Empirical machine learning for 3He spin-filter polarisation decay

Résumé :

A lightweight, empirical machine-learning framework is presented for predicting the time evolution of the nuclear polarisation of 3He spin-filter cells under realistic neutron beam line conditions. The method bridges the gap between detailed microscopic modelling of polarisation relaxation and purely heuristic correction schemes by combining supervised learning with basic physics-informed constraints on spin relaxation. This strategy ensures physically consistent polarisation forecasts while maintaining computational efficiency and ease of implementation.

The framework is tailored to the limited number of heterogeneous experimental datasets typically encountered during routine instrument operation. Rather than explicitly modelling all relaxation mechanisms, it learns an effective representation of polarisation decay directly from experimental data. Within the domain spanned by the training dataset, the model demonstrates stable predictive performance while preserving the expected exponential relaxation behaviour and accommodating non-linear, history-dependent effects.

The resulting predictions enable reliable, time-dependent corrections of neutron scattering data affected by polarisation efficiency drift. Compared with traditional single-parameter relaxation models, the approach accounts for the combined influence of multiple experimental parameters and their temporal evolution. Although extrapolation beyond the training domain requires careful validation, the method provides an operationally efficient and reproducible tool for polarisation monitoring and correction in realistic experimental environments.

College 6 Secretary

Gabriel Cuello

External visitors may ask for a site access to Brigitte Dubouloz (dubouloz@ill.fr)