This report addresses challenges in simulating bridge behaviour and highlights limitations in existing models. It explores the potential of artificial intelligence for strain prediction using acceleration data, aiming to contribute to the field. Limitations include studying one bridge, the Banafjäl bridge, and using simulated data, but the neural network is adaptable for broader applications. The report outlines a methodology for predicting strains from accelerations using machine learning, employing simulated data using Fryba’s method and under two load cases, a point load and a train load. Three setups are explored, predicting strains from maximal acceleration points, complete strain signals from complete acceleration signals and future strain point from past acceleration points. Python is used to build neural networks evaluated using mean squared errors and R2 scores. The customized algorithms for each setup and point load result in 12 distinct neural networks to optimize efficiency and data separation. In the first setup, strain points at maximal acceleration point, the predictions fail, with all scenarios yielding in negative R2 scores. However, the low mean squared error suggests prediction values are relatively close to the true valued to the data’s small scale. In the complete signal setup, most R2 scores are positive, indicating correlations between the data and predictions. Negative R2 scores result from specific scenarios, potentially due to overfitting caused by a small training data size, extended training time or model complexity. In the last setup, predicting future strain value from past acceleration signal, results show partial negative R2 scores. For the point loads, all scenarios except for one yield a positive R2 score while the train load is not as successful. In summary, the results show that parameter finetuning is crucial for success and that the pursuit of accurate bridge strain predictions is ongoing and dynamic while offering potential for future safety improvement. The report suggests as a conclusion more extensive parameter testing for optimization. Further research can involve real bridge data, adapting machine learning for complex signals and comparing predictions with finite element models.