S3: Bridging sciml and numerical methods: unified frameworks and fair comparisons
Co-chairs:
Sean De Marco* (Imperial College London)
Divij Ghose (Imperial College London)
Lloyd Fung (Imperial College London)
Ben Moseley (Imperial College London)
Keywords: Numerical Methods, Scientific Machine Learning, Hybrid Modelling, Benchmarking and Reproducibility, Scientific Computing.
ABSTRACT
Scientific machine learning (SciML) has developed rapidly in recent years. It offers promising ways to accelerate simulation-based research, including physics-informed networks, neural operators and, most recently, foundation models for PDEs. However, it has largely developed in isolation from classical numerical analysis. Recent meta-studies show that many reported speed-ups over numerical solvers shrink or vanish against strong baselines, and SciML methods still face challenges in interpretability, robustness and generalisation. At the same time, many of the most recent and operationalised SciML methods increasingly resemble numerical methods. Hybrid approaches such as learned closures and correctors, ML-generated preconditioners and initial guesses, neural basis functions within Galerkin and domain-decomposition schemes, and differentiable solvers are gaining traction because they offer better interpretability, generalisation and performance.
This mini-symposium has three aims. The first is to bring together numerical analysts and SciML researchers and to foster a common language between the two fields. The second is to promote unified frameworks that view SciML methods as numerical methods, and vice versa. The third is to encourage credible evaluation, in which solvers are compared at matched accuracy and on matched hardware, with costs that include training. We welcome work that places ML and numerical methods within a single computational or theoretical framework, fair benchmarks across paradigms (including negative results), and hybrid methods that make their trade-offs explicit. Together, such contributions will clarify when learning helps, when classical methods remain preferable, and what becomes possible when the strengths of both are combined.
REFERENCES
[1] Brunton, S.L., Kutz, J.N. “Promising directions of machine learning for partial differential equations.” Nat Comput Sci 4, 483–494 (2024).
[2] Koehler, Felix, et al. “APEBench: A benchmark for autoregressive neural emulators of PDEs.” Advances in Neural Information Processing Systems 37 (2024): 120252-120310.
[3] McGreivy, N., Hakim, A. “Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations.” Nat Mach Intell 6, 1256–1269 (2024).
[4] Kochkov, D., Yuval, J., Langmore, I. et al., “Neural general circulation models for weather and climate.” Nature 632, 1060–1066 (2024).
[5] Zhang, E., Kahana, A., Kopaničáková, A. et al. “Blending neural operators and relaxation methods in PDE numerical solvers.” Nat Mach Intell 6, 1303–1313 (2024).