Conference Agenda

This JLESC project is focused on the development of Machine Learning (ML)-based methods for the prediction of flow fields. The research consists of three main directions of work:

• ML-based reduced order models (ROMs) for the reconstruction of flow fields,
• Physics-informed neural networks (PINNs) and,
• Graph convolutional neural networks (GCNNs).

For the PINNs work, a Deep Operator Network (DeepONet) is developed, which allows the prediction of flow fields under arbitrary flow conditions, such as the Reynolds number Re and inflow velocity distributions. This approach is expected to facilitate the performance of a large number of simulations for various scenarios at a low cost. Canonical flow configurations of a backward facing step with varying Re and a channel flow with varying inlet velocity profile are implemented to assess the generalization of the DeepONet. This work has been submitted to a journal and is currently under review. This talk provides an update on this activity and future plans.

GCNNs are developed in this project for different applications, such as for respiratory flows, urban flows and combustion. For respiratory flows, one of the applications is to investigate GCNN models for respiratory flow simulations, that are capable of predicting the air resistance in the nasal cavity.
Additionally, a data augmentation strategy for generating large virtual populations from a limited number of real patient geometries has been developed. This work has recently been published in the Physics of Fluids journal¹. The project talk will feature a detailed update on this work and provide an overview on other activites with respect to the GCNN sub-project.

¹H. Calmet, J. Calafell, R. Puri, B. Johanning-Meiners, A. G. Peiró, R. Sarma, M. Rüttgers, A. Lintermann, G. Houzeaux, Virtual nasal cavity populations for flow prediction with distributed graph convolutional neural networks, Physics of Fluids 38 (2) (2026) 027107. doi:10.1063/5.0304463

Flood dynamics are transitions between low-flow stages which result in small wet areas and high-flow stages which naturally result in large flooded areas. The response of the dynamics of a flood to the time-varying forcing (may it be a hydrograph or precipitation) is precisely what flood models attempt to simulate. Therefore, it is a priori unknown.

The computational load of 2D shallow water simulators is strongly dependent on the number of flooded cells, and thus the flooded area. Consequently, the dynamics of the flooded area translates into time-varying computational demands: low flow stages can be simulated with fewer resources, whereas peak-flow stages demand significantly higher computational capacity. Typically, modellers will choose a set of computational resources which suits the problem size and demands based on experience and preliminary tests. However, these static (used throughout the simulation) resource sets either slow down computations when they are too small for the high flow stages, or make inefficient use of resources when they are too large for the low flow stages. It follows that dynamic resource allocations, based on the computational demands, would be optimal.

In this contribution we present the integration of the SERGHEI-SWE hydrodynamic model with the Dynamic Management of Resources library (DMRlib) to enable malleability —i.e., the runtime adjustment of MPI process counts and computational resources— to improve computational efficiency in shallow-flow simulations. By coupling SERGHEI-SWE with DMRlib, we enable the solver to dynamically expand or shrink its resource set during execution, adapting to these changing computational needs based on minimal heuristics.

SERGHEI-SWE is a high-performance, exascale-ready, scalable shallow water solver supporting CPUs and GPUs. DMRlib extends it with lightweight runtime support for process-level malleability, coordinating with the MPI runtime and job scheduler to manage resource adaptations. Within SERGHEI-SWE, resource reconfiguration is fundamentally a generalization of dynamic domain decomposition, to allow both the size and number of subdomains to change during execution. As a proof-of-concept, we implement minimal heuristics to trigger malleability based on wet-cell fractions: as flooded areas increase, additional resources are requested; when they decrease, resources are released.

The malleable SERGHEI-SWE was evaluated using dam-breaks, river flood, and catchment runoff tests. Numerical accuracy was preserved, with negligible differences relative to static (non-malleable) runs. Dynamic resource management improved computational efficiency relative to minimal fixed-resource configurations. However, performance remained below the best-case static maximum-resource setup, and communication overheads limited gains in low-demand phases. Nonetheless, the proof-of-concept demonstrates both feasibility and potential at larger scales.

The approach is accurate, robust, and promising for improving resource utilization in large-scale hydrodynamic modeling. Future work will focus on refining reconfiguration heuristics, improving understanding of overheads, and combining malleability with dynamic load balancing to better exploit scalable HPC environments.

This talk introduces the progresses made in the past year within our project. After a first review of algorithmic differentiation and its source transformation variant, I will give an overview of the advances in the sparsity detection of the Jacobian within Tapenade, in the differentiation of Julia code, in the creation of a library of differentiated BLAS routines and on improving the integration of OptimalControl.jl with the optimization solvers from the MadSuite, with a particular focus on our automatic differentiation tool ExaModels.jl
I will conclude with current ongoing work and future directions in source transform AD in general and within this project in particular.

Scientific instruments, simulation and data analytics applications are generating or using extremely large datasets that are difficult (if not impossible) to move, store and transform in full. Compression of scientific data, either lossless or lossy, is considered as one of the solutions to address this problem. Already several applications, including ECP applications, are relying on compression to reduce their data. The purpose of the collaboration is first to explore the compression needs at ANL and RCCS from instruments (in particular light sources like the APS and Spring-8), simulations and data analytics involving ML/DL. Second, the collaboration will design new compression algorithms or adapt existing ones responding to ANL and RCCS application needs. Third, the collaboration will explore advanced implementations in FPGA and GPU of the designed compression pipelines.