Hybrid quantum-classical models for methane emissions forcasting

Overview

North Africa is a climate hotspot, and Algeria, Egypt and Libya are among the main methane emitters because of oil and gas production. Satellites such as Sentinel-5P (TROPOMI) produce large methane time series, but classical models struggle with their complexity.

This project, part of a national research effort on climate and agriculture, asks whether quantum machine learning can forecast methane better. It had not been applied to greenhouse gas forecasting in this region before. The work was published in Quantum Information Processing.

Key components and technical implementation

• Data: Sentinel-5P CH4 column mixing ratio from February 2019 to March 2024, pulled through Google Earth Engine, for 27 cities in Algeria, Egypt, Libya, Morocco and Tunisia, split into four sub-datasets (Central, West and South Algeria, and the other countries).
• Task: forecast the next day’s methane from a sequence of 3 days, with z-score normalization. Training uses data before 1 January 2023, and the test provinces were never seen in training, so models are tested on unseen time and unseen places.
• QLSTM and QGRU: the LSTM and GRU gates are replaced by variational quantum circuits (6 and 3 circuits per cell). Each circuit uses Hadamard and arctan-based angle encoding, CNOT entanglement, trainable rotations and Pauli-Z measurement.
• Proposed hybrids: QCNN-LSTM and QCNN-GRU, where a 1D quantum convolution (kernel size 3, two ansatz designs including a general SU(4) gate) extracts features before the recurrent layer.
• Baselines and tuning: classical LSTM, GRU, CNN-LSTM and CNN-GRU, each tuned separately. Quantum models trained for 15 epochs with Adagrad. Implemented with PennyLane (default.qubit) and PyTorch, classical models with TensorFlow/Keras. Metrics are MAE, MSE, RMSE and R².

Key research findings and achievements

• Among the quantum models, the proposed QCNN-GRU and QCNN-LSTM gave the lowest RMSE in every region. QGRU often had the highest R².
• Against the best classical model per region, the quantum hybrids had lower RMSE in three of four sub-datasets, for example 1.76 vs 7.50 on the largest one (other countries) and 1.56 vs 3.13 in West Algeria. On South Algeria, the classical LSTM was slightly better (2.13 vs 2.17).
• Quantum models reached these results in 15 epochs, while the classical baselines needed 200 to 500.
• Classical models kept higher R² in several regions, so the result is a lower error with faster training, not a win on every metric.
• All results are from simulation, not real quantum hardware. Next steps are Transformer baselines, explainable AI, real devices and other gases such as CO₂ and NOx.
• Achievement: peer-reviewed journal paper (Springer) with a new quantum-convolutional recurrent architecture for environmental time series.