NTH

Sample-Efficient Control of Magnetoconvective Flows via Differentiable Simulation

AuthorsJannis Becktepe, Dmitry Krasnov, Nils Thuerey, Jörg Schumacher, Sebastian Peitz

AffiliationsTU Dortmund University & Lamarr Institute for Machine Learning and Artificial Intelligence and Institute of Thermodynamics and Fluid Mechanics, TU Ilmenau and Physics-based Simulation Group, TU Munich & Munich Center for Machine Learning and Institute of Thermodynamics and Fluid Mechanics, TU Ilmenau and Scientific Machine Learning Group, Paderborn University

November 21, 2026 2 min read
Watch on YouTube
The one-line take

A differentiable simulator helps researchers control magnetoconvective flows more efficiently and design better heat-transfer outcomes.

Key results

20
Average solver speedup

× over PICT on FluidGym benchmarks.

28.59%
Heat-transfer improvement

BPTT improvement in Nusselt number at Ra=20Q.

1000
BPTT convergence

Approximate forward-equivalent steps to reach final reward.

What the paper found

ϕ-PICT is a differentiable, GPU-accelerated solver for low-magnetic-Reynolds-number magnetohydrodynamics, built to make liquid-metal flow control and inverse design practical. It extends a PISO solver with electric-potential and Lorentz-force calculations, conservative current-density formulation, and analytical gradients; an adjoint solve propagates gradients through the electric-potential Poisson equation without memory growing with the number of linear-solver iterations. On FluidGym benchmarks, the optimized solver achieved a 20× average speedup over PICT, measured on an NVIDIA A100. In closed-loop convection control, backpropagation through time (BPTT) raised heat transfer by 28.59% at Ra=20Q and reached its final reward after about 1,000 forward-equivalent steps, outperforming Soft Actor-Critic (SAC) and Short-Horizon Actor-Critic (SHAC) in final reward. For inverse design, L-BFGS-B used gradients from ϕ-PICT to infer defective wall conductivities from sparse velocity measurements, producing an objective roughly three orders of magnitude smaller than CMA-ES within the same time limit. The results show how differentiable simulation can reduce the data demands of both physical control and parameter recovery.

Original abstract

Magnetohydrodynamic (MHD) fluid flows are central to demanding engineering applications, most prominently fusion blanket cooling, where flow control is a promising tool to optimize key properties such as heat transfer. While active flow control (AFC) has recently advanced through machine learning, its application to MHD flows remains largely unexplored. In this work, we present $φ$-PICT, a novel differentiable GPU-accelerated MHD solver targeted at AFC. We validate $φ$-PICT against MHD benchmarks and demonstrate its high computational efficiency. Integrated into FluidGym, $φ$-PICT enables sample-efficient gradient-based algorithms to improve heat transfer by $>25$% and allows optimizers to achieve $10^3\times$ higher accuracy in inverse design than black-box methods. $φ$-PICT is publicly available at https://github.com/upb-sciml/phiPICT.

Read the original paper

More in AI for Science

Browse all 58 papers →