Towards smarter particle accelerators
For science and society
GANIL / CNRS
2025-11-12
EU : ARTIFACT | the network
41 partners, 14 countries
EU : ARTIFACT | SCIANCE
Overview 
EU : ARTIFACT | SCIANCE
Pilot Areas
EU : ARTIFACT | IRIS
- EMI Device Fingerprints
Electromagnetic interference signatures for equipment identification and monitoring
- Reinforcement learning for accelerator control
Extended capabilities for autonomous beam optimization
- AI driven lattice design
Neural network-based optimization of beam optics
EU : ARTIFACT | TWINRISE
EU : beyond ARTIFACT
FR : M4CAST | The network
FR : M4CAST | MLACC
Highlighs
- Indigo open collaborative space at CC IN2P3 to share data and benchmarks
- Prototyping the deployment of a feature store at IJCLab and CC IN2P3
Fig. Collaborative space access. Credit : T. KachelhofferFR : M4CAST | Case studies
Neural networks based virtual observers
Infering the heat dissipation of SPIRAL2 superconducting cavities with stacked DNN, CNN, LSTM, LSTM attention.Interpretable classification of anomalies
Low Level RF postmortem for the SPIRAL2 LINAC.
Figure : Prediction error of the standard deviation of SPIRAL2 cavities heat load observers. Credit : C. Lassalle.
FR : M4CAST | Case studies
- PointNet inspired architecture for predicting beam parameters and particle clouds applied to ThomX.
- Physics constraints for increased interpretability.
Figure : Prediction performances of LINACnet on particle cloud for ThomX. Credit : E. Goutiere / H. Guler.
.
FR : M4CAST | Case studies
- Beam profile monitors may contain the necessary information to reconstruct an accelerator model.
- They also serve as a critical diagnostic for machine monitors1
Figure : BPM anomaly detection workflow for SuperKEK-B. Credit : Q. Bruant, B. Dalena, F. Bugiotti, C. Ndegwa.
FR : M4CAST | Case studies
- Surrogate models (NN, GP, XGB) are trained on PIC simulations to predict key beam properties across plasma–laser parameter space.
- Enables fast parameter scans and optimisation for the PALLAS laser–plasma injector.
Figure: Prediction maps of beam parameters using several surrogate modelling approaches (NN, GP, XGB, interpolation) for the PALLAS laser–plasma accelerator. C redit: K. Cassou.
FR : M4CAST | Case studies
- Anomalies detection in time series data for the C70XP ARRONAX cyclotron
- Hybrid model combining auto-encoders and isolation forest used
- Initial validation on protons intensities data
- Ongoing extension to multi-variate time series for different accelerator subsytems
