Sequential learning based PINNs to overcome temporal domain complexities in unsteady flow past flapping wings
Rahul Sundar, Didier Lucor, Sunetra Sarkar · 2025
Read Sequential learning based PINNs to overcome temporal domain complexities in unsteady flow past flapping wings by Rahul Sundar, Didier Lucor, Sunetra Sarkar (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-15679.txt.
Introduction
Rahul Sundar, Didier Lucor, and Sunetra Sarkar improve a modeling method for airflow around flapping wings in this two thousand twenty-five paper. Their tool blends neural networks with the laws of physics. Earlier work by the group let such networks handle moving boundaries without reshaping the frame of reference. Long runs of time, though, strain these networks. Data grow sparse, spectra grow rich, and errors build up.
The authors add sequential learning to the framework. One strategy grows the time window step by step as training proceeds. Another restarts the learning in stages across the full span. They test velocity and pressure recovery for the flapping flow. The goal is a stable model of unsteady flow over long times. Such models could aid the design of small flying machines and wind devices.
Catalog record
| CREATOR | Rahul Sundar, Didier Lucor, Sunetra Sarkar |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 84 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-15679
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Sequential learning based PINNs to overcome temporal domain complexities in unsteady flow past flapping wings Rahul Sundar, Didier Lucor, Sunetra Sarkar Abstract For a data-driven and physics combined modelling of unsteady flow systems with moving immersed boundaries, Sundar et al. [ 1 ] introduced an immersed boundary-aware (IBA) framework, combining Physics-Informed Neural Networks (PINNs) and the immersed boundary method (IBM). This approach was beneficial because it avoided case-specific transformations to a body-attached reference frame. Building on this, we now address the challenges of long time integration in velocity reconstruction and pressure recovery by extending this IBA framework with sequential learning strategies. Key difficulties for PINNs in long time integration include temporal sparsity, long temporal domains and rich spectral content. To tackle these, a moving boundary-enabled PINN is developed, proposing two sequential learning strategies: - a time marching with gradual increase in time domain size, however, this approach struggles with error accumulation over long time domains; and - a time decomposition which divides the temporal domain into smaller segments, combined with transfer learning it effectively reduces error propagation and computational complexity. The key findings for modelling of incompressible unsteady flows past a flapping airfoil include: - for quasi-periodic flows, the time decomposition approach with preferential spatio-temporal sampling improves accuracy and efficiency for pressure recovery and aerodynamic load reconstruction, and, - for long time domains, decomposing it into smaller temporal segments and employing multiple sub-networks, simplifies the problem ensuring stability and reduced network sizes. This study highlights the limitations of traditional PINNs for long time integration of flow-structure interaction problems and demonstrates the benefits of decomposition-based strategies for addressing error accumulation, computational cost, and complex dynamics. 1 Introduction Physics-informed neural networks [ 2 , 3 ] are a hybrid class of supervised and unsupervised learning models capable of learning from labeled data while encoding prior physical knowledge into its loss functions. This has allowed PINNs to solve complex forward and inverse problems in fluid dynamics and other domains. Their framework might also be particularly well-suited in a context of multi-query problems, such as optimization, data assimilation and uncertainty quantification [ 4 ] , and parametric exploration, as they provide a flexible and efficient framework for solving complex time-dependent systems across varying inputs without the need for repeated computationally expensive simulations. However, standard PINNs [ 3 ] could be difficult to train due to various failure modes such as competing objectives [ 5 ] , spectral bias [ 6 ] , and propagation failures [ 7 ] , especially in the case of forward problems. Many authors have proposed adaptive loss component weighting [ 5 , 8 ] , domain decomposition strategies [ 9 , 10 ] , adaptive activation functions [ 11 ] , and modified architectures [ 5 ] , to mitigate the above failure modes. Systems with moving boundaries are more difficult for PINNs to capture because the domain evolves over time, complicating the enforcement of boundary conditions and increasing computational complexity. Additionally, moving boundaries often involve coupled nonlinear dynamics and can introduce sharp gradients or discontinuities in the solution, which are challenging for neural networks to approximate accurately. In this context, some previous works relied on domain transformations to body-attached frame of reference [ 12 ] . But this can be limiting when handling multiple moving bodies or flexible structures. Hence, in the context of unsteady flows past moving/flapping wing-like bodies, the use of a fixed background Eulerian grid would be enviable to train PINNs. This served as the motivation for some of the recent works [ 13 , 14 , 15 , 1 , 16 , 17 ] . Huang et al. [ 13 ] proposed a direct forcing immersed boundary method based PINN (IB-PINN) which was validated for a canonical steady-state flow past a fixed cylinder. Here, IB-PINN used the modified NS equations which included a forcing term in the momentum conservation equation. While the IB-PINN performance was discussed, a comparison with a PINN formulation considering the standard Navier-Stokes equations was missing. Recently, Calicchia et al. [ 15 ] used PINNs based on standard Navier-stokes equations to recover pressure from planar velocity field data obtained from PIV experiments for flow past swimming fish. In addition to the physics loss, the authors used data losses to fit the PIV velocity measurement data in the interior bulk, and a no penetration boundary condition loss at the moving interface. The results were promising, however, a deeper understanding of the relative merits of standard NS-based PINNs or IBM-based PINNs was missing. Towards this, Sundar et al. [ 1 ] proposed an immersed boundary-aware framework for a moving boundary enabled PINN, that included a standard NS equations based formulation (MB-PINN) and an IBM based formulation (MB-IBM-PINN). Sundar et al. [ 1 ] showed MB-PINNs to be useful in scenarios where body position and velocity are known, and MB-IBM-PINN having potential in scenarios where such information might not be available apriori . Additionally, a physics-based vorticity-cutoff based sampling technique was proposed to improve the data sample efficiency of the models while maintaining good accuracy levels. In a connected work [ 18 ] , using a novel zonal loss gradient splitting technique, it was shown how vorticity-cutoff based sampling technique alleviated vanishing gradients. More recently, Zhu et al. [ 17 ] demonstrated the capability of NS-based PINNs (same as MB-PINNs) in modeling forward and hidden physics recovery problems for flow past multiple flexible cylinders with predefined trajectories in three-dimensional spatial domains. Forward problems are not attempted in the current study given their well known difficulties in training PINNs over a long time domain in an error free manner [ 7 , 19 , 17 ] . These difficulties get further accentuated for moving body problems [ 1 , 17 ] . This is especially true when the dynamics strongly depends on the accuracy of the flow-field time history that is being resolved. For forward problems, traditional CFD solvers always triumph over PINNs, which are yet to achieve a similar level of proficiency. However, PINNs work very well for inverse problems such as hidden physics recovery with limited data, which purely data-driven methods cannot achieve. The major challenges, when working with unsteady flow-field data past moving bodies, arise in the temporal domain of the problem. An understanding of the dynamical behavior of the flow and its interaction with the moving body depends on long time integration of the flow-field, posing a challenge for PINNs as they do not train well on long-time domains [ 7 ] . Often, due to memory/storage constraints, the temporal resolution of the obtained data can be coarser than the minimum required Nyquist sampling criterion. In such a scenarios, standard interpolation techniques might not suffice for flow-field reconstruction at intermediate time stamps. Moreover, it might not be feasible to re-run the simulations or experiments for the entire spatial domain when the region of interest is often a truncated spatial subdomain encompassing the near-field region and the moving body. Hence, a physics-informed continuous time-space interpolator like a PINN would be beneficial as it can work under data-sparse conditions [ 20 ] , and with any arbitrary domain [ 21 ] . Importantly, sparsity is a computational source of complexity for PINNs, whereas,