Speaker
Description
In preparation for the forthcoming operation of RFX-mod2 in reversed-field pinch configuration, a major upgrade to the experiment’s real-time control system is the transition from a simplified cylindrical approximation to the more accurate toroidal geometry.
However, the existing toroidal reconstruction code is only currently available for post-shot analysis, as it does not satisfy the timing constraints of the control system.
We propose a surrogate model based on Neural Networks (NNs) that serves as the first component in a chain-of-models architecture for real-time magnetic perturbation estimation.
This model approximates the radial profiles of the poloidal and toroidal components of the equilibrium magnetic field, as well as the Shafranov shift.
To enable implementation on FPGA hardware, we apply quantization-aware training (QAT) to convert the model weights from standard 32-bit floating-point representation to fixed-point formats.
By treating the model’s weight bit-width as an additional optimization parameter and using Effective Bit Operations (EBOPs) as a proxy for silicon usage, we perform a multi-objective optimization that explores the trade-off between accuracy and resource cost.
In a subsequent stage, we further analyze the relationship between latency and FPGA resource usage, tuning latency to meet the required real-time performance.
Finally, we test the implemented model across the range of input parameters typically encountered in the experiment, demonstrating that it reproduces target quantities within an acceptable margin of error while meeting timing constraints and achieving substantial savings in FPGA resources compared with more straightforward approaches.
| Minioral | Yes |
|---|---|
| IEEE Member | Yes |
| Are you a student? | Yes |