The past decade has witnessed a paradigm shift in meteor science, driven by rapidly expanding observational networks and advances in machine learning. Optical systems now generate millions of meteor orbits annually, volumes that exceed the capacity of traditional analysis pipelines. This work reviews the application of machine learning (ML) and deep learning (DL) algorithms to meteor detection, classification, shower identification, and physical modeling. CNN-based classifiers now routinely achieve high recall with sub-percent false-positive rates on optical data, while CNNs trained on synthetic radar data transfer effectively across multiple high-power large-aperture facilities without requiring labeled data at each site. Unsupervised density-based clustering (DBSCAN, HDBSCAN) has begun to supersede the D -criterion paradigm for meteor shower identification. Recent automated statistical frameworks have delivered the first rigorous uncertainty quantification for meteoroid physical properties from ablation model fits, though neural network surrogates and physics-informed architectures remain unexplored. Significant challenges persist in class imbalance handling, the absence of standardized benchmark datasets, limited model interpretability, and poor reproducibility. Because datasets, class distributions, and evaluation protocols differ substantially across studies, the performance figures reported in the literature cannot be directly compared; establishing community benchmarks is therefore an urgent priority. We identify future directions including standardized data sets, physics-informed neural networks for ablation modeling, and real-time edge computing deployment. These directions, along with a systematic attention to benchmarking and reproducibility, will change the path from isolated classification methods to an integrated ML bundle for meteor science.
Simon Anghel (2026) studied this question.