Single channel recordings obtained with the patch-clamp technique have made an enormous impact to our understanding of ion channel function, allowing researchers to obtain detailed and precise information about the kinetic behaviour of ion channels. A major limitation in the current systems used for analysis is that automatic analysis is unreliable and manual analysis is extremely laborious. We present here an “artificial intelligence” (AI) deep learning approach to automatically process large collections of ion channel data. Recently, deep learning, a machine learning development, has begun to revolutionise automatic analysis of data. Here, we have generated semi-synthetic single-channel data and developed deep learning models to detect ion channel current events re-construct idealised records and estimate open probability (Po). <!–![endif]—-> Methods: Ion channel events were simulated from published stochastic gating models by the Gillespie (1977) method, passed through a HEKA patch-clamp EPC7 amplifier (List-Medical, Darmstadt, Germany) and recorded back to file with Signal (CED, Cambridge UK) via an Axon (Molecular Devices, CA, USA) electronic “model cell”. Two different stochastic gating models were used to generate single channel and multi-channel records (up to 5 channels). To validate the proposed DL model, 6 different datasets were used in total throughout the experiments: 3 datasets for single and 3 datasets for multi-channel recordings. Two different models were also compared, with recurrent neural network (LSTM) and convolutional layers. During training, a stochastic gradient descent optimizer was used with 80/20% training/testing split. To reduce overfitting, a dropout layer (0.2) was applied to each layer after batch normalization. Due to a class-imbalance in the datasets, a class weighting scheme was used and loss measured with categorical cross entropy. Results: Our best models correctly detected channel events with 0.98 ± 0.04 accuracy (98%) when a single channel was present in the dataset and 0.91 ± 0.06 accuracy (91%) in datasets with a few single channels present. Open probability (Po) was 99 ± 3% accurate compared to the ground truth/fiducial input. When compared to a traditional method of single channel analysis using QuB software, mean Po was remarkably similar; 0.044±0.017 with traditional software (QuB), vs 0.0436±0.018 was achieved with AI detection. Early models are highly promising. This model will further be adapted with different network architectures and compared the detection power and generalisability. Furthermore, the developed models will be validated to robustly detect transition events in both single and multiple ion channels in a range of biological and biosynthetic datasets without need for re-training. <!–![endif]—->