Handling time-series astronomical data with SNN architecture for exoplanet detection and characterization

NABS

For entertainment purposes, I call it NABS - Neural Astro-Biosignature Scanner, inspired by how many times Mr. Spock says, “The scanner can’t identify any life forms, captain. Fascinating...”

---

🔭 Project phase: exploration

Summary

This study explores the utilization of Spiking Neural Network (SNN) architecture for the analysis of time-series astronomical data, focusing on exoplanet detection and characterization. While SNNs are relatively novel in the domain of astronomical data analysis, this research presents a prototype inspired by Kaggle and snnTorch tutorials as its first model. The primary goal is to train small-scale SNN models capable of discerning subtle changes in time-series data during exoplanet transits and to evaluate its performance across epochs.

Overview

Spiking Neural Networks are a type of artificial neural network inspired by the biological neurons in the brain. Unlike traditional artificial neural networks that use continuous-valued activations, SNNs operate on spikes, which are discrete events that represent the firing of a neuron. This spike-based communication allows SNNs to process temporal information and handle asynchronous input. In this project, I want to explore how SNNs can be utilized as the underlying architecture for a deep-learning model aimed at exoplanet detection.

Problem

The transit method, although highly succesful in detecting exoplanets, can present challenges when considering the practicality of AI-enhanced probes in deep space. The transits are often infrequent and brief, which makes gathering sufficient data time and power-consuming, thus rendering space exploration unsustainable.

Potential solution

To address this problem, I propose an experiment leveraging Spiking Neural Networks (SNNs). SNNs offer advantages in processing sparse and temporal data, potentially effective in analyzing intermittent transit signals.

Goals

• preprocess available data;
• study baseline models and existing methods used for exoplanet detection;
• augment the data for imbalance;
• train small-scale SNN models to analyze subtle changes in the spectrum during transit;
• evaluate models' performance over each epoch using sensitivity, specificity, test loss and AUC ROC;
• optimize models;
• compare SNN's methods with other methods;
• compare the performance of SNNs with traditional machine learning algorithms;
• documenting experiments and delivering results.

Note

I'm looking to refine and narrow down the project's objectives and structure. That's why I split it into "phases". Right now, I'm in the exploration phase, tapping into time-series modeling, acquainting myself with various forms of astronomical data, and reviewing literature.

Milestones

• First model (light curve data from Kepler and snnTorch package)
• Classification testing on the first model
• Second model (lightkurve package and tsai)
• Studying random forests and decision trees for benchmarking
• Studying suitable machine learning models for time series forecasting tasks
• Third model (atmospheric spectra)