Speaker
Description
Integrated modelling of tokamaks combines a host of different codes to self-consistently model plasma discharges with key applications in modelling plasma scenarios and reactor designs as well as analysing experimental discharges. Fast and accurate integrated modeling is crucial to enable rapid iteration and efficient use of limited computational resources. Currently, a major bottleneck in integrated modelling is the running of gyro-kinetic simulations to calculate transport fluxes for each time-step, with even the fastest reduced models using the quasi-linear approximation still being orders of magnitude too slow for routine use. To address this challenge, machine learning surrogate models have been explored as a means of significantly accelerating transport predictions. Previous work has demonstrated the use of real time capable Neural Network (NN) surrogate models trained on QuaLiKiz [1] simulations to predict core tokamak transport quasi-linear heat and particle fluxes [2,3] as well as local linear stability and mode eigenvalues [4].
We investigate the viability of another machine learning algorithm: Gradient Boosted Decision Trees (XGBoost [5]) which have been shown to have superior performance to neural networks in tasks involving tabular data such as that from simulations, and for predicting functions with sharp transitions [6] such as the threshold for the onset of mode instability in plasmas. We compare NNs and XGBoost models trained on linear QuaLiKiz simulations spanning a 22D JET experimental space [3] to predict the stability of a local simulation as well as the growth rates of unstable modes. The performance of both algorithms is examined in a focused comparison study using 100,000 training points and across a broader range of training set sizes, to capture scaling trends, ranging from 100 to 3 million training points.
Both models achieve comparable accuracy using 100,000 data points, however XGBoost has a significant edge in inference speed by a factor of approximately 50 while maintaining smaller model variance due to random initialisation, as well as greatly reduced training and hyperparameter optimisation times. These last two aspects are particularly important when considering active learning within the training pipeline in which models are frequently retrained as new data is acquired. Early scaling results shown in Figure 1 suggest advantages of XGBoost in low-data regimes showing increased robustness to hyperparameter variation.
References:
1] QuaLiKiz homepage: http://qualikiz.com
[2] K.L van de Plassche, J. Citrin, C. Bourdelle, Y. Camenen et al., Phys. Plasmas, 27 (2020)
022310
[3] A. Ho, J. Citrin, C. Bourdelle, Y. Camenen, F. J. Casson, K. L. van de Plassche, H. Weisen, JET Contributors; Neural network surrogate of QuaLiKiz using JET experimental data to populate training space. Phys. Plasmas 1 March 2021; 28 (3): 032305. https://doi.org/10.1063/5.0038290
[4] E. Fransson, A. Gillgren, A. Ho, J. Borsander, O. Lindberg, W. Rieck, M. Åqvist, P. Strand; A fast neural network surrogate model for the eigenvalues of QuaLiKiz. Phys. Plasmas 1 December 2023; 30 (12): 123904. https://doi.org/10.1063/5.0174643
[5] Chen, T., & Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). New York, NY, USA: ACM. https://doi.org/10.1145/2939672.2939785
[6] Léo Grinsztajn, Edouard Oyallon, Gaël Varoquaux. Why do tree-based models still outperform deep learning on typical tabular data?. 36th Conference on Neural Information Processing Systems (NeurIPS 2022) Track on Datasets and Benchmarks, Nov 2022, New Orleans, United States. hal-03723551v3
