REPEATABLE SCIENCE IN SEISMIC MODEL TRAINING

Keeping track of the training and evaluation of deep learning models has always been a challenge. Large-scale models require a significant number of experiments to determine the best-performing model, which is usually accomplished through a hyperparameter search of various aspects of the model, such as learning rate, batch size, model architecture, dataset versions, and more. With this growing complexity, the need for reproducibility and tracking becomes important. Similar to DevOps in software development, machine learning operations (MLOps) is a paradigm to accomplish reproducible, reliable, and efficient machine learning models through data management, experiment tracking, and model storage. Our team is developing a seismic foundation model equivalent to large language models for processing text. This development requires a large group of diverse subject matter experts, and this produces the need for consistent record-keeping. The team uses the open-source Python package Curifactory for experiment tracking. The software is structured to enhance experiment reproducibility and tracking. We create parameter files that outline the input parameters that are controllable during an experiment. These parameter files also allow for structured hyperparameter search using grid search instead of ad hoc experiments, which allows comparison across experiment configurations, improvement of model performance, and evaluation of the model performance on seismic data of interest. Curifactory also helps keep track of experiments through automated experiment reporting and caching. These features allowed the team to minimize the time to reproduce results from specific experiments and to facilitate more downstream tasks to improve seismic data analysis work. Furthermore, this structured approach reduced the time to establish a baseline and to create more specific experiments. Using the Stanford Earthquake Dataset (STEAD) to train and evaluate deep learning model architectures, the team has trained a multitask deep learning model that reduces the noise of the signal and measures the arrival time from three-component seismograms simultaneously.