The characterization of complex microwave electric fields remains challenging for precision microwave electrometry using Rydberg atoms. When multiple microwave fields interact simultaneously with Rydberg atoms, multiple dressed-state splittings and overlapping spectral features are produced. The resulting nonlinear relationships between atomic spectra and microwave parameters complicate analytical inversion using semiclassical methods. To address this challenge, a multitask deep learning model combining a residual convolutional neural network (CNN) with a Transformer is proposed. The model exploits spectral responses arising from quantum coherence and microwave dressing. Atomic spectra acquired under the simultaneous application of a known reference microwave field and a field under test are used as model inputs. The model is trained using a reference dataset to predict the frequency and electric field strength of the field under test. The architecture combines residual connections, shared features, and multiple task-specific outputs. A one-dimensional CNN extracts local spectral features, including peak positions, linewidths, and splitting intervals, while residual connections facilitate feature propagation through deeper layers. The Transformer captures global correlations across different spectral regions, integrating local features with the overall spectral structure. Shared features are fed into separate multilayer perceptron output heads for frequency and electric field strength prediction, enabling end-to-end inference from measured atomic spectra.
An experimental reference dataset was constructed using atomic spectra acquired with a reference field at 17.038 GHz and a field under test near 28.916 GHz. Spectra were analyzed for fields under test at 28.914, 28.916, 28.918 and 28.920 GHz, with the reference field fixed at 17.038 GHz. On a held-out test set, the model achieved a frequency classification accuracy of 99.82% for these four frequencies, with an adjacent frequency spacing of 2 MHz. Electric field strength predictions yielded a mean relative error of 0.89% and a coefficient of determination (R
2) of 0.9981. These results indicate that the model effectively learns the nonlinear mapping between dressed-state spectral features and microwave parameters. Noise robustness was numerically evaluated by adding zero-mean Gaussian white noise to the measured spectra, with training and test noise standard deviations varied independently. At a training noise standard deviation of 2.5, frequency classification accuracy was 94.1-99.8% and R
2 for electric field strength prediction was 0.978-0.997 across nine test noise levels (standard deviations 0-4.0). The results indicate that both frequency and electric field strength can still be predicted under different noise levels. To assess applicability to different frequency combinations and larger numbers of microwave fields under test, the experiments were extended to a multifrequency configuration. A reference field at 17.042 GHz was combined with two fields under test at 28.916 and 39.385 GHz. Using a reference dataset constructed for this configuration, the model simultaneously predicted the electric field strengths of both fields under test. For the fields at 28.916 and 39.385 GHz, the R
2 values were 0.9855 and 0.9886, respectively, with corresponding mean relative errors of 2.47% and 2.09%. These results demonstrate the model’s adaptability to different microwave frequencies and its scalability to larger numbers of fields under test.
This work provides a data-driven approach to Rydberg-atom-based multifrequency microwave sensing and parameter extraction for complex microwave electric fields. In future work, a standardized reference dataset will be constructed using atomic spectra acquired under standard electric-field conditions in a microwave anechoic chamber. The applicability of the model will be further investigated for different frequency combinations and configurations involving a larger number of simultaneously applied microwave fields under test. Comparisons with microwave metrology standards will also be conducted for validation.