Abstract Background: Current ablation therapy for atrial fibrillation is sub-optimal and long-term response is challenging to predict. Clinical trials identify bedside properties that provide only modest prediction of long-term response in populations, while patient-specific models in small cohorts primarily explain acute response to ablation. We aimed to predict long-term atrial fibrillation recurrence after ablation in large cohorts, by using machine learning to complement biophysical simulations by encoding more inter-individual variability. Methods: Patient-specific models were constructed for 100 atrial fibrillation patients (43 paroxysmal, 41 persistent, 16 long-standing persistent), undergoing first ablation. Patients were followed for 1-year using ambulatory ECG monitoring. Each patient-specific biophysical model combined differing fibrosis patterns, fibre orientation maps, electrical properties and ablation patterns to capture uncertainty in atrial properties and to test the ability of the tissue to sustain fibrillation. These simulation stress tests of different model variants were post-processed to calculate atrial fibrillation simulation metrics. Machine learning classifiers were trained to predict atrial fibrillation recurrence using features from the patient history, imaging and atrial fibrillation simulation metrics. Results: We performed 1100 atrial fibrillation ablation simulations across 100 patient-specific models. Models based on simulation stress tests alone showed a maximum accuracy of 0.63 for predicting long-term fibrillation recurrence. Classifiers trained to history, imaging and simulation stress tests (average ten-fold cross-validation area under the curve 0.85 �� 0.09, recall 0.80 �� 0.13, precision 0.74 �� 0.13) outperformed those trained to history and imaging (area under the curve 0.66 �� 0.17), or history alone (area under the curve 0.61 �� 0.14). Conclusion: A novel computational pipeline accurately predicted long-term atrial fibrillation recurrence in individual patients by combining outcome data with patient-specific acute simulation response. This technique could help to personalise selection for atrial fibrillation ablation. Dataset Description: We include surface meshes in vtk format, consisting of the nodes, triangular elements, the atrial coordinate fields defined on the nodes, and the endocardial and epicardial fibre fields defined on the elements. We also include universal atrial coordinate fields alpha and beta, which are a lateral-septal coordinate and posterior-anterior coordinate for the LA. More details on the coordinate construction are given in our manuscript and https://www.ncbi.nlm.nih.gov/pubmed/31026761. These coordinates can be used for registering datasets. Publication: https://pubmed.ncbi.nlm.nih.gov/35089057/