Researchers' L-FNO neural operator beats baselines on 11 event benchmarks

Rare, bursty events—like disease outbreaks or semiconductor defects—are notoriously hard to predict. Standard neural operators are typically trained as regression-style function-to-function models, which makes them poor at estimating conditional intensity in sparse event regimes. To fix this, researchers Songhee Kang and Jihoon Kang introduced L-FNO (Lorentzian Fourier Neural Operator), a stochastic neural operator architecture designed specifically for point-process data. L-FNO integrates three components: an FNO-style covariate path for exogenous inputs, Lorentzian spectral kernels to model history-dependent self-excitation, and a likelihood-based training objective instead of the usual regression loss. This combination gives the model a structured spectral memory that naturally captures the bursty, self-triggering dynamics common in operational environments.

The team evaluated L-FNO on eight synthetic point-process benchmarks and three real-world datasets covering disease outbreak prediction and semiconductor fault detection. Across these 11 scenarios, L-FNO outperformed both regression-based and likelihood-based neural operator baselines in event likelihood, calibration diagnostics, and rare-event detection accuracy. The results demonstrate that structured spectral memory and likelihood-based learning offer strong inductive biases for neural operators in stochastic event forecasting. For practitioners, L-FNO provides a principled way to model sparse and bursty events—potentially improving early-warning systems in public health and reliability monitoring in semiconductor manufacturing. The paper is available on arXiv (2608.13562) with code access via linked tools.