TLDR: A new neural operator called STONe (Spatio-Temporal Operator Network) has been developed to accurately forecast high-altitude radiation dose from sparse ground-based neutron sensor data over long periods (up to 180 days). Unlike previous models that require input and output data to be in the same domain and often suffer from error accumulation, STONe learns a direct, non-autoregressive mapping between different physical domains, enabling stable, real-time predictions for critical applications like aviation safety.
Forecasting critical environmental factors, especially those that are difficult to measure directly, has long been a challenge in scientific research. A new study introduces a groundbreaking approach to predict high-altitude radiation doses from sparse, ground-based neutron sensor data, addressing a significant gap in current forecasting capabilities.
The research, titled “CROSS-DOMAIN LONG-TERM FORECASTING: RADIATION DOSE FROM SPARSE NEUTRON SENSOR VIA SPATIO-TEMPORAL OPERATOR NETWORK”, was conducted by Jay Phil Yoo, Kazuma Kobayashi, Souvik Chakraborty, and Syed Bahauddin Alam. Their work presents a novel neural operator framework called the Spatio-Temporal Operator Network (STONe), designed to overcome limitations of existing models that struggle with data from different physical domains and long-term predictions.
Cosmic radiation, particularly Galactic Cosmic Radiation (GCR), poses substantial health risks, especially at aviation altitudes (10–12 km), where dose rates can be 50–100 times higher than at sea level. Aircrew and frequent flyers are particularly vulnerable, with annual doses potentially comparable to or exceeding those of nuclear industry workers. Despite decades of progress in modeling cosmic rays, real-time forecasting of high-altitude radiation fields has remained elusive due to the inherent challenges of sensing and prediction occurring on distinct physical manifolds and over long timescales.
Traditional weather forecasting models, while advanced, typically rely on dense input and output fields that exist on the same spatial grid and within the same physical domain. This fundamental constraint makes them unsuitable for problems like radiation dose forecasting, where ground-based neutron measurements are sparse and the target is a dense, high-altitude radiation field. Furthermore, physics-based atmospheric cascade models are too slow for real-time applications, taking hours to days for accurate forecasts.
Introducing STONe: A New Paradigm
STONe redefines operator learning by introducing a non-autoregressive neural operator that learns a stable functional mapping between heterogeneous domains. Instead of iteratively predicting one step at a time and accumulating errors, STONe predicts the entire future spatiotemporal evolution of the dense output field in a single forward pass. This direct sequence-to-sequence operator bypasses the architectural constraints of existing models.
The framework operates by learning an inverse operator: it directly maps observable proxies (ground-based neutron counts) to the desired high-altitude dose fields, rather than simulating complex intermediate physical processes. This allows for fast, resolution-agnostic reconstruction of global radiation distribution from real-world sensor data.
The STONe architecture consists of two main parts: a branch network and a trunk network. The branch network processes time-series measurements from a global network of 12 sparse, ground-based neutron monitors, spanning 23 years (2001-2023). This network encodes the temporal dependencies. The trunk network, on the other hand, takes spatial query coordinates (latitude and longitude) and generates basis functions, effectively reconstructing the full radiation dose field at high altitudes.
Key Innovations and Advantages
The STONe framework offers several significant contributions:
-
Cross-Domain Operator Learning: It moves beyond models that require input and output functions to be on the same grid, enabling learning between physically distinct manifolds (e.g., ground sensors to high-altitude fields).
-
Inverse Operator Inference: It reframes forecasting as an inverse problem, learning mappings from observable proxies to latent fields without explicitly solving complex governing equations, making real-time inference feasible.
-
Non-Autoregressive Stability: By predicting the full spatiotemporal trajectory in a single pass, STONe eliminates recursive feedback and achieves stable 180-day forecasting without compounding errors, a common issue in iterative models.
-
Real-Time Inference: It achieves millisecond-scale inference latency, providing an operationally viable surrogate for radiation dose forecasting, crucial for applications like flight planning.
-
Universal Operator Paradigm: STONe establishes a general framework that can be applied across various scientific domains, such as climate modeling and geophysical inversion, for data-sparse systems.
Performance and Efficiency
The researchers evaluated STONe using different temporal encoders for its branch network: a Fully Connected Network (FCN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and a Transformer. All models demonstrated competent performance, producing forecasts visually almost indistinguishable from ground truth, especially for short-term predictions.
However, for long-range forecasting (up to 180 days), the GRU-based architecture consistently emerged as the most accurate and stable model, achieving the lowest average errors across various metrics. The Transformer model also showed strong performance, particularly in capturing mid-range dependencies. This highlights the importance of memory-enabled architectures for extended temporal dependencies.
Crucially, all tested architectures maintained sub-millisecond inference times for a full 180-day forecast, making them suitable for real-time operational deployment. This efficiency is vital for applications where rapid response is essential, such as in digital twin systems for predictive monitoring and anomaly detection.
Also Read:
- Probabilistic Kolmogorov-Arnold Networks Enhance Satellite Traffic Forecasting
- Advancing PDE Solutions with Agent-Mediated Neural Operators
Future Directions
While STONe represents a significant advancement, the authors acknowledge areas for future development. These include integrating uncertainty quantification for probabilistic forecasts, addressing memory scalability for even longer horizons or finer resolutions, enhancing generalization to rare extreme space weather events, and gaining a deeper understanding of the model’s learned representations to yield valuable physical insights.
This research provides a foundational contribution to scientific forecasting, extending the theoretical scope of operator learning to reconstruct latent dynamical systems previously inaccessible to direct measurement. The full research paper can be accessed here.
