量将有助于从 CMB 宇宙学参数估计中消除它作为一个令人讨厌的参数。以前提取τ 的建议从未来的 21 cm 数据集中，通常使用半数值模型生成的 21 cm 功率谱来重建再电离历史。我们在这里提出了一种不同的方法，它使用在 21 cm EoR 信号的模拟图像上训练的卷积神经网络 (CNN) 来提取τ。我们构建了一个改进先前提出的架构的 CNN，并执行自动超参数优化。我们表明，训练有素的 CNN 能够准确预测τ，即使去除了预计会被 21 cm 信号的明亮前景污染破坏的傅立叶模式。优化网络的典型随机误差小于 3.06%，偏差因子小几个。虽然是初步的，但这种方法可能会对τ产生限制改进了对 CMB的低ℓ EE 观测值的样本方差限制测量，使这种方法成为对更传统的推断τ方法的有价值的补充。