This is a project which is currently making use of HPC facilities at Newcastle University. It is active.
For further information about this project, please contact:
This MSc dissertation project trains encoder-decoder neural networks on simulated DNA sequences to simultaneously minimise DNA synthesis cycles and correct multiple sequencing errors of mixed types. DNA data storage is an emerging technology with significant potential for long-term archival storage, and this project addresses an open empirical question at the intersection of coding theory and machine learning. Two architectures are evaluated: a convolutional encoder with multi-scale decoder, and a bidirectional GRU encoder-decoder. Experiments are conducted across three redundancy levels, two error types (deletions and substitutions), two error rates, and a systematic sweep of a loss weighting parameter to generate an empirical trade-off curve between synthesis efficiency and error correction capability. The full experimental design requires approximately 50-60 training runs including reproducibility repeats on key configurations.
The project is implemented in Python using PyTorch with GPU acceleration via CUDA. Core architectures include a bidirectional GRU encoder paired with a BiGRU decoder, and a convolutional encoder with a multi-scale transposed convolutional decoder. Gumbel-Softmax sampling is used to provide differentiable discrete base selection during training. A custom differentiable synthesis time function is implemented as a PyTorch tensor operation. Experiments will be submitted as parallel GPU batch jobs to enable concurrent execution of independent configurations, with an estimated total of 21-25 GPU hours across the full sweep. Storage requirements are modest, with datasets consisting of generated sequence tensors requiring less than 1GB.