Optimizing the ALMA Research Proposal Process with Machine Learning

Every year, astronomers from around the world submit research proposals to the Atacama Large Millimeter Array (ALMA), the largest radio telescope array in the world. The aim of the current work is to streamline the proposal process for astronomers submitting projects to ALMA by suggesting frequency ranges that may be relevant to their research based on their proposal text. We introduce a pipeline of supervised and unsupervised machine learning models, each using various representations of the title and abstract of an incoming proposal. First, a logistic regression filters out proposed projects that are not expected to need specific technical setups. Second, if a technical setup is deemed necessary, our pipeline assigns an incoming project to one of 50 "similar project" groups, defined by topics generated from Latent Dirichlet Allocation (LDA). Third, we apply Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) to mine patterns in measurements ("areas of interest") made in previous projects, for each one of the 50 "similar project" groups. In parallel to the aforementioned topic modeling and HDBSCAN mining, we employ a Multinomial Naive Bayes classifier to predict the broad frequency range defined by the technical limitations of ALMA (frequency band) that we expect a project to make measurements in. Finally, we offer researchers a list of the mined "areas of interest" filtered by the predictions of the Multinomial Naive Bayes classifier. Ultimately, given a proposed project title and abstract, our pipeline generates several recommended "areas of interest" that one should consider measuring in.Regarding the performance of our models, we find that 67.17% of test projects match at least one of the recommended "areas of interest", with an average hit rate of 44.72% across measurements within each test project, when limiting to the top two band predictions. When we disregard band predictions, 88.81% of test projects match at least one recommended "area of interest" with an average hit rate of 60.00% across measurements within each test project. While the latter approach gives better results in terms of hit rates, we believe the combination of models provides a good balance of accuracy and precision.