Using numerically generated datasets of qubit dynamics based on the spin-boson and pure dephasing models, the authors train Random Forests, Support Vector Regressors and Feedforward Neural Networks to infer key properties of the surrounding environment directly from the time evolution of the system’s density matrix.
The models accurately classify the spectral “Ohmicity” of the environment and regress coupling strengths and bath timescales across regimes ranging from weak to strong coupling, and from fast (Markovian) to slow (non-Markovian) baths.
The authors also introduce a practical trace-distance-based metric that allows them to train machine-learning models to quantify non-Markovianity without computing more demanding standard measures.
Overall, the study shows that combining advanced open-quantum-system simulations with machine learning enables automated “quantum noise spectroscopy”, and could help characterize and control realistic quantum devices using experimental time-series data alone.